Bioinformatics

This book contains the latest material in the subject, covering next generation sequencing
(NGS) applications and ­meeting the requirements of a complete semester course. This book
digs deep into analysis, providing both concept and practice to satisfy the exact need of
researchers seeking to understand and use NGS data reprocessing, genome assembly, vari-
ant discovery, gene profiling, epigenetics, and metagenomics. The book does not introduce
the analysis pipelines in a black box, but with detailed analysis steps to provide readers
with the scientific and technical backgrounds required to enable them to conduct analysis
with confidence and understanding. The book is primarily designed as a companion for
researchers and graduate students using sequencing data analysis but will also serve as a
textbook for teachers and students in biology and bioscience.
Chapman & Hall/CRC Computational Biology Series
About the Series: This series aims to capture new developments in computational ­biology, as well
as high-quality work summarizing or contributing to more established t­ opics. Publishing a broad
range of reference works, textbooks, and handbooks, the series is designed to appeal to students,
researchers, and professionals in all areas of computational biology, including genomics, pro-
teomics, and cancer computational biology, as well as interdisciplinary researchers involved in
associated fields, such as bioinformatics and ­systems biology.

Metabolomics: Practical Guide to Design and Analysis

Ron Wehrens and Reza Salek
An Introduction to Systems Biology: Design Principles of Biological Circuits, Second Edition
Uri Alon
Computational Biology: A Statistical Mechanics Perspective, Second Edition
Ralf Blossey
Stochastic Modelling for Systems Biology, Third Edition
Darren J. Wilkinson
Computational Genomics with R
Altuna Akalin, Bora Uyar, Vedran Franke, and Jonathan Ronen
An Introduction to Computational Systems Biology: Systems-Level Modelling of Cellular Networks
Karthik Raman
Virus Bioinformatics
Dmitrij Frishman and Manuela Marz
Multivariate Data Integration Using R: Methods and Applications with the mixOmics Package
Kim-Anh LeCao and Zoe Marie Welham
Bioinformatics: A Practical Guide to NCBI Databases and Sequence Alignments
Hamid D. Ismail
Data Integration, Manipulation and Visualization of Phylogenetic Trees
Guangchuang Yu
Bioinformatics Methods: From Omics to Next Generation Sequencing
Shili Lin, Denise Scholtens, and Sujay Datta

For more information about this series please visit: https://www.routledge.com/Chapman-

HallCRC-Computational-Biology-Series/book-series/CRCCBS
 Bioinformatics
A Practical Guide to Next Generation
­Sequencing Data Analysis

Hamid D. Ismail
First edition published 2023
by CRC Press
4 Park Square, Milton Park, Abingdon, Oxon, OX14 4RN

and by CRC Press

6000 Broken Sound Parkway NW, Suite 300, Boca Raton, FL 33487-2742

© 2023 Hamid D. Ismail

CRC Press is an imprint of Informa UK Limited

The right of Hamid D. Ismail to be identified as author of this work has been asserted in accordance with sections 77
and 78 of the Copyright, Designs and Patents Act 1988.

All rights reserved. No part of this book may be reprinted or reproduced or utilised in any form or by any electronic,
mechanical, or other means, now known or hereafter invented, including photocopying and recording, or in any
­information storage or retrieval system, without permission in writing from the publishers.

For permission to photocopy or use material electronically from this work, access www.copyright.com or contact the
Copyright Clearance Center, Inc. (CCC), 222 Rosewood Drive, Danvers, MA 01923, 978-750-8400. For works that are
not available on CCC please contact [email protected]

Trademark notice: Product or corporate names may be trademarks or registered trademarks, and are used only for
identification and explanation without intent to infringe.

Library of Congress Cataloging-in-Publication Data

Names: Ismail, Hamid D., 1967- author.
Title: Bioinformatics : a practical guide to next generation sequencing data analysis / Hamid D. Ismail.
Description: First edition. | Boca Raton : CRC Press, 2023. |
Series: Computational biology series | Includes bibliographical references and index.
Identifiers: LCCN 2022048850 (print) | LCCN 2022048851 (ebook) | ISBN 9781032409009 (hbk) |
ISBN 9781032408910 (pbk) | ISBN 9781003355205 (ebk)
Subjects: LCSH: Bioinformatics. | Genetics—Technique. | Nucleotide sequence.
Classification: LCC QH324.2.I86 2023 (print) | LCC QH324.2 (ebook) | DDC 570.285—dc23/eng/20221014
LC record available at https://lccn.loc.gov/2022048850
LC ebook record available at https://lccn.loc.gov/2022048851

ISBN: 978-1-032-40900-9 (hbk)

ISBN: 978-1-032-40891-0 (pbk)
ISBN: 978-1-003-35520-5 (ebk)

DOI: 10.1201/9781003355205

Typeset in Minion
by codeMantra
Contents

Preface, xiii
Author, xvii

Chapter 1   ◾   Sequencing and Raw Sequence Data Quality Control 1

1.1 NUCLEIC ACIDS 1
1.2 SEQUENCING 3
1.2.1 First-Generation Sequencing 3
1.2.2 Next-Generation Sequencing 4
1.2.2.1 Roche 454 Technology 5
1.2.2.2 Ion Torrent Technology 6
1.2.2.3 AB SOLiD Technology 6
1.2.2.4 Illumina Technology 7
1.2.3 Third-Generation Sequencing 8
1.2.3.1 PacBio Technology 9
1.2.3.2 Oxford Nanopore Technology 10
1.3 SEQUENCING DEPTH AND READ QUALITY 11
1.3.1 Sequencing Depth 11
1.3.2 Base Call Quality 11
1.4 FASTQ FILES 13
1.5 FASTQ READ QUALITY ASSESSMENT 18
1.5.1 Basic Statistics 23
1.5.2 Per Base Sequence Quality 24
1.5.3 Per Tile Sequence Quality 25
1.5.4 Per Sequence Quality Scores 28
1.5.5 Per Base Sequence Content 28
1.5.6 Per Sequence GC Content 28
1.5.7 Per Base N Content 30

v
vi   ◾    Contents

1.5.8 Sequence Length Distribution 30

1.5.9 Sequence Duplication Levels 31
1.5.10 Overrepresented Sequences 31
1.5.11 Adapter Content 32
1.5.12 K-mer Content 33
1.6 PREPROCESSING OF THE FASTQ READS 34
1.7 SUMMARY 45
REFERENCES 46

Chapter 2   ◾    Mapping of Sequence Reads to the Reference Genomes 49

2.1 INTRODUCTION TO SEQUENCE MAPPING 49
2.2 READ MAPPING 55
2.2.1 Trie 56
2.2.2 Suffix Tree 56
2.2.3 Suffix Arrays 57
2.2.4 Burrows–Wheeler Transform 58
2.2.5 FM-Index 62
2.3 READ SEQUENCE ALIGNMENT AND ALIGNERS 63
2.3.1 SAM and BAM File Formats 65
2.3.2 Read Aligners 70
2.3.2.1 Burrows–Wheeler Aligner 71
2.3.2.2 Bowtie2 75
2.3.2.3 STAR 76
2.4 MANIPULATING ALIGNMENTS IN SAM/BAM FILES 79
2.4.1 Samtools 79
2.4.1.1 SAM/BAM Format Conversion 79
2.4.1.2 Sorting Alignment 80
2.4.1.3 Indexing BAM File 80
2.4.1.4 Extracting Alignments of a Chromosome 81
2.4.1.5 Filtering and Counting Alignment in SAM/BAM Files 81
2.4.1.6 Removing Duplicate Reads 82
2.4.1.7 Descriptive Statistics 83
2.5 REFERENCE-GUIDED GENOME ASSEMBLY 83
2.6 SUMMARY 85
REFERENCES 86
 Contents   ◾    vii

Chapter 3   ◾    De Novo Genome Assembly 89

3.1 INTRODUCTION TO DE NOVO GENOME ASSEMBLY 89
3.1.1 Greedy Algorithm 90
3.1.2 Overlap-Consensus Graphs 90
3.1.3 De Bruijn Graphs 91
3.2 EXAMPLES OF DE NOVO ASSEMBLERS 93
3.2.1 ABySS 93
3.2.2 SPAdes 97
3.3 GENOME ASSEMBLY QUALITY ASSESSMENT 99
3.3.1 Statistical Assessment for Genome Assembly 100
3.3.2 Evolutionary Assessment for De Novo Genome Assembly 103
3.4 SUMMARY 106
REFERENCES 107

Chapter 4   ◾    Variant Discovery 109

4.1 INTRODUCTION TO GENETIC VARIATIONS 109
4.1.1 VCF File Format 110
4.1.2. Variant Calling and Analysis 113
4.2 VARIANT CALLING PROGRAMS 114
4.2.1 Consensus-Based Variant Callers 114
4.2.1.1 BCF Tools Variant Calling Pipeline 115
4.2.2 Haplotype-Based Variant Callers 125
4.2.2.1 FreeBayes Variant Calling Pipeline 127
4.2.2.2 GATK Variant Calling Pipeline 129
4.3 VISUALIZING VARIANTS 143
4.4 VARIANT ANNOTATION AND PRIORITIZATION 143
4.4.1 SIFT 145
4.4.2 SnpEff 148
4.3.3 ANNOVAR 151
4.3.3.1 Annotation Databases 153
4.3.3.2 ANNOVAR Input Files 156
4.5 SUMMARY 160
REFERENCES 161
viii   ◾    Contents

Chapter 5   ◾    RNA-Seq Data Analysis 163

5.1 INTRODUCTION TO RNA-SEQ 163
5.2 RNA-SEQ APPLICATIONS 165
5.3 RNA-SEQ DATA ANALYSIS WORKFLOW 166
5.3.1 Acquiring RNA-Seq Data 166
5.3.2 Read Mapping 167
5.3.3 Alignment Quality Assessment 171
5.3.4 Quantification 172
5.3.5 Normalization 174
5.3.5.1 RPKM and FPKM 174
5.3.5.2 Transcripts per Million 175
5.3.5.3 Counts per Million Mapped Reads 175
5.3.5.4 Trimmed Mean of M-values 175
5.3.5.5 Relative Expression 176
5.3.5.6 Upper Quartile 176
5.3.6 Differential Expression Analysis 176
5.3.7 Using EdgeR for Differential Analysis 180
5.3.7.1 Data Preparation 181
5.3.7.2 Annotation 183
5.3.7.3 Design Matrix 184
5.3.7.4 Filtering Low-Expressed Genes 185
5.3.7.5 Normalization 186
5.3.7.6 Estimating Dispersions 186
5.3.7.7 Exploring the Data 189
5.3.7.8 Model Fitting 194
5.3.7.9 Ontology and Pathways 202
5.3.8 Visualizing RNA-Seq Data 204
5.3.8.1 Visualizing Distribution with Boxplots 206
5.3.8.2 Scatter Plot 207
5.3.8.3 Mean-Average Plot (MA Plot) 208
5.3.8.4 Volcano Plots 209
5.4 SUMMARY 209
REFERENCES 211
 Contents   ◾    ix

Chapter 6   ◾    Chromatin Immunoprecipitation Sequencing 213

6.1 INTRODUCTION TO CHROMATIN IMMUNOPRECIPITATION 213
6.2 CHIP SEQUENCING 214
6.3 CHIP-SEQ ANALYSIS WORKFLOW 215
6.3.1 Downloading the Raw Data 217
6.3.2 Quality Control 218
6.3.3 ChIP-Seq and Input Read Mapping 219
6.3.4 ChIP-Seq Peak Calling with MACS3 223
6.3.5 Visualizing ChIP-Seq Enrichment in Genome Browser 226
6.3.6 Visualizing Peaks Distribution 229
6.3.6.1 ChIP-Seq Peaks’ Coverage Plot 230
6.3.6.2 Distribution of Peaks in Transcription Start Site (TSS)
Regions 233
6.3.6.3 Profile of Peaks along Gene Regions 234
6.3.7 Peak Annotation 235
6.3.7.1 Writing Annotations to Files 237
6.3.8 ChIP-Seq Functional Analysis 239
6.3.9 Motif Discovery 243
6.4 SUMMARY 250
REFERENCES 251

Chapter 7   ◾    Targeted Gene Metagenomic Data Analysis 253

7.1. INTRODUCTION TO METAGENOMICS 253
7.2 ANALYSIS WORKFLOW 254
7.2.1 Raw Data Preprocessing 254
7.2.2 Metagenomic Features 255
7.2.2.1 Clustering 255
7.2.2.2 Denoising 256
7.2.3 Taxonomy Assignment 258
7.2.3.1 Basic Local Alignment Search Tool 258
7.2.3.2 VSEARCH 259
7.2.3.3 Ribosomal Database Project 259
7.2.4 Construction of Phylogenetic Trees 260
7.2.5 Microbial Diversity Analysis 261
7.2.5.1 Alpha Diversity Indices 262
7.2.5.2 Beta Diversity 262
x   ◾    Contents

7.3 DATA ANALYSIS WITH QIIME2 263

7.3.1 QIIME2 Input Files 265
7.3.1.1 Importing Sequence Data 265
7.3.1.2 Metadata 269
7.3.2 Demultiplexing 269
7.3.3 Downloading and Preparing the Example Data 271
7.3.3.1 Downloading the Raw Data 271
7.3.3.2 Creating the Sample Metadata File 272
7.3.3.3 Importing Microbiome Yoga Data 274
7.3.4 Raw Data Preprocessing 275
7.3.4.1 Quality Assessment and Quality Control 275
7.3.4.2 Clustering and Denoising 278
7.3.5 Taxonomic Assignment with QIIME2 289
7.3.5.1 Using Alignment-Based Classifiers 289
7.3.5.2 Using Machine Learning Classifiers 291
7.3.6 Construction of Phylogenetic Tree 297
7.3.6.1 De Novo Phylogenetic Tree 297
7.3.6.2 Fragment-Insertion Phylogenetic Tree 298
7.3.7 Alpha and Beta Diversity Analysis 298
7.4 SUMMARY 300
REFERENCES 301

Chapter 8   ◾    Shotgun Metagenomic Data Analysis 303

8.1 INTRODUCTION 303
8.2 SHOTGUN METAGENOMIC ANALYSIS WORKFLOW 305
8.2.1 Data Acquisition 305
8.2.2 Quality Assessment and Processing 305
8.2.3 Removing Host DNA Reads 306
8.2.3.1 Download Human Reference Genome 306
8.2.3.2 Mapping Reads to the Reference Genome 307
8.2.3.3 Converting SAM to BAM Format 307
8.2.3.4 Separating Metagenomic Reads in BAM Files 307
8.2.3.5 Creating Paired-End FASTQ Files from BAM Files 308
8.2.4 Assembly-Free Taxonomic Profiling 310
8.2.4 Assembly of Metagenomes 315
 Contents   ◾    xi

8.2.5 Assembly Evaluation 317

8.2.6 Mapping Reads to the Assemblies 318
8.2.7 Binning 321
8.2.8 Bin Evaluation 323
8.2.9 Prediction of Protein-Coding Region 324
8.3 SUMMARY 325
REFERENCES 326

INDEX, 327
Preface

T he use of next-generation sequencing data analysis is the only analysis that can
make sense of the massive genomic data produced by the high-throughput sequenc-
ing technologies and accumulated in gigabytes and terabytes in our hard drives and cloud
databases. With the presence of computational resources and elegant algorithms for NGS
data analysis, scientists need to know how to master the tools of these analyses to achieve
the goals of their research. Learning NGS data analysis techniques has already become one
of the most important assets that bioinformaticians and biologists must acquire to keep
abreast of the progress in the modern biology and to avail of the genomic technologies and
resources that have become the de facto in bioscience research and applications including
diagnosis, drug and vaccine discovery, medical studies, and the investigations of pathways
that give clues to many biological activities and pathogenicity of diseases.
In the last two decades, the progress of next-generation sequencing has made a strong
positive impact on human life and a forward stride in human civilization. Introduction of
new sequencing technologies revolutionizes the bioscience. As a result, a new field of biol-
ogy called genomics has emerged. Genomics focuses on the composition, structure, func-
tional units, evolution, and manipulation of genomes, and it generates massive amount
of data that need to be ingested and analyzed. As a consequence, bioinformatics has also
emerged as an interdisciplinary field of science to address the specific needs in data acqui-
sition, storage, processing, analysis, and integration of that data into a broad pool to enrich
the genomic research.
This book is designed primarily to be a companion for the researchers and graduate
students who use sequencing data analysis in their research, and it also serves as a text-
book for teachers and students in biology and bioscience. It contains an updated material
in the subject covering most NGS applications and meeting the requirements of a complete
semester course. The reader will find that this book is digging deep in the analysis, pro-
viding both concept and practice to satisfy the exact need of the researchers who seek to
understand and use NGS data reprocessing, genome assembly, variant discovery, gene pro-
filing, epigenetics, and metagenomics. The book does not introduce the analysis pipelines
in a black box as the existing books do, but with the analysis steps, it pervades each topic in
detail to provide the readers with the scientific and technical background that enable them
to conduct the analysis with confidence and understanding.
The book consists of eight chapters. All chapters include real-world worked examples
that demonstrate the steps of the analysis workflow with real data downloadable from the

xiii
xiv   ◾    Preface

databases. Most programs used in this book are open-source Unix/Linux-based programs.
Others can be used in Anaconda environments.
Chapter 1 discusses sequencing data acquisition from NGS technologies and databases,
FASTQ file format, and Phred base call quality. The chapter covers the quality assessment
of the FASTQ and read quality metrics in some detail so that the readers can diagnose
potential problems in raw data and learn how to fix any possible quality problem before
analysis.
Chapter 2 discusses read alignment/mapping to reference genomes. The strategies of
both reference genome indexing algorithms and read mapping algorithms are discussed in
detail with illustrations so that the readers can understand how mapping process works,
the different indexing and alignment algorithms currently used, and which aligners are
good for RNA sequencing applications. The chapter discusses indexing and searching
algorithms like suffix tree, suffix arrays, Burrow-Wheeler Transform (BWT), FM-index,
and hashing, which are the algorithms used by aligners. The chapter then discusses the
mapping process and aligners like BWA, Bowtie, STAR, etc. The SAM/BAM file format
is discussed in detail so that the reader can understand how alignment information are
stored in fields in the SAM/BAM file. Finally, the chapter discusses the manipulation of
alignments in SAM/BAM files using Samtools programs for different purposes, including
SAM to BAM conversion, alignment sorting, indexing BAM files, extracting alignments of
a chromosome or a specific region, filtering and counting alignment, removing duplicate
reads, and generating descriptive statistics.
Chapter 3 discusses de novo genome assembly and de novo assembly algorithms includ-
ing greedy algorithm, overlap-consensus graphs, and de Bruijn graphs. The quality assess-
ment of the assembled genome is discussed through two approaches: statistical approach
and evolutionary approach.
Chapter 4 covers variant calling (SNPs and InDels) in detail. The introduction of this
chapter discusses variants, variant file format (VCF), and the general workflow of the vari-
ant calling. The chapter then discusses both consensus-based variant calling and hap-
lotype-based variant calling and example callers from each group including BCFTools,
FreeBayes, and GATK best practice variant calling pipelines. Finally, the chapter discusses
variant annotation and prioritization and annotation programs including SIFT, SnpEff,
and ANNOVAR.
Chapter 5 discusses RNA-Seq data analysis. The introduction includes RNA-Seq basics
and applications. The chapter then discusses the steps of RNA-Seq analysis workflow,
including data acquisition, read alignment, alignment quality control, quantification,
RNA-Seq data normalization, statistical modeling and differential expression analysis,
using R packages for differential analysis, and visualization of RNA-Seq data.
Chapter 6 covers ChIP-Seq data analysis. It discusses in detail the workflow of ChIP-Seq
data analysis including data acquisition, quality control, read mapping, peak calling, visu-
alizing peak enrichment and peak distribution, peak annotation, peak functional analysis,
and motif discovery.
Chapter 7 discusses targeted gene metagenomic data analysis (amplicon-based microbial
analysis) for environmental and clinical samples. The chapter covers raw data preprocessing
 Preface   ◾    xv

(demultiplexing and quality control), metagenomic feature processing (OTU clustering,

DADA2 denoising, Deblur denoising, and UNOISE2 denoising), taxonomy assignment
and construction of phylogenetic trees, microbial diversity analysis (alpha diversity and
beta diversity), and data analysis with QIIME2.
Chapter 8 discusses the shotgun metagenomic data analysis. The discussion includes
data acquisition, quality control, removal of host DNA, assembly-free taxonomic profil-
ing, assembly evaluation, de novo assembly of metagenomes, binning, bin evaluation, and
prediction of gene transcript and gene annotation.
The computer codes used in the book chapters and their updates will be available at
“https://github.com/hamiddi/ngs”.
Author

Hamid D. Ismail, PhD, earned an MSc and a PhD in computational science at North
Carolina Agricultural and Technical State University (NC A&T), USA, and DVM and BSc
at the University of Khartoum, Sudan. He also earned several professional certifications,
including SAS Advanced Programmer and SQL Expert Programmer. Currently, he is a
postdoc scholar at Michigan Technological University and an adjunct professor of data
science and bioinformatics at NC A&T State University. Dr. Ismail is a bioinformatician,
biologist, data scientist, statistician, and machine learning specialist. He has contributed
widely to the field of bioinformatics by developing bioinformatics tools and methods for
applications of machine learning on genomic data.

xvii
 Chapter 1

Sequencing and
Raw Sequence Data
Quality Control

1.1 NUCLEIC ACIDS

Nucleic acids are the chemical molecules that every living organism must have. They carry
information that directs biological activities in cells and determines the inherited charac-
teristics of the living organism. The two main kinds of nucleic acids are deoxyribonucleic
acid (DNA) and ribonucleic acid (RNA). DNA is the master blueprint for life or the book of
life, and it constitutes the genetic material in prokaryotic and eukaryotic cells and virions.
The RNA is the main genetic material of the RNA viruses, but it is found in other organisms
as molecules transcribed by DNA to play important biological roles such as protein syn-
thesis and gene regulation. The set of the DNA particles in both prokaryotic and eukary-
otic cells is called the genome. RNA is the genome of only some viruses (RNA viruses).
A nucleic acid (DNA/RNA) is a polymer made up of four building blocks called nucleo-
tides. A molecule of the nucleotide consists of (i) a sugar molecule (either deoxyribose in
DNA or ribose in RNA) attached to a phosphate group and (ii) a nitrogen-containing base
called nucleobase. In general, the nucleic acid sequence is made up of four nucleotides dis-
tinguished from one another only by the nitrogen-containing bases (Adenine (A), Cytosine
(C), Guanine (G), and Thymine (T) in the DNA molecule and Adenine (A), Cytosine (C),
Guanine (G), and Uracil (U) in the RNA molecule). Those four nucleobases are divided
into pyrimidine and purine bases. Pyrimidine bases include cytosine, thymine, and uracil;
they are aromatic heterocyclic organic compound with a single ring. Purine bases include
adenine and guanine which have two heterocyclic ring structures. A DNA molecule exists
in the form of two complementary strands (forward and reverse) that wind around each
other forming a double-helix structure. The two strands are held together by hydrogen
bonds formed between the bases (adenine is a base pair of thymine (A/T), and cytosine is a

DOI: 10.1201/9781003355205-1 1
2   ◾    Bioinformatics

base pair of guanine (C/G)) as shown in Figure 1.1. Adenine and thymine form two hydro-
gen bonds (weak bond), while cytosine and guanine form three hydrogen bonds (strong
bond). Those base pairings are specific so that a sequence of a strand is predicted from the
other one. The length of a DNA sequence is given in base pair (bp), kilobase pair (kbp),
or megabase pair (Mbp). The RNA exists in a single strand; however, it sometimes forms
double-stranded secondary structure with itself to perform specific function.
The genome of an organism is the book of life for that organism. It determines the living
aspects and biological activities of cells. A genome contains coding regions known as genes
that carry information for protein synthesis. Genes are transcribed into messenger RNA
(mRNA), which is translated into proteins and the proteins control most of the biological
processes in the living organisms.
A gene consists of coding regions, non-coding regions, and a regulatory region. The cod-
ing regions in the eukaryotic genes are not continuous, but non-coding sequences (called
introns) are found between the coding sequences (called exons). These introns are removed
from the transcribed transcripts before protein translation, leaving only the exons which
form the coding region called the open reading frame (ORF). Each eukaryotic gene has its
own regulatory region that controls its expression. In prokaryotic cells, a group of genes,
called an operon, are regulated by a single regulatory region. The viruses, which fall in
the margin between living organisms and chemical particles, function and replicate only
inside host cells by using the host cells machineries such as ribosomes to create structural
and non-structural proteins of viruses and to replicate to create new virions.

H H O
N

N N
N N H

O Thymine
N N
Adenine H

N
O H

H N N
N N

H Cytosine
N O
N N
Guanine H

FIGURE 1.1 Base pairing and hydrogen bonds between pairs of the DNA nucleotides.
 Sequencing and Raw Sequence Data Quality Control   ◾    3

1.2 SEQUENCING
DNA/RNA sequencing is the determination of the order of the four nucleotides in a nucleic
acid molecule. The recovered order of the nucleotides in a genome of an organism is called
a sequence. Sequencing of the DNA helps scientists to investigate the functions of genes,
roles of mutations in traits and diseases, species, evolutionary relationships between spe-
cies, diagnosis of diseases caused by genetic factors, development of gene therapy, criminal
investigations and legal problems, and more. Since the nucleotides are distinguished by the
bases, the DNA and RNA sequences are represented in bioinformatics by the sequences of
the four-nucleobase single-character symbols (A, C, G, and T for DNA and A, C, G, and
U for RNA).
The attempts to sequence nucleic acid began immediately after the landmark discovery
in 1953 of the double-helix structure of the DNA by James Watson and Francis Crick.
The alanine tRNA was the first nucleic acid sequenced in 1965 by the Nobel prize winner
Robert Holley. Holley used two ribonuclease enzymes to split the tRNA at specific nucleo-
tide positions, and the order of the nucleotides was determined manually [1]. The first DNA
molecule was sequenced in 1972 by Walter Fiers. That DNA molecule was the gene that
codes the coat protein of the bacteriophage MS2, and the sequencing was made by using
enzymes to break the bacteriophage RNA into pieces and separating the fragments with
electrophoresis and chromatography [2]. The sequencing of the alanine tRNA by Robert
Holley and the sequencing of the gene of the bacteriophage MSE coat protein are ones of
the major milestones in the history of genomics and DNA sequencing. They paved the way
for the first-generation sequencing.

1.2.1 First-Generation Sequencing

The early 1970s witnessed the emergence of the first-generation sequencing when the
American biologists Allan M. Maxam and Walter Gilbert developed a chemical method
for sequencing, followed by the English biochemist Frederick Sanger who developed the
chain-terminator method. The Sanger method became the more commonly used first-­
generation sequencing method to this date. Both methods were used in the shotgun
sequencing, which involves breaking genome into DNA fragments and sequencing of the
fragments individually. The genome sequence is then assembled based on the overlaps after
aligning the fragment sequences.
The Maxam–Gilbert sequencing method is based on the chemical modification of DNA
molecules and subsequent cleavage at specific bases. In the Maxam–Gilbert sequencing
method, first, the DNA is denatured (separation of the DNA strands) by heating or helicase
enzyme into single-stranded DNA (ssDNA) molecules. The ssDNA is run in the gel elec-
trophoresis to separate the two DNA strands into two bands. Any one of the bands (strand)
can be cut from the gel and sequenced. In the sequencing step, the solution with ssDNA is
then divided into four reaction tubes labeled A+G, G, C+T, and C. The ssDNA in each tube
is labeled chemically with an isotope and treated with a specific chemical that breaks the
DNA strand at a specific nucleotide according to the tube labels. After the reaction, poly-
acrylamide gel is then used for running the four reactions in four separate lanes (A+G, G,
4   ◾    Bioinformatics

C+T, and C). The order of the nucleotides (A, C, G, and T) in the DNA sequence can then
be solved from the bands on the gel.
On the other hand, the steps of the Sanger sequencing method are similar to that of the
polymerase chain reaction (PCR) including denaturing, primer annealing, and comple-
mentary strand synthesis by polymerase. However, in the Sanger sequencing, the sample
DNA is divided into four reaction tubes labeled ddATP, ddGTP, ddCTP, and ddTTP. In the
four reaction tubes, the four types of deoxynucleotides triphosphates (dATP, dGTP, dCTP,
and dTTP) are added as in the PCR but one of the four radio-labeled dideoxynucleotide
triphosphates (ddATP, ddGTP, ddCTP, or ddTTP) is also added to the reactions, as labeled,
to terminate the DNA synthesis at certain positions of known nucleotides. The synthesis
termination results in DNA fragments of varying lengths ending with the labeled ddNTPs.
Those fragments are then separated by size using gel electrophoresis on a denaturing poly-
acrylamide-urea gel with each of the four reactions running in a separate lane labeled A,
T, G, and C. The DNA fragments will be separated by lengths; the smaller fragments will
move faster in the gel. The DNA bands are then graphed by autoradiography, and the order
of the nucleotide bases on the DNA sequence can be directly read from the X-ray film or
the gel image.

1.2.2 Next-Generation Sequencing

The next-generation sequencing (NGS) was invented a few decades after the invention of
the first-generation sequencing. Unlike the first-generation sequencing, NGS produces
massive number of sequences from a single sample in a short period of time, with lower
costs, and it can process multiple samples simultaneously. Millions to billions of DNA
nucleotides are sequenced in parallel, yielding substantially massive sequences. With the
NGS, millions of prokaryotic, eukaryotic, and viral genomes were sequenced. Rather
than chain termination, the NGS uses library or fragmented DNA to solve the order of
the nucleotide in a targeted sequence. The NGS is used in many applications including
the sequencing of the whole genome, whole transcriptome, targeted genes or transcripts,
sequencing of the genomic regions where the epigenetic modifications or protein interac-
tions take place. Hence, the NGS can be used for genome assembly, mutation or variant
discovery, gene expression studies, epigenetics, and metagenomics. Those applications are
discussed in detail in the next chapters.
After DNA or RNA sample collection, the step of the NGS process is the library prepa-
ration in which the sequencing libraries are constructed for the DNA or RNA sample of
interest. The RNA is converted into complementary DNA (cDNA) before library prepara-
tion. The library preparation involves breaking the DNA into small fragments (fragmen-
tation step) using sonication or enzymes. The size of the fragment can be adjusted to a
specific length or range. The fragmentation is followed by repairing or blunting the ends of
the fragments which have unpaired or overhanging nucleotides (end-repair step). The next
step is the ligation of the adaptors to the ends of the DNA fragments. The adaptors are arti-
ficially synthesized sequences that include certain parts to serve specific purposes. The free
end of a ligated adaptor is made of an anchoring sequence that can attach to the surface of
the flow cell slide where sequencing takes place. The adaptor also includes universal primers
 Sequencing and Raw Sequence Data Quality Control   ◾    5

for PCR DNA strand synthesis and barcode sequence for indexing the sample DNA. This
allows multiple samples (multiplexing) to be sequenced in a single run; the DNA fragments
of each sample will have a unique barcode. Later, after sequencing, the sample sequences
can be separated in the analysis by demultiplexing. In the sequencing of some application
like gene expression (RNA-Seq) and epigenetics, an enrichment step is usually included
to amplify or to separate only the targeted sequences. In RNA application, enrichment is
performed to separate mRNA from the other types of RNA. In epigenetics, the genomic
regions where protein interaction is taking place can also be enriched. The enrichment is
usually performed with PCR, but there are other means as well. The library preparation of
the DNA/RNA is similar for all NGS technologies but the sequencing process is different
from one to another. The sequences produced by the NGS technologies range between 75
and 400 base pairs (bp) in length. These sequences are called short reads. In general, short-
read sequencing (SRS) can either be a single-end sequencing, which sequences the forward
strand only, or paired-end sequencing, which sequences both forward and reverse strands.
The latter reduces the chance of making basal error in the resulted sequence. The DNA or
RNA reads consist of the four nucleobase characters A, C, G, and T. However, the sequence
may also include N for an unresolved base.

1.2.2.1 Roche 454 Technology

Roche 454 pioneered the NGS, when 454 was used to sequence the whole genome of
Mycoplasma genitalium in 2006 [3]. The 454 technology uses pyrosequencing, which
depends on the sequential addition and incorporation of nucleotides in the DNA template.
The signal of the added nucleotide is quantitated by conversion of released pyrophosphate
into a light signal in the real time. The pyrosequencing is based on a series of enzymatic
reactions that lead to the DNA synthesis and release of the inorganic pyrophosphate every
time a nucleotide is incorporated by polymerase in the DNA chain. The density of the light
generated by the reactions can be detected by a charge-coupled device camera. The order
of the nucleotides in the DNA template is determined by quantitating the light density.
In the pyrosequencing, the DNA is fragmented and denatured into ssDNA. Two adap-
tors (A and B) are ligated to both ends of the fragments. Beads of soluble particles with
single-stranded primers complementing adaptor A are added to the reactions. The adaptor
A attached to ssDNA template complements the bead primers, which initiate the synthe-
sis of the complementary strand. This step can be repeated several times for enrichment
(PCR). Then, the beads with the ssDNA templates are placed into wells where sequencing
takes place. A primer is added to complement the adaptor B and to initiate the addition of
new nucleotides to the complementary strand. However, this time, known nucleotides are
added. Every time a nucleotide is incorporated into the complementary strand, a hydroxyl
group of the last nucleotide reacts with the alpha phosphate of the incorporated nucleo-
tide releasing a two-phosphate compound called the inorganic pyrophosphate (PPI). The
PPI contains a high amount of energy that converts the adenosine monophosphate (AMP)
into adenosine triphosphate (ATP) with the help of ammonium persulfate (APS) and sul-
furylase which are added to the reaction. Finally, luciferin and luciferase are added to
the ATP so the luciferin forms light. Every time a nucleotide is added, a light is emitted
6   ◾    Bioinformatics

and captured by a sensor or a camera. Figure 1.2 shows the steps of pyrosequencing after
adding the DNA fragments to the wells, where sequencing takes place by incorporating a
known nucleotide to the growing complementary strand each time and releasing of PPI
that is translated into a light signal, which is detected by a camera.

1.2.2.2 Ion Torrent Technology

The Ion Torrent sequencers are manufactured by Thermo Fisher Scientific. The Ion Torrent
sequencing method is similar to the Roche 454 sequencing in the first steps: fragmenta-
tion, DNA denaturing, adaptor ligation, and the use of beads in microwells with primers.
However, instead of solving the nucleotide order of the DNA template by capturing the
light released by inorganic pyrophosphate as in Roche 454 sequencing, the Ion Torrent
method depends on the pH signal of the hydrogen ions released every time a nucleotide
is incorporated in the DNA template. The microwells are in a microchip provided with
the so-called ion sensitive layer (pH sensor). This layer is able to detect the hydrogen ions
(protons). There are four solutions for the four kinds of nucleotides (dATP, dCTP, dGTP,
and dTTP). During the Ion Torrent sequencing, the four nucleotide solutions are added
sequentially. If the nucleotide is incorporated, a hydrogen ion (H+) will be released chang-
ing the pH of the solution and a signal will be generated; otherwise, no change in the pH
or a signal (Figure 1.3). The base call is based on this signal. Each time, the microwells are
washed to remove the free nucleotides before adding a new nucleotide solution.

1.2.2.3 AB SOLiD Technology

The Applied Biosystem SOLiD is based on the sequential ligation with dye-labeled oligos.
SOLiD stands for Sequencing by Oligonucleotide Ligation and Detection. This technology
was developed by Fisher Scientific. It depends on the DNA ligase that has the ability to
incorporate nucleotides in the DNA template. The DNA library preparation is similar to

FIGURE 1.2 Pyrosequencing and base calling.

 Sequencing and Raw Sequence Data Quality Control   ◾    7

FIGURE 1.3 Ion Torrent sequencing.

what has been discussed above. In the library preparation step, the DNA is broken down
into fragments. The ends of these fragments are repaired and then adaptor sequences are
ligated to the ends. As Roche 454 and Ion Torrent sequencing, beads are also used and the
fragments attached to the beads are amplified using PCR. After the amplification, specific
primers are used to synthesize strands complementing the ssDNA templates and provid-
ing a free 5′ phosphate group (instead of 3′ hydroxyl group) that can ligate to one of a set
of fluorescently labeled probes. A probe is an 8-mer oligonucleotide with a fluorescent dye
at the 5′-end and a hydroxyl group at the 3′-end. The first two nucleotides of the probe spe-
cifically complement the nucleotides on the sequenced fragments. The next three nucleo-
tides are universal that can bind to any one of the four nucleotides. The remaining three
nucleotides of the probe (at the 5′-end) are also universal but are fluorescently labeled and
they are able to be cleaved during the sequencing leaving only the other five nucleotides
binding to the template strand. The set of probes consists of a combination of 16 possible
­2-nucleotides that can ligate to the primer sequence on the fragments. Every time two
specific ­nucleotides ligate, the last three nucleotides of the probe are cleaved and the fluo-
rescence specific to the ligated probe is emitted, captured, and translated into the corre-
sponding two bases. This step is followed by removing the fluorescently labeled nucleotide
and regenerating the 5′ phosphate group. The process (ligation, detection, and cleavage)
is repeated multiple times for extending the complementary strand. The strand extension
products are then removed, and a new primer is used for the second round of ligation,
detection, and cleavage. The cycles are then repeated and every time a new primer is used
for the remaining of the DNA fragments.

1.2.2.4 Illumina Technology

Illumina uses sequencing by synthesis (SBS) approach (Figure 1.4), in which it uses four
fluorescently labeled nucleotides to sequence the DNA fragments on the flow cell surface in
multiple cycles. In each cycle, a single fluorescently labeled nucleotide (dATP, dCTP, dGTP,
or dTTP) is added to the DNA growing chain. These labeled nucleotides serve as chain
terminators that stop chain extension and also the incorporation of the labeled nucleotide
in the chain causes color emission or a signal of a specific nucleotide. The signal of the
incorporated nucleotide is then captured by an imaging camera to identify that nucleotide
(or base). The terminator nucleotide is then cleaved to allow the incorporation of the next
8   ◾    Bioinformatics

nucleotide. The base call is based on the intensity of the signals during chain synthesis and
labeled nucleotide incorporation.
The DNA library preparation, as described above, includes the DNA fragmentation,
DNA end repair, and adaptor ligation. Sequencing is carried out on the solid surface of a
flow cell divided into lanes. On the surface, there are two types of oligonucleotides that
complement the anchor sequence in the adaptors attached to the DNA template. When the
DNA fragments are added to the cell flow, the anchor sequences in the fragments anneal
to (complement) the surface oligonucleotides forming bridges. The oligos also act as prim-
ers that initiate the synthesis of complementary strands generating clusters of the DNA
fragments. The sequencing step begins by denaturing the DNA strands and adding fluo-
rescently labeled nucleotides. Only one nucleotide is incorporated in the DNA template at
a time. After the addition of each nucleotide, the clusters are excited by a light source and
a signal is emitted. The signal intensity determines the base call.

1.2.3 Third-Generation Sequencing

The third-generation sequencing (TGS) is relatively new sequencing technology developed
as a result of the need for long reads to improve the accuracy and resolution of sequencing

FIGURE 1.4 Illumina sequencing.

 Sequencing and Raw Sequence Data Quality Control   ◾    9

applications like the de novo genome assembly, variant discovery, and epigenetics. Usually,
genomes have long repeated sequences ranging from hundreds to thousands of bases that
are hard to cover with short reads produced by the NGS. The foundation for TGS was
emerged in 2003, when the DNA polymerase was used to obtain a sequence of 5 bp from
a single DNA molecule by using fluorescent microscopy [4]. The single-molecule sequenc-
ing (SMS) then evolved to include (i) direct imaging of individual DNA molecules using
advanced microscopy techniques and (ii) nanopore sequencing technologies in which a
single molecule of DNA is threaded through a nanopore and molecule bases are detected
as they pass through the nanopore. Although TGS provides long reads (from a few hun-
dred to thousands of base pairs), that may come at the expense of the accuracy. However,
lately, the accuracy of the TGS has been greatly improved. The TGS provides long reads
that can enhance de novo assembly and enable direct detection of haplotypes and higher
consensus accuracy for better variant discovery. In general, there are two TGS technolo-
gies that are currently available: (i) Pacific Bioscience (PacBio) single-molecule real-time
(SMRT) sequencing and (ii) Oxford Nanopore Technologies (ONTs).

1.2.3.1 PacBio Technology

The Pacific Biosciences (PacBio) sequencing can provide long reads that range between
500 and 50,000 bp. The PacBio sequencing has been improved since it made debut in 2011.
The underlying technology of the PacBio is based on the SMRT sequencing, in which a
single DNA molecule is sequenced and the base calling is given in the real time, while the
sequencing is in progress [5]. The sequencing steps include fragmentation and ligation of
adaptors to the DNA template for library generation. Special loop adaptors are ligated to
the ssDNA produced from the double-stranded DNA (dsDNA). The loop adaptors link
both strands forming structures called linear DNA SMRTbells. The sequencing takes place
on nano wells on a flow cell. The nano wells are made of silicon dioxide chips called zero-
mode waveguides (ZMWs) [6]. A cell contains thousands of ZMWs. A ZMW is around
70 nm in diameter and 100 nm in depth, and it allows laser light to come through the bot-
tom to excite the fluorescent dye. A DNA polymerase is attached to the bottom of the nano
well. When a DNA single fragment is added to the well, the DNA polymerase is attached
to it. The polymerase has the strand displacement capability that converts the DNA
SMRTbells into circular structure called circular DNA SMRTbell (Figure 1.5). Then, the
DNA polymerase continues adding nucleotides to form a complementary strand for both
forward and reverse strands. PacBio uses SBS approach. Four fluorescently labeled nucleo-
tides (dNTPs) are added to the reactions. The dNTPs are fluorescently labeled by attach-
ing the fluorescent dyes to phosphate chain of the nucleotides. Each time a fluorescently
labeled nucleotide is incorporated, a fluorescent dye is cleaved from the growing nucleic
acid chain before the next nucleotide is added. The fluorescence is then excited by the light
coming through the bottom of the well and detected in the real time. The real-time identi-
fication of the incorporated labeled nucleotides allows the base call. This process of adding
nucleotides and fluorescence detection continues until the entire fragment is sequenced.
This pass can be repeated different times to generate more accurate reads by the circular
consensus sequences (CCS). Ten passes produce reads with 99.9% accuracy. These reads are
10   ◾    Bioinformatics

FIGURE 1.5 Pacific Bioscience (PacBio) sequencing.

called highly accurate long sequencing reads (HiFi reads). The ­sequencing is ­carried out in
parallel on thousands of nanophotonic wells, generating long reads. PacBio can produce
long reads with 99% accuracy (>Q20) and uniform coverage.

1.2.3.2 Oxford Nanopore Technology

The Oxford Nanopore Technology (ONT) was announced in 2012, but it was made com-
mercially available in 2015 as MinION technology. ONT uses a flow cell to allow mas-
sive parallel sequencing. Their flow cell is made of an electrical resistant membrane that
contains thousands of tiny pores, each with a diameter of one nanometer. The sequencing
principle is based on applying a steady current to the nanopores. The magnitude of current
depends on the size and shape of the nanopore. During sequencing, a ssDNA is allowed to
pass through the nanopore like a needle and thread so that each nucleotide passes at a time.
When a nucleotide passes the opening of the pore, the voltage magnitude is changed. We
can use known DNA sequence to capture the unique electrical change for each nucleotide
and then allow the DNA molecule that we need to sequence to go through the nanopore
opening. The order of the nucleotides on the DNA sequence is determined by matching
the patterns. ONT has a higher error rate because the DNA fragments tend to pass faster
through the nanopores that may generate errors.
 Sequencing and Raw Sequence Data Quality Control   ◾    11

1.3 SEQUENCING DEPTH AND READ QUALITY

1.3.1 Sequencing Depth
The biological results and interpretation of sequencing data for the different sequencing
applications are greatly affected by the number of sequenced reads that cover the genomic
regions. Usually, multiple sequences overlap over certain regions of the genome. The
sequencing depth measures the average read abundance and it is calculated as the number
of bases of all sequenced short reads that match a genome divided by the length of that
genome if the genome size is known. If the reads are equal in length, the sequencing depth
is calculated as

read length ( bp ) × number of reads

Coverage = (1.1)
genome size ( bp )

If the reads are not equal in length, the coverage is calculated as

∑
n
length of read i
Coverage = i =1
(1.2)
genome size ( bp )
where n is the number of sequenced reads.
The sequencing coverage is expressed as the number of times the genome (e.g., 1X, 2X,
20X,…, etc.).
The sequencing depth affects the genomic assembly completeness, accuracy of de novo
assembly and reference-guided assembly, number of detected genes, gene expression lev-
els in RNA-Seq, variant calling, genotyping in the whole genome sequencing, microbial
identification and diversity analysis in metagenomics, and identification of protein–DNA
interaction in epigenetics. Therefore, it is important to investigate sequencing depth before
sequence analysis. The higher the number of times that bases are sequenced, the better the
quality of the data.

1.3.2 Base Call Quality

We have already discussed the different sequencing technologies which have different
sequencing approaches. However, at the end, each of these technologies attempts to infer
the order of the nucleic acid studied. The process of inferring the base (A, C, G, or T) at
specific position of the sequenced DNA fragment during the sequencing process is called
base calling. The sequencing platforms are not perfect and errors may occur during the
sequencing process when the machine tries to infer a base from each measured signal. For
all platforms, the strength of the signals and other characteristic features are measured
and interpreted by the base caller software. Errors affect the sequence data directly and
make them less reliable. Therefore, it is critical to know the probability of such errors so
that users can know the quality of their sequence data and can figure out how to deal with
those quality errors. Most platforms are equipped with base calling programs that assign
12   ◾    Bioinformatics

a Phred quality score to measure the accuracy of each base called. The Phred quality score
(Q-score) transforms the probability of calling a base wrongly into an integer score that is
easy to interpret. The Phred score is defined as

p = 10− Q /10 (1.3)

Q = −10log10 ( p ) (1.4)

where p is the probability of the base call being wrong as estimated by the caller software.
The Phred quality scores are encoded using ASCII single characters. All ASCII char-
acters have a decimal number associated with them. However, since the first 32 ASCII
characters are non-printable and the integer 33, which is the decimal number for the
exclamation mark ASCII character “!”, the Q=0 is the exclamation mark and the encod-
ing that begins with “!” as zero is called Phred+33 encoding. Illumina 1.8 and later ver-
sions use this Phred+33 encoding (Q33) to encode the base call quality in FASTQ files. The
older Illumina versions (e.g., Solexa) used Phred+64 encoding, in which the character “@”,
whose decimal number is 64, corresponds to Q=0. Table 1.1 shows the Phred quality score
(Q), corresponding probability (P), and the decimal number and ASCII code. For instance,
when the probability of calling a base is 0.1, the Phred score will be 10 (Q=10), but instead
of giving the number 10, that quality score is encoded as the plus sign “+”.
Higher Q scores indicate a smaller probability of error and lower Q scores indicate
low qualities of the base called which is more likely that the base was called wrongly. For
instance, a quality score of 20 indicates the chance of making an error rate (1 error) in 100,
corresponding to 99% call accuracy. In general, the Q-score of 30 is considered a benchmark

TABLE 1.1 Phred Quality Score and ASCII_BASE 33 (Q33)

Q P ASCII Q p ASCII Q p ASCII
0 1.00000 33 ! 15 0.03162 48 0 30 0.00100 63 ?
1 0.79433 34 “ 16 0.02512 49 1 31 0.00079 64 @
2 0.63096 35 # 17 0.01995 50 2 32 0.00063 65 A
3 0.50119 36 $ 18 0.01585 51 3 33 0.00050 66 B
4 0.39811 37 % 19 0.01259 52 4 34 0.00040 67 C
5 0.31623 38 & 20 0.01000 53 5 35 0.00032 68 D
6 0.25119 39 ‘ 21 0.00794 54 6 36 0.00025 69 E
7 0.19953 40 ( 22 0.00631 55 7 37 0.00020 70 F
8 0.15849 41 ) 23 0.00501 56 8 38 0.00016 71 G
9 0.12589 42 * 24 0.00398 57 9 39 0.00013 72 H
10 0.10000 43 + 25 0.00316 58 : 40 0.00010 73 I
11 0.07943 44 , 26 0.00251 59 ; 41 0.00008 74 J
12 0.06310 45 - 27 0.00200 60 < 42 0.00006 75 K
13 0.05012 46 . 28 0.00158 61 = 43 0.00005 76 L
14 0.03981 47 / 29 0.00126 62 > 44 0.00004 77 M
 Sequencing and Raw Sequence Data Quality Control   ◾    13

TABLE 1.2 Phred Quality Score and Error Probability and Base Call Accuracy
Q Error Probability Base Call Accuracy (%) Interpretation
10 0.1 90 1 error in 10 calls
20 0.01 99 1 error in 100 calls
30 0.001 99.9 1 error in 1,000 calls
40 0.0001 99.99 1 error in 10,000 calls
50 0.00001 99.999 1 error in 100,000 calls
60 0.000001 99.9999 1 error in 1000,000 calls

for good quality in the high-throughput sequencing (HTS). Table 1.2 shows some of the Q
scores and corresponding error probability, base call accuracy, and interpretation.

1.4 FASTQ FILES

The sequencing technologies like Illumina are provided with the Real-Time Analysis (RTA)
software that stores individual base call data in intermediate files called BCL files. When
the sequencing run completes, these BCL files are filtered, demultiplexed if the samples are
multiplexed, and then converted into a sequence file format called FASTQ. There will be a
single FASTQ file for each sample for a single-end run and two FASTQ files (R1 and R2)
for each sample for a paired-end run: R1 file for forward reads and R2 for reverse reads.
The FASTQ files are usually compressed and they may have the file extension “*.fastq.gz”.
A FASTQ [7] file is a human-readable file format that has become de facto standard for
storing the output of most HTS technologies. A FASTQ file consists of a number of records,
with each record having four lines of data as shown in Figure 1.6.
The first line of each record of a FASTQ file begins with the “@” symbol and this line is
called the read identifier since it identifies the sequence (read). A typical FASTQ identifier
line of the reads generated by an illumine instrument looks as follows:

@<instrument>:<run num>:<flowcell ID>:<lane>:<tile>:<x>:<y>:<UMI>

<read>:<filtered>:<control num>:<index>

Table 1.3 describes the elements of the Illumina FASTQ identifier line and Figure 1.6
shows an example FASTQ file with three read records. The sequence observed in the index
sequence (part of the adaptor) is written to the FASTQ header in place of the sample num-
ber. This information can be useful for troubleshooting and demultiplexing. However,
these metadata elements may be altered or replaced by other elements especially when they
are submitted to a database or altered by users.
The second line of the FASTQ file contains the bases inferred by the sequencer. The
bases include A, C, G, and T for Adenine, Cytosine, Guanine, and Thymine, respectively.
The character N may be included if the base in a position is ambiguous (was not deter-
mined due to a sequencing fault).
The third line starts with a plus sign “+”, and it may contain other additional metadata
or the same identifier line elements.
14   ◾    Bioinformatics

TABLE 1.3 Illumina FASTQ Identifier Line Elements [8]

Identifier Line Element Description
@ The beginning of the read identifier line
<instrument> Instrument ID or sequence ID
<run num> The number of the run on the instrument
<flowcell ID> The flow cell ID
<lane> The number of lane where the read was sequenced
<tile> The number of the tile where the read was sequenced
<x> The X-coordinate of the DNA cluster
<y> The Y-coordinate of the DNA cluster
<UMI> Only if a unique molecular identifier (UMI) is used
<read> The read number (1 for single read or 2 for paired end)
<filtered> Y if the read passed the filter and N if didn’t pass
<control num> 0 (none of the control bits are on) or an even number
<index> The sample number or read index

FIGURE 1.6 A FASTQ file format showing three records.

The fourth line of the FASTQ file contains the ASCII-coded string that represents the
per base Phred quality scores. The numeric value of each ASCII character corresponds to
the quality score of a base in the sequence line.
Researchers usually acquire raw sequencing data for their own research from a sequenc-
ing instrument. Raw sequencing data can also be downloaded from a database, where sci-
entists and research institutions deposit their raw data and make it available for public. In
either case, the raw sequencing data is usually obtained in FASTQ files. The NCBI SRA
database is one of the largest databases of raw data for hundreds of species. The FASTQ
files are stored in Sequence Read Archive (SRA) format, and they can be downloaded and
extracted using SRA-toolkit [9], which is a collection of programs developed by the NCBI
and can be downloaded and installed by the instructions available at “https://trace.ncbi.
nlm.nih.gov/Traces/sra/sra.cgi”.
For the purpose of demonstration, we will download raw data from the NCBI SRA data-
base. We will use a single-end FASTQ file with the run ID “SRR030834”, whose size is
3.5G. The FASTQ file contains reads sequenced from an ancient hair tuft of 4000-year-old
male individual from an extinct Saqqaq Palaeo-Eskimo, excavated directly from culturally
deposited permafrozen sediments at Qeqertasussuk, Greenland. To keep file organized, you
can create the directory “fastqs” and then download the FASTQ file using “fasterq-dump”
 Sequencing and Raw Sequence Data Quality Control   ◾    15

command (make sure that you have installed the SRA-toolkit on your computer and it is
on the path):

mkdir fastqs
cd fastqs
mkdir single
cd single
fasterq-dump --verbose SRR030834

As shown in Figure 1.7, the FASTQ file “SRR030834.fastq” has been downloaded to the
directory, and we will use that file to show how to use some Linux commands to perform
some operations with that file.
FASTQ files may contain up to millions of entries, and their sizes can be several mega-
bytes or gigabytes, which often make them too large to open in a normal text editor. In
general, no need to open a FASTQ file unless it is necessary for troubleshooting or out of
curiosity. To display a large FASTQ file, we can use some Unix or Linux commands such
as “less” or “more” to display very large text file page by page or “cat” to display the content
of the file.

less SRR030834.fastq
more SRR030834.fastq
cat SRR030834.fastq

If a FASTQ file name ends with the “.gz” extension, that means the file is compressed with
“gzip” program. In this case, instead of “less”, “more”, and “cat” commands, use “zless”,
“zmore”, and “zcat” commands, respectively, without decompressing the files.
We can also use “head” and “tail” to display the first lines and last lines, respectively.
The following command will display the first 15 lines of the file:

head -15 SRR030834.fastq

If a FASTQ file is large, we can compress it with the “gzip” program to reduce its size more
than three times. Compressing the “SRR030834.fastq” file with gzip will reduce its size to
less than one gigabyte.

gzip SRR030834.fastq

The file name will become “SRR030834.fastq.gz”.

FIGURE 1.7 Downloading a FASTQ file from the NCBI SRA database.
16   ◾    Bioinformatics

Use “gzip -d” to decompress a compressed file.

gzip -d SRR030834.fastq.gz

If you need to know the number of records in a FASTQ file, you can use a combination of
“cat” or “zcat” and “wc -l”, which counts the number of lines in a text file. Remember that
a record in a FASTQ file has 4 lines. We can use the Unix pipe symbol “|” to transfer the
output of the “cat” command to the “wc -l” command. The following command line will
count the number of records stored in the FASTQ files:

cat SRR030834.fastq | echo $((`wc -l`/4))

If we need to display the file name and read count for multiple files, with the “.fastq” file
name extension, in a directory, we can use the following script:

for filename in *.fastq;

do
echo -e “$filename\t `cat $filename | wc -l | awk ‘{print $1 /
4}’`”
done

To display a FASTQ file in a tabular format, you can use the “cat” command and then use
the Unix pipe to transfer the output to the “paste” command, which converts the four lines
of the FASTQ records into tabular format.

cat SRR030834.fastq | paste - - - - > SRR030834_tab.txt

The command will store the new tabular file in a new file “SRR030834_tab.txt”. You can
open this file in any spreadsheet, or you can display it as follows:

less -S SRR030834_tab.txt

Creating a tabular file from a FASTQ file will help us to perform several operations such as
sorting of the entries, filtering out the duplicate reads, extracting read IDs, sequences, or
quality scores, and creating a FASTA file. We expect that the format of the identifier lines
of a FASTQ file is consistent. If you display “SRR030834_tab.txt”, you will notice that some
of the identifier line fields are separated by spaces, and if we consider the space as a column
separator, the IDs will be in the first column and the sequence will be in the fourth column.
However, this column order may be different in tabular files extracted from other FASTQ
files. Assume that we wish to extract only the IDs and sequences from “SRR030834_tab.
txt” in a separate text file, then we can use the “awk” command as follows:

awk ‘{print $1 “\t” $4}’ SRR030834_tab.txt > SRR030834_seq.txt

 Sequencing and Raw Sequence Data Quality Control   ◾    17

The “awk” command extracts the first column and fourth column from “SRR030834_tab.
txt” and prints the two columns separated by a tab. The output is directed to a new text file
“SRR030834_seq.txt” (Figure 1.8).
Linux commands allow us to do multi-step operations. Assume that we want to create a
FASTA file from the FASTQ file; we can do that in multiple steps. First, we need to extract
both IDs and sequences in a file as we did above, then we can remove “@” symbol leaving
only the IDs, then we need to add “>” in the beginning of each line with no space between
the “>” and the IDs, and finally, we separate the two columns, forming the definition line
(defline) of FASTA and the sequence, store them in a file, and delete the temporary files.

cat SRR030834.fastq | paste - - - - \

  > SRR030834_tab.tmp
awk ‘{print $1 “\t” $4}’ SRR030834_tab.tmp \
| sed ‘s/@//g’ > SRR030834_seq.tmp
sed -i ‘s/^/>/’ SRR030834_seq.tmp
awk ‘{print $1, “\n” $2}’ SRR030834_seq.tmp \
> SRR030834.fasta
rm *.tmp

In the FASTA format, as shown in Figure 1.9, each entry contains a definition line and a
sequence. The defline begins with “>” and can contain an identifier immediately after “>”
(no whitespace in between).

FIGURE 1.8 Extracting IDs and sequence of a FASTQ file.

FIGURE 1.9 Extracting FASTA sequence from the FASTQ file.

18   ◾    Bioinformatics

1.5 FASTQ READ QUALITY ASSESSMENT

The quality control for obtaining good sequencing data begins by isolation of pure nucleic
acid from the samples of interest. After isolation, the nucleic acid quality is usually assessed
before sequencing. The nucleic acid must be pure and have sufficient concentration as rec-
ommended by the kits used for library preparation. If the Nanodrop is used, the purity of
the nucleic acid is measured by the ratio of light absorbances at 260 and 280 nm. A ratio of
~1.8 is generally accepted as pure for DNA and a ratio of ~2.0 is generally accepted as pure
for RNA. Similarly, absorbance at 230 nm is usually due to other contamination. For pure
nucleic acid, the 260/230 value is often higher than the respective 260/280 value. The purity
of the RNA is better to be measured by the Bioanalyzer which provides the RNA integrity
(RIN). The RIN ranges from 1 to 10, where 10 is the best RNA integrity or non-degrading
RNA. Most commercially available RNA library preparation kits require an RNA of a RIN
value greater than 7 (RIN>7) for proper library construction.
Another quality control check point is performed after library preparation and before
sequencing. The quality of the nucleic acid libraries can also be assessed to ensure that
there are no remaining adaptor primers that may cause adaptor dimers in the sequenced
DNA fragments. The adaptor dimers are DNA produced from complete adaptor sequences
that can bind and cluster on the flow cell and generate contaminating reads, which nega-
tively impact the quality of sequencing data.
The quality assurances in the steps of nucleic acid isolation and removing of adaptor
primers after library preparation help in producing high-quality reads. However, errors
can also be generated during the sequencing. In Illumina instruments, errors in base call
may be made due to failure to terminate the synthesis (polymerase remains attached),
forming leading strands that are longer than normal, or due to failure to reverse synthesis
termination in the washing cycle, forming lagging strands. The base call software converts
fluorescence signals into actual sequence data with quality scores, which will be stored in
FASTQ files.
A post-sequencing quality check must be carried out to assess the read quality to ensure
that the sequence data looks good and there are no problems or biases that lead to inaccu-
rate or misleading results. FastQC [10] is the most popular program for assessing the qual-
ity of the sequencing data produced by high-throughput instruments. FastQC provides a
simple way to assess the quality of the raw data and generates summary and simple graphi-
cal reports that we can use to have a clear idea about the overall quality of the raw data and
any potential quality problem. FastQC can be downloaded for all platforms from “https://
www.bioinformatics.babraham.ac.uk/projects/fastqc/”. You can use the following steps to
install the latest FastQC version, as of this time, on Linux (Ubuntu):
Use “wget” command to download the current version; the current version is v0.11.9,
which may change in the feature.

wget https://www.bioinformatics.babraham.ac.uk/projects/fastqc/
fastqc_v0.11.9.zip

Decompress the zipped file with unzip:

 Sequencing and Raw Sequence Data Quality Control   ◾    19

unzip fastqc_v0.11.9.zip

A directory named “FastQC” containing the program files will be created.

Change to that directory with:

cd FastQC

This directory contains several files but we will focus only on “fastqc” file, which is the pro-
gram command that we will use and “Configuration” directory which contains three files
“adapter_list.txt”, “contaminant_list.txt”, and “limits.txt”, which can be used to configure
the QC report.
Use the Linux command “chmod” to make “fastqc” file executable:

chmod 755 fastqc

You may need to add the program to the Linux path so that you can use it from any direc-
tory. To do that, open the file “~/.bashrc” with any text editor and add the program path to
the end of the file, save the file and exit.

export PATH=”/home/mypath/FastQC”:$PATH

You may need to replace “/home/mypath” in the above statement with the path of your
downloaded FastQC.
You need to restart the terminal or you can use “source” command to make that change
active.

source ~/.bashrc

Finally, test the FastQC from any directory with the following:

fasqc -h

If the path was set correctly, you will see the program help screen.
On Linux, the FastQC program can also be installed simply by running “sudo apt-get
install fastqc”. Once it has been installed, you do not need any further configuration as we
did above. Although any one of the two installation methods works, in the first method,
you will know where you download the program and you can reach the “Configuration”
directory easily to make changes to any of the three configuration files stored there.
The FastQC program can be run in one of two modes: either as (i) an interactive graphi-
cal user interface (GUI) in which we can load FASTQ files and view their results or (ii) as a
non-interactive mode on a command line where we can specify the FASTQ files and then it
will generate an HTML report for each file. However, if we run it non-interactively, to display
the report, we will need to open the HTML files on an Internet browser. The command-line
mode allows FastQC to be run as part of an analysis pipeline. FastQC uses a FASTQ file as
20   ◾    Bioinformatics

an input and generates quality assessment reports including per base sequence quality, per
tile sequence quality, per sequence quality scores, per base sequence content, per sequence
GC content, per base N content, sequence length distribution, sequence duplication levels,
overrepresented sequences, adaptor content, and k-mer content. FastQC supports all vari-
ants of FASTQ formats and gzip-compressed FASTQ files.
We will download some public single-end FASTQ files from an NCBI BioProject with
an accession “PRJNA176149” for practicing purpose. The SRA files of this project contain
genomic single-end reads of Escherichia coli str. K-12. To keep the files organized, we can
create the directory “ecoli” using “mkdir ecoli” and then move it inside this directory “cd
ecoli” and save the following IDs (each in a line) in a text file with the file name “ids.txt”
using any text editor:

SRR653520
SRR653521
SRR576933
SRR576934
SRR576935
SRR576936
SRR576937
SRR576938

Then, run the following script to create the subdirectory “fastQC” and to download the
FASTQ files associated with the IDs stored in the “ids.txt” file into the directory:

mkdir fastQC
while read f;
do
fasterq-dump \
--outdir fastQC “$f” \
--progress \
--threads 4
done < ids.txt

Once the raw FASTQ files have been downloaded, we can use the command “ls -lh fastQC”
to display the file names as shown in Figure 1.10.

FIGURE 1.10 The names of the downloaded FASTQ files.

 Sequencing and Raw Sequence Data Quality Control   ◾    21

On Linux terminal, we can use FastQC non-interactively and later we will display the
generated reports on an Internet browser. But before running FastQC, it is important to
know about the “limits.txt” file in the “Configuration” directory. This file contains the
default values for the FastQC options, and we can use it to determine which report to gen-
erate. Use a text editor of your choice to open that file and study its content. In most cases,
no change is needed. At this point, we will change only

kmer ignore 1  to    kmer ignore 0

Then, save the file and exit. This change is necessary to include k-mer report when we run
the program.
The following is a simple syntax for running the FastQC program non-interactively on
the command line:

fastqc seqfile1 seqfile2 .. seqfileN

The input can be a single FASTQ file name or multiple file names separated by whitespaces.
The FastQC program has several options that can be displayed using the following
command:

fastqc --help

Since we have downloaded the eight E. coli raw FASTQ files above and stored them in the
“fastQC” directory, we can either run the program for each file or provide all file names
as input as shown in the above syntax. However, the efficient way is to use the bash com-
mands if we are using a Linux/Unix platform. The following bash script creates a directory
“qc”, changes to “fastQC” directory where the FASTQ files are stored, stores the file names
in a variable “filename”, then runs the FastQC program non-interactively, and finally saves
the QC reports in the “qc” directory:

mkdir qc
cd fastQC
filenames=$(ls *.fastq)
fastqc $filenames \
--outdir ../qc \
--threads 3
cd ..

We can also simply use the following command:

mkdir qc
cd fastQC
fastqc *.fastq --outdir ../qc --threads 3

The QC reports of the FASTQ files will be stored in the “qc” directory. FastQC will gen-
erate an HTML file “*_fastqc.html” and a zipped file “*_fastqc.zip” for each FASTQ file.
22   ◾    Bioinformatics

The zipped file contains the images and the text files that will be displayed on the HTML
when it is opened on an Internet browser. We can display the HTML files from the Linux
­terminal using Firefox. If Firefox is not installed, use “sudo apt-get install firefox” to
install it. If you are using Putty to access a remote computer, you may need to enable “X11
Forwarding”.
Use the following syntax to display the QC reports:

firefox HTMLfile1 HTMLfile2 .. HTMLfileN

Alternatively, you can use a bash script as follows:

cd qc
htmlfiles=$(ls *.html)
firefox $htmlfiles
cd ..

The Firefox Internet browser will display the QC FastQC reports of all FASTQ files, each
on a separate tab as shown in Figure 1.11. Click a tab to move from a report to another.
A QC FastQC report includes a summary, basic statistics, and the quality control met-
rics that we will discuss in some detail. Summary, on the left-hand side, provides a quick
visual report to the different QC metrics available in the report and allows us to identify
any existing problem. With each metric title, Summary displays a sign indicating whether
there is any potential quality problem; the green color indicates a normal metric, orange
indicates a slightly abnormal metric, and red indicates a quality problem associated with
the metric. A metric with an orange warning requires an attention and correction if it is
possible and a metric with a red warning indicates that there is a clear fault in the sequence
data and it is recommended to fix it before proceeding to the analysis step. Summary also

FIGURE 1.11 QC FastQC reports.

 Sequencing and Raw Sequence Data Quality Control   ◾    23

acts as a table of content. Clicking a Summary item will take you to that item graph. In the
following, we will discuss each item that may be included on the QC report.

1.5.1 Basic Statistics

Basic Statistics table, as shown in Figure 1.11, includes summary information about the
FASTQ file analyzed with the FastQC. Basic Statistics title is always green and does not
show a warning sign. The information in the basic statistics includes the file name, file type,
encoding, total sequences (number of reads), sequence flagged, sequence length, and %GC
content. “Filename” field indicates the name of the FASTQ file analyzed. “File type” field
indicates whether the analyzed file was generated by the conventional base calling or by
another means. “Encoding” field indicates the ASCII encoding for the Phred quality score.
Since the encoding in the table is “Sanger / Illumina 1.9”, the Phred+33 encoding (Q33)
was used to encode the per base quality scores (as described above). “Total sequences”
field shows the number of records in the FASTQ file. “Filtered Sequences” field shows the
number of removed reads if the FASTQ file is in Casava format (Casava FASTQ file will
be discussed in Chapter 7). “Sequence Length” field shows either the length of the shortest
and longest sequence if the length of the reads is variable or a single value if all reads in
the FASTQ file have the same length. The “%GC” field shows the overall percentage of GC
content of all bases in all reads.
We should pay attention to “Total Sequences” if the analyzed FASTQ files are paired
end; the number of sequences must be the same in both files (forward and reverse files)
as shown in Figure 1.12. Otherwise, some programs used in the analysis may complain
because they expect matched pairs and in the same order in both files.
We should also pay attention to “Sequence length”. A range of numbers as shown in
Figure 1.13 indicate that the lengths of the sequences in the file are not equal and they can
be any length in the range between the two numbers. Some programs used in the analysis
may expect that all reads in a FASTQ file have equal length.

FIGURE 1.12 The basic statistics of the QC report.

24   ◾    Bioinformatics

FIGURE 1.13 The basic statistics of the paired-end FASTQ files.

The GC content (%GC) is important for showing sequencing problem due to bias. There
is a relationship between GC content and read coverage across a genome. Some short-read
sequencers may tend to sequence the region with higher GC content more or less than the
region with the lower GC content [11]. The average of the GC content of bacterial genome
varies from less than 15% to more than 75% [12], and the average genomic GC content of
most eukaryotes lies somewhere within 40%–50% [13]. Very small or very large GC con-
tent may indicate a potential sequencing bias problem that we may need to fix.

1.5.2 Per Base Sequence Quality

Per Base Sequence Quality graphs (Figure 1.14) show box plots of the quality distributions
on each position across all bases of the reads in the forward and reverse FASTQ files. The
position indexes are plotted in the x-axis against the Phred quality scores in the y-axis. The
higher the quality score, the better the base call.
The maximum limit of the x-axis indexes depends on the length of the reads if they have
the same length; otherwise, it will be the number of bases in the longest read. Each box
plot displays the five-number summary of the base quality scores in that position. The five-
number summary includes the minimum quality score (the lower end of lower whisker),
first quartile (the lower end of the box), median (the red line), third quartile (the upper end
of the box), and maximum quality score (upper end of the upper whisker). The yellow box of
a box plot represents the interquartile range (IQR), which is the middle 50% of the quality
scores of the bases in that position. The longer the box, the more the spread of the quality
scores. The blue line on the graph represents the mean of the quality scores.
The background of the graph is divided into three regions. The quality scores on the
green region (Q>28) are very good, the scores on the orange region (20 < Q <28) are reason-
able, and the scores on the red region (Q<20) are poor. In general, for the reads produced
by Illumina, the per base qualities degrade toward the end of the reads.
If the lower quartile of the quality scores for any position is less than 20, an orange
warning will be displayed in front of “Per Base Sequence Quality” as shown in Figure 1.14.
The red warning (failure) is displayed if the lower quartile for any position is less than 5.
We can notice that the average quality of the base drops toward the end of the reads and
it is greatly deteriorated in the 36th base. To avoid using low-quality sequence data in the
 Sequencing and Raw Sequence Data Quality Control   ◾    25

FIGURE 1.14 Per base sequence quality with warning.

analysis, we may need to filter the reads that have low-quality bases or to trim the ends
of the reads beginning from the 34th base. Figure 1.15 shows a per base sequence quality
graph without warning.

1.5.3 Per Tile Sequence Quality

The per tile sequence quality is represented by a heatmap graph that is available only if an
Illumina sequencer was used and the reads in the FASTQ file retain the original identi-
fiers including the IDs of the flow cell tiles on which reads were sequenced. Figure 1.16
shows the first two records of the FASTQ file “SRR576933.fastq”. Notice that the identifier
line of each record contains the sequence ID, the flow cell ID, lane number, tile number,
x-coordinate and y-coordinate of the tile, and read length.
In the graph, the base position indexes are plotted in the x-axis against the physical posi-
tions on the flow cells (tile numbers) in the y-axis. The base quality is represented by a color
scale from blue (cold) to red (hot). The blue color indicates that the quality of the base from
the tile is at or above the average for the base in the run. The red color indicates the qual-
ity for the base in that tile is worse than the quality for the same base from the other tiles.
The graph provides an easy way to track the average quality scores from each tile across all
26   ◾    Bioinformatics

FIGURE 1.15 Good per base sequence quality (with no warning).

FIGURE 1.16 The last records of the FASTQ file “SRR030834.fastq”.

bases in a base position and to quickly detect if there is any poor quality associated with
any tile. A graph with a completely blue color indicates an overall good quality base call
from all tiles (Figure 1.17a).
The quality losses associated with the physical positions on the flow cell are of different
causes. A random low quality at different positions is usually due to a technical problem
with the run such as the flow cell overlapping (Figure 1.17b).
 Sequencing and Raw Sequence Data Quality Control   ◾    27

FIGURE 1.17 Per tile sequence quality graphs, some with quality problems.

Reads with such random scattered hot areas (errors) are hard to be fixed. Another
cause of the loss of quality is the failure of the imaging system at the edges of the flow
cell to read signals (Figure 1.17b). However, the reads in such case can be still useable.
The loss of quality that begins from somewhere and continues until the end of the run
(Figure 1.17c) is usually due to an obstruction to the imaging system (because of a dirt
on the surface of the flow cell). The reads with low quality, in this case, can be filtered
out. A final type of errors associated with tiles is a temporary loss of quality in specific
base positions (Figure 1.17d) due to an obstruction to few cycles (because of bubbles). The
obstruction usually stops the imaging and also prevents the reagents from getting to the
DNA template clusters. Such quality problem may introduce false insertion to the reads
and it cannot be fixed with sequence trimming since the low-quality bases are not at the
end of the reads.
Indeed, not every quality problem associated with tiles can be removed by trimming
or filtering out reads originating from the affected tiles. However, the last one may affect
several tiles and may extend several reads without lowering the overall sequence score. We
should watch out to this kind of fault especially for reads used for variant discovery.
28   ◾    Bioinformatics

1.5.4 Per Sequence Quality Scores

The per sequence quality score graph is created by plotting the mean sequence quality
(Phred scores) in the x-axis against read count (frequency) in the y-axis. The graph allows
us to see if a subset of reads have an overall low quality. The ideal curve is the one that shows
the majority of the reads having an overall quality score at or over 30 (Figure 1.18a); a peak
is toward the end of the x-axis. The presence of a large number of reads with an overall low
quality will indicate a systematic problem in the run. A warning sign is displayed if the
mean quality score of the majority of reads is below 27 (Figure 1.18b). An error is displayed
if the average quality score of the majority of the reads is below 20. The low-quality reads
can be filtered out to keep only the reads that pass a quality threshold.

1.5.5 Per Base Sequence Content

The per base sequence content graph depicts the percentage of each of the four bases (A, C,
G, and T) called at each position across all reads in a FASTQ file. The positions are plotted in
the x-axis against the base percentage in the y-axis. If there is no bias and library sequenc-
ing is random, we will expect no big difference between the distributions of the four bases
in each position. The percentage of each base is expected to be close to 25% and the four
lines will run approximately parallel to each other as shown in Figure 1.19a. Any deviation
from that, such as a bias or a systematic fault, will be suspected, and hence, some sequences
may be overrepresented as shown in Figure 1.19b. Higher percentage of some bases at the
beginning of the x-axis may indicate contaminating remnants of adaptor sequences or
other contaminating sequences. A warning is displayed if the difference between any of the
four bases is greater than 10% in any position and the failure of this metric occurs if the
difference between any four of bases is greater than 20% in any position.

1.5.6 Per Sequence GC Content

The per sequence GC content graph plots the number of reads in the y-axis against the
mean GC percentage per read in the x-axis (Figure 1.20). It depicts the distribution of GC

FIGURE 1.18 Per sequence quality scores.

 Sequencing and Raw Sequence Data Quality Control   ◾    29

FIGURE 1.19 Normal and abnormal per base sequence content.

FIGURE 1.20 Normal and abnormal per sequence GC content.

content across the reads in a FASTQ file and then compares it to the theoretical normal
distribution of the GC content, which is estimated from the observed data. If there is no
sequencing bias and the library is random, we will expect that the observed distribution of
the GC content of the reads to be approximately normal and roughly similar to the theo-
retical distribution in which the central peak corresponds to the overall GC content of the
underlying genome. Deviation of the distribution of the per sequence GC content from the
normal distribution is an indication of a contaminated library or a fault in the sequencing
process. However, a bell-shaped normal curve that deviates from theoretical curve may or
may not be biased. In this case, there is a chance that the observed distribution may repre-
sent the actual distribution of the genome of the organism; therefore, no warning will be
issued.
30   ◾    Bioinformatics

A warning sign is displayed if the observed distribution deviates from normal distribu-
tion by a sum of more than 15% of the reads. A failure sign will be displayed if the distribu-
tion deviates by a sum of more than 30% of reads.

1.5.7 Per Base N Content

During the sequencing process, a base is called with a high confidence. However, for some
fault, the machine may fail to call any base at a specific position. The “N” character is then
placed at that position as an indication of call failure. A few call failures are tolerable;
however, if the frequency of “N” is high, that may pose a quality problem. The per base N
content graph shows the distribution of “N” at each base position. The N percentages are
plotted in the y-axis against positions in the x-axis. A warning is issued if any position
shows an N content of greater than 5% and a failure sign if any position shows an N content
of greater than 20%. Figure 1.21 shows the per base N content with no problem.

1.5.8 Sequence Length Distribution

In the library preparation step, DNA molecules are cut into equal fragments to generate
reads with equal lengths. Most sequencing instruments run quality control to keep the

FIGURE 1.21 Per base N content.

 Sequencing and Raw Sequence Data Quality Control   ◾    31

FIGURE 1.22 Sequence length distribution graphs (equal length and variable lengths).

length of the reads equal. However, sometimes reads with unequal reads are generated
especially if the reads are trimmed to remove low-quality bases at the beginning or ends
of the reads. The sequence length distribution graph shows the read length distribution.
If the reads are of the same length, the graph will be simple with a single peak at a bar
indicating a single value (Figure 1.22a). When reads are of a variable length, the graph will
show the relative read count of each read length (Figure 1.22b). A warning is displayed if
the reads do not have the same length.

1.5.9 Sequence Duplication Levels

The PCR may be used in sequencing step especially if the concentration of DNA is low,
in RNA-Seq and ChIP-Seq for enrichment. The PCR will increase the number of DNA
fragments; a single fragment is duplicated several times (exact match). However, well-cali-
brated sequencing instrument will produce, at the end, a single read for each of the library
fragments. Low sequence duplication level may indicate a high level of coverage. In con-
trast, the high level of duplication indicates a bias due to PCR amplification. The graph of
sequence duplication levels plots the percentages of reads against the sequence duplication
levels (number of duplicates). Only the first 200,000 reads in a FASTQ file are checked for
duplication to save computer memory. The number of duplicates is counted for each read.
A big rise may indicate the presence of a large number of reads with high levels of dupli-
cation. A warning is displayed if the number of duplicated reads is more than 20% of the
total. A failure sign is shown if the number of duplicate reads is more than 50% of the total.
Figure 1.23a shows that the majority of reads are unique. However, the number of dupli-
cated reads is more than 20% of the total reads; therefore, a warning is issued. Figure 1.23b
shows that the number of duplicated reads is more than 50% of the total; therefore, the
metric failed.

1.5.10 Overrepresented Sequences

The overrepresented sequences of genomic DNA will indicate a clear bias or contamination
due to adaptor dimers. However, in RNA-Seq, the overrepresented sequences can also be
32   ◾    Bioinformatics

FIGURE 1.23 Sequence duplication levels (warning and failure).

FIGURE 1.24 Overrepresented sequences.

due to the gene expression and that overrepresentation can be of a biological importance
rather than a bias. The overrepresented sequences report is a table that shows the over-
represented sequences, counts, percentage, and possible source. To save memory, only the
first 200,000 reads are checked in the FASTQ file; therefore, the list is not exhaustive and
other overrepresented sequences may skip the check. For each overrepresented sequence,
the FastQC program will search on a database of known contaminants and report the best
match that is at least 20 bases in length and has no more than a single mismatch. A warn-
ing will be issued if a sequence is overrepresented more than 0.1% of the total and failure
will occur if the overrepresentation is more than 1% of the total. As shown in Figure 1.24,
five overrepresented sequences are found, three of which are contaminating adaptors and
two sequences have no hits. The count and percentage reflect the significance of each of
these overrepresented sequences. The count of the first sequence in the table represents
29.4% of the total count of the reads in the FASTQ file. It is clear that this sequence is origi-
nated from a primer contamination and it must be removed before analysis.

1.5.11 Adapter Content

The full-length adaptor primers may cause contaminating adaptor dimers of a significant
number of reads. The adaptor content graph shows the cumulative percentage count of
 Sequencing and Raw Sequence Data Quality Control   ◾    33

FIGURE 1.25 Some known adaptor sequences.

FIGURE 1.26 Adaptor content graphs.

the proportion of the reads which contain the adaptor sequences at each position. Known
adaptor sequences and description are stored in the “adapter_list.txt” file as shown in
Figure 1.25.
K-mers (sequences of k size of bases) are formed from the adaptor sequences in the
“adapter_list.txt” file and then the program searches for these k-mers to report the total
percentage of the reads which contain these k-mers. The report may discover the sources of
bias due to contaminating adaptor dimers in the library.
A warning is raised if any sequence is present in more than 5% of all reads, and a failure
occurs if any sequence is present in more than 10% of all reads.
Figure 1.26 shows a FASTQ file with raw reads without adaptor content (left) and a
FASTQ file with reads with failed metric due to significant content of Illumina Universal
Adaptor.

1.5.12 K-mer Content

The K-mer content graph plots the count of each short nucleotide of length k (default k = 7)
against positions in reads. In a normal k-mer content, k-mers are expected to be repre-
sented evenly across the length of the reads. The k-mer content graph shows the positions
for the only six most significant k-mers (Figure 1.27). The list of k-mers which are present
at specific position with significant abundance will be reported in a table including k-mer
sequences, counts, p-values, and expected position. Caution is required when the reads are
from RNA-Seq libraries; the significant k-mers may be due to highly expressed gene, and
hence, they will have a biological importance.
34   ◾    Bioinformatics

FIGURE 1.27 Failed k-mer content.

1.6 PREPROCESSING OF THE FASTQ READS

In the above, we discussed the assessment of the quality of the reads produced by the HTS
instruments to understand the potential errors and biases that may arise from warnings or
failures of the quality metrics. Before moving on to the next step for data analysis, errors
and biases should be adjusted to avoid incorrect results and misleading interpretation. In
general, there are three common approaches to fix the biases resulted from the quality
metrics. Those three approaches include (i) trimming the ends of the reads, (ii) removing
low-quality reads, and (iii) masking low-quality bases. The use of any of those approaches
depends on the quality problem. In the following, we will discuss the most commonly used
programs to deal with read quality issues.
The most commonly used software for the processing of raw sequence reads in FASTQ
files is FASTX-toolkit [14], which is a collection of command-line programs. The installa-
tion instructions of FASTX-toolkit are available at “http://hannonlab.cshl.edu/fastx_tool-
kit/download.html”. We can download and install it on Linux using the following steps:
Create a directory in which you can download the FASTX-toolkit compressed file:

mkdir fastxtoolkit
cd fastxtoolkit

Download the compressed program file and decompress it:

wget http://hannonlab.cshl.edu/fastx_toolkit/fastx_toolkit_0.0.13_
binaries_Linux_2.6_amd64.tar.bz2
tar xvf fastx_toolkit_0.0.13_binaries_Linux_2.6_amd64.tar.bz2

Copy the program files from the “bin” directory to “/usr/local/bin” so that it can be exe-
cuted from any directory on the computer:

sudo cp ./bin/* /usr/local/bin

 Sequencing and Raw Sequence Data Quality Control   ◾    35

FIGURE 1.28 FASTX-toolkit programs.

TABLE 1.4 FASTX-Toolkit Programs and Descriptions

Command Name Description
fastq_to_fasta converts FASTQ files to FASTA files
fastx_quality_stats charts Quality Statistics and Nucleotide Distribution
fastx_collapser collapses identical sequences into a single sequence
fastx_uncollapser expands collapsed identical sequences
fastx_trimmer trims reads in a FASTQ files (removing barcodes or noise)
fastx_renamer renames the sequence identifiers in FASTQ/A file
fastx_clipper removes sequencing adapters/linkers
fasta_clipping_histogram.pl creates a Linker Clipping Information Histogram
fastq_quality_boxplot_graph.sh creates quality boxplot
fastx_nucleotide_distribution_graph.sh creates nucleotide distribution graph
fastx_nucleotide_distribution_line_graph.sh creates nucleotide distribution line graph
fastx_reverse_complement produces the Reverse-complement of each sequence
fastx_barcode_splitter.pl splits a FASTQ/FASTA files containing multiple samples
fasta_formatter changes the width of sequences line in a FASTA file
fasta_nucleotide_changer converts FASTA sequences from/to RNA/DNA
fastq_quality_filter filters sequences based on quality
fastq_quality_trimmer trims (cuts) sequences based on quality
fastx_artifacts_filter FASTQ/A Artifacts Filter
fastq_masker masks nucleotides with “N” based on quality

Figure 1.28 shows the FASTX-toolkit tools in the “bin” directory after downloading and
extracting the compressed archive file.
FASTX-toolkit includes several tools for the processing of FASTQ files as described in
Table 1.4. You can display the usage and options of any of the executable programs by
entering the program name with “-h” option on the command-line prompt. For instance,
to display the help for “fastq_quality_filter”, simply enter the following on the command
line:

fastq_quality_filter -h

To show how FASTQ files are processed, we will download a raw FASTQ file from the NCBI
SRA database and modify its name for the practice. The following commands create the
36   ◾    Bioinformatics

directory “preprocessing”, then download the FASTQ file from the NCBI SRA d ­ atabase,
and finally rename it to “bad.fastq” file for the practice purpose. The script then generates
the QC FastQC report and displays the report on the Firefox browser.

mkdir preprocessing
cd preprocessing
fasterq-dump --verbose SRR957824
rm SRR957824_1.fastq
mv SRR957824_2.fastq bad.fastq
fqfile=$(ls *.fastq)
fastqc $fqfile
htmlfile=$(ls *.html)
firefox $htmlfile

When all the commands have been executed sequentially without an error, the QC report
will be displayed on the Firefox Internet browser. Study the reports carefully and identify
any potential problems on the quality metrics that we have discussed in the previous sec-
tion. Figure 1.29 shows that the reads in the file have three failures and a single warning.
Next, we will try to fix these problems as possible.
Using such FASTQ file in the downstream analysis without fixing some of the quality
problems will definitely impact the results negatively and may lead to misleading results.
The good strategy whenever there are warnings or failures is to try the available ways to fix
the problems as possible, and if there is any unfixable problem, you may need to be aware
of it and to know how it may affect the results.

FIGURE 1.29 The QC report summary and per base sequence quality for “bad.fastq” file.
 Sequencing and Raw Sequence Data Quality Control   ◾    37

Figure 1.29 shows a common deterioration of the quality of bases toward the end of the
reads produced by short-read sequencing instruments. We can also notice that some qual-
ity scores in some positions are low as 2 Phred (probability of error is 0.6).
The report shows three failures and a single warning: failed per base sequence quality
(Figure 1.29), failed per base sequence content and failed k-mer content (Figure 1.30), and
overrepresented sequences warning (Figure 1.31).
The QC processing strategies are different from a FASTQ file to another depending on
the failed metrics. Understanding the problem always gives a good idea about which kinds
of QC processing to perform. In our example file, we will begin by filtering the low-quality
reads and clipping the overrepresented sequences and then we will run FastQC again to see
how the quality is improved.
First, we will try to fix the per base quality score of the reads in the FASTQ file by
using “fastq_quality_filter” to keep the reads that have 80% of the bases which have qual-
ity scores equal or greater than 28. The following script performs filtering (the output file
is “filtered.fastq”), runs FastQC to generate the new QC report, and then runs Firefox to
display the QC report on the Internet browser:

fastq_quality_filter \
-i bad.fastq \
-q 28 \

FIGURE 1.30 Failed per base sequence content and k-mer content.

FIGURE 1.31 Overrepresented sequences that raised a warning.

38   ◾    Bioinformatics

-p 80 \
-o filtered.fastq \
-Q33
fastqc filtered.fastq
firefox filtered_fastqc.html

In the above script, “-i” option specifies the input FASTQ file, “-q” specifies the minimum
Phred quality threshold, “-p” specifies the percentage of bases of the reads that have at least
the specified threshold quality, “-o” specifies a name of the output FASTQ file where the
filtered reads are stored, and “-Q33” is to tell the program that the FASTQ quality encod-
ing is Phred+33 (the default is “-Q64”; therefore, we must use “-Q33” for FASTQ files with
Illumina 1.9 encoding or later).
Figure 1.32 shows the per base sequence quality graph of the filtered FASTQ file. The
filtering process removed 499,970 reads, which did not meet the criteria. The per base
sequence quality, which is the most important metric, has been improved and per base
sequence content has been also improved. However, some positions at the ends of the reads
have still low Phred quality scores. We can trim the low-quality bases from the ends of the
reads by using the “fastq_quality_trimmer” program. Instead of removing the reads that

FIGURE 1.32 A graph of the filtered “bad.fastq” file with low-quality bases at the read ends.
 Sequencing and Raw Sequence Data Quality Control   ◾    39

do not meet the criteria entirely, this program cuts only the bases, whose quality scores are
less than the specified threshold, from the ends of the reads.

fastq_quality_trimmer \
-i bad_filt.fastq \
-t 28 \
-o bad_filt_trim.fastq \
-Q33
fastqc bad_filt_trim.fastq
htmlfiles=$(ls *.html)
firefox $htmlfiles

The “-t” option specifies the quality threshold, which is the minimum quality score below
which the bases will be trimmed from the ends of the reads. When trimming is performed,
the resulted reads may be of unequal lengths, which may not be accepted by some pro-
grams used in following steps of analysis. As shown in Figure 1.33, although the per base
sequence quality has been improved by trimming, it also raised a sequence length distribu-
tion warning since trimming resulted in reads with unequal lengths. We may need to filter
reads by length.

FIGURE 1.33 The QC report of the filtered and trimmed “bad.fastq” file.
40   ◾    Bioinformatics

The sequence length distribution warning also can be raised by clipping a­ daptors or
­ verrepresented sequences. Thus, we can use “fastx_clipper” first to remove the overrep-
o
resented sequences (see Figure 1.31). The following script removes a contaminating over-
represented sequence:

fastx_clipper \
-a ATCGGGAGAGGGGCGGGGAGGGGAAGAGGGGAGAATTCGGGGGGGGCCGG \
-i bad_filt_trim.fastq \
-o bad_filt_trim_clip.fastq \
-v \
-Q33
fastqc bad_filt_trim_clip.fastq
htmlfiles=$(ls *.html)
firefox $htmlfiles

Since some aligners in the next step of analysis may not accept sequences with unequal
lengths, we can use a bash script to filter out the short reads. Figure 1.34 shows sequence
length distribution. If the aligner that we intend to use does not accept unequal read
lengths, then we can filter out all reads whose length is less than 150 bases using the fol-
lowing script:

FIGURE 1.34 Sequence length distribution (different lengths).

 Sequencing and Raw Sequence Data Quality Control   ◾    41

paste <(cat bad_filt_trim_clip.fastq | paste - - - -) \

| awk -v FS=’\t’ ‘length($2) >= 150 && length($4) >= 150’ \
| tee >(cut -f 1-4 | tr ‘\t’ ‘\n’ > bad_filt_trim_clip_eq.fastq)
fastqc bad_filt_trim_clip_eq.fastq
htmlfiles=$(ls *.html)
firefox $htmlfiles

If you use the above script for other FASTQ files, you may need to change the read length
and the numbers of the columns. In our example FASTQ file, “$2” is for the sequence col-
umn and “$4” is for the quality column. The numbers of these columns may vary depend-
ing on the content of the FASTQ definition line.
Figure 1.35 shows the per base sequence quality of the final FASTQ file. You should
remember that you may not be able to fix all quality problems, and that filtering and clip-
ping may compromise the sequencing depth. Fortunately, most of the problems other than
base quality errors can be tolerated by the majority of the aligning programs. However, we
should try to fix the failed metrics as possible before continuing to the subsequent step of
the analysis.

FIGURE 1.35 Cleaned FASTQ file.

42   ◾    Bioinformatics

The FASTX-toolkit tools listed in Table 1.4 are used for quality assessment and quality
adjustment. The major limitation is that “fastq_quality_filter” of FASTX-toolkit does not
process the paired-end FASTQ files together and that usually results in singletons or reads
without pairs in any of the two paired-end FASTQ files. Most aligners do not accept to pro-
cess paired-end FASTQ files with singletons. The FASTX-toolkit solution to the singleton
problem is to mask the low-quality bases instead of removing the reads with low-quality
bases. Thus, “fastq_masker” program is used instead of “fastq_quality_filter” to mask the
bases of Phred quality score less than a user-defined threshold “-q”.

fastq_masker \
-q 20 \
-i bad.fastq \
-o bad_masked.fastq \
-Q33
fastqc bad_masked.fastq
firefox bad_masked_fastqc.html

The above “fastq_masker” command masks the bases with quality lower than 20 Phred
quality score “-q 20”; therefore, they will be ignored by aligners and assemblers.
For paired-end FASTQ files produced by an Illumina instrument, there is another
FASTQ processing program, developed by Illumina for paired-end FASTQ files, called
Trimmomatic [15]. It is a multithreaded command-line Java-based program and is more
modern than FASTX-toolkit. It was developed by Illumina to perform several operations,
including detection and removing the known adaptor fragments (adapter.clip), trim-
ming low-quality regions from the beginning of the reads (trim.leading), trimming low-
quality regions from the end of the reads (trim.trailing), filtering out short reads (min.
read.length), in addition to other operations with different quality-filtering strategies for
dropping low-quality bases in the reads (max.info and sliding.window). Trimmomatic can
be used in two modes: simple and palindrome modes. In the simple mode, for removing
adaptor sequences, the pairwise local alignment between adaptor sequence and reads is
used to scan reads from 5′ ends to 3′ ends using seed and extend approach. If a score of
a match exceeds a user-defined threshold, both the matched region and the region after
alignment will be removed. The entire read is removed if an alignment covers all the read.
The simple Trimmomatic approach may not be able to detect the short adaptor sequence.
Therefore, the palindrome model is used because it is able to detect and remove short
fragment sequences of adaptors. Palindrome is used only for the paired-end data. Both
forward and reverse reads will have equal number of valid bases and each read comple-
ments another. The valid reads are followed by the contaminating bases from the adaptors.
The tool uses the two complementary reads to identify the adaptor fragment or any other
contaminating technical sequence by globally aligning the forward and reverse reads. An
alignment score that is greater than a user-defined threshold indicates that the first parts
of each read reversely complement one another and the remaining read fragments which
match the adaptor sequence will be removed.
 Sequencing and Raw Sequence Data Quality Control   ◾    43

Trimmomatic is available at “http://www.usadellab.org/cms/index.php?page=

trimmomatic”. You can download it from the website and unzip it using the following
script:

$ wget http://www.usadellab.org/cms/uploads/supplementary/
Trimmomatic/Trimmomatic-0.39.zip
$ unzip Trimmomatic-0.39.zip

Notice that the version may change in the future. The unzipped directory is
“Trimmomatic-0.39”, where there will be two files (“LICENSE” and “trimmomatic-0.39.
jar”) and a directory (“adapters”). The file “trimmomatic-0.39.jar” is the Java executable
program that performs the preprocessing tasks and the directory “adaptors” contains the
known adaptor sequences in FASTA files. The following script uses Trimmomatic to repro-
cess the paired-end FASTQ files, then runs FastQC to generate QC reports, and finally
displays the reports on the Firefox browser:

java -jar ../Trimmomatic-0.39/trimmomatic-0.39.jar \

PE SRR957824_1.fastq SRR957824_2.fastq \
out_PE_SRR957824_1.fastq out_UPE_SRR957824_1.fastq \
out_PE_SRR957824_2.fastq out_UPE_SRR957824_2.fastq \
ILLUMINACLIP:TruSeq3-PE.fa:2:30:10:2:True \
LEADING:3 \
TRAILING:3 \
ILLUMINACLIP:TruSeq2-PE.fa:2:30:10 \
SLIDINGWINDOW:5:30 \
MINLEN:35
fastqc out_PE_SRR957824_1.fastq out_PE_SRR957824_2.fastq
firefox out_PE_SRR957824_1_fastqc.html out_PE_SRR957824_2_fastqc.
html

The option “PE” is used for paired end, and then the two paired-end FASTQ files
“SRR957824_1.fastq” and “SRR957824_2.fastq” were provided as inputs. The adaptors that
were detected and removed from the reads are stored in the “TruSeq3-PE.fa” file in the
“adaptors” directory. Hence, “ILLUMINACLIP:TruSeq2-PE.fa” is used to specify the file
in which the adaptor sequences are stored. The program removed the leading and trail-
ing edges of reads with low quality that is below 3 Phred quality score. The “SLIDING-
WINDOW:5:35” is used so that the program can scan the read with a 5-base wide sliding
window and remove a read when the window per base average quality score declines to
below 30. Finally, the program removes the reads that are shorter than 35 bases.
In Figures 1.36 and 1.37, notice how the quality of the two files have been improved and
also notice that the total sequence is equal in both files. However, the read lengths vary.
If for any reason we need reads of the same length as some aligners may require, we can
set “MINLEN:” to the maximum length. Since the maximum read length is 150 bases, we
can use “MINLEN:151” as follows:
44   ◾    Bioinformatics

java -jar ../Trimmomatic-0.39/trimmomatic-0.39.jar \

PE SRR957824_1.fastq SRR957824_2.fastq \
out_PE_SRR957824_1.fastq out_UPE_SRR957824_1.fastq \
out_PE_SRR957824_2.fastq out_UPE_SRR957824_2.fastq \
ILLUMINACLIP:TruSeq3-PE.fa:2:30:10:2:True \
LEADING:3 \
TRAILING:3 \
ILLUMINACLIP:TruSeq2-PE.fa:2:30:10 \
SLIDINGWINDOW:5:30 \
MINLEN:150
fastqc out_PE_SRR957824_1.fastq out_PE_SRR957824_2.fastq
firefox out_PE_SRR957824_1_fastqc.html out_PE_SRR957824_2_fastqc.
html

Check out the reports of the two paired-end FASTQ files to see that only the reads with
equal length are left.
There are several other programs that can be used, as well, for the quality improvement
of FASTQ reads. For instance, Fastp can be used for the filtering of low-quality reads and

FIGURE 1.36 Trimmomatic processed forward FASTQ file.

 Sequencing and Raw Sequence Data Quality Control   ◾    45

FIGURE 1.37 Trimmomatic processed reverse FASTQ file.

adaptor trimming. This program is specifically fast and easy to use as part of a pipeline.
Moreover, it is able to identify adaptor sequences and trim them without the need of pro-
viding adaptor sequences [16].

1.7 SUMMARY
The NGS produces short reads that are widely used for the different sequencing applica-
tions for the high accuracy and low cost. However, the long reads produced by the TGS
(Pacific Bioscience and Oxford Nanopore Technologies) have also gained some popularity
in applications like de novo assembly, metagenomics, and epigenetics. The accuracy of the
long-read technologies has been substantially improved, but the cost is still high and less
affordable when they are compared to short-read technologies. The sequencing depth and
base call quality are the two crucial factors for most applications, and the analysts must
keep looking at them before proceeding with the analysis. Most HTS instruments per-
form quality control before delivering raw sequence data in FASTQ files. However, per base
qualities and other quality metrics must be assessed before using raw data in any analysis.
46   ◾    Bioinformatics

There are several programs for quality assessment, but FastQC is the most popular one.
FastQC is a user-friendly program to assess the quality of the reads generated by any of the
sequencing technologies, and it produces a report that summarizes the results in graphs
that are easy to interpret. The potential quality problems include low-quality bases, pres-
ence of adaptor sequences connected to the reads, presence of adaptor dimers or other
technical contaminating sequences, overrepresented PCR sequences, sequence length dis-
tribution, per base sequence content, per sequence GC content, per base N content, and
k-mer content. The per base sequence quality and adaptor content are the most important
metrics that we should look at and take the appropriate action. The ideal sequencing data
are the one without warnings or failed metrics. Therefore, we should try to fix the prob-
lems as possible. However, some problem may not be solved. If the unsolved problem does
not affect the reads severely, that data still can be used in the analysis. However, we must
be aware that unsolved problems may have some negative impact in the results. The read
quality problems can be solved based on the failed metrics by removing low-quality reads,
trimming the reads from the beginning and the end of the reads, and masking the bases
with low-quality scores. There are several programs for the processing of raw sequence
data. FASTX-toolkit is the most popular one for single-end FASTQ files, and Trimmomatic
is more sophisticated and can be used for both single-end and paired-end raw data. Fastp
filters low-quality reads and automatically recognizes and trims adaptor sequences. It is
important to process the paired-end FASTQ files (forward and reverse) together to avoid
leaving out singletons, which may not be accepted by almost all aligners. In this chapter, we
discussed the command-line programs for quality controls. However, those programs or
similar ones are implemented in Python, R, and other programing languages, but under-
standing the general principle for checking the raw data quality and solving potential qual-
ity problems are the same. Most sequencing applications use these kinds of QC processing,
but when we cover the metagenomic data analysis, you will learn how to preprocess micro-
bial raw data using different programs. Once the raw sequencing data are cleaned, then we
can move safely to the next step of sequence data analysis depending on the application
workflow that we are adopting.

REFERENCES
1. Holley RW, Everett GA, Madison JT, Zamir A: Nucleotide sequences in the yeast alanine
transfer ribonucleic acid. J Biol Chem 1965, 240: 2122–2128.
2. Jou WM, Haegeman G, Ysebaert M, Fiers W: Nucleotide sequence of the gene coding for the
bacteriophage MS2 coat protein. Nature 1972, 237(5350):82–88.
3. Margulies M, Egholm M, Altman WE, Attiya S, Bader JS, Bemben LA, Berka J, Braverman
MS, Chen Y-J, Chen Z et al: Genome sequencing in microfabricated high-density picolitre
reactors. Nature 2005, 437(7057):376–380.
4. Braslavsky I, Hebert B, Kartalov E, Quake SR: Sequence information can be obtained from sin-
gle DNA molecules. Proceedings of the National Academy of Sciences 2003, 100(7):3960–3964.
5. Rhoads A, Au KF: PacBio sequencing and its applications. Genomics, Proteomics &
Bioinformatics 2015, 13(5):278–289.
6. Levene MJ, Korlach J, Turner SW, Foquet M, Craighead HG, Webb WW: Zero-mode wave-
guides for single-molecule analysis at high concentrations. Science 2003, 299(5607):682–686.
 Sequencing and Raw Sequence Data Quality Control   ◾    47

7. Cock PJ, Fields CJ, Goto N, Heuer ML, Rice PM: The Sanger FASTQ file format for sequences
with quality scores, and the Solexa/Illumina FASTQ variants. Nucleic Acids Res 2010,
38(6):1767–1771.
8. FASTQ Files [https://support.illumina.com/help/BaseSpace_OLH_009008/Content/Source/
Informatics/BS/FASTQFiles_Intro_swBS.htm]
9. Leinonen R, Sugawara H, Shumway M: The sequence read archive. Nucleic Acids Res 2011,
39(Database issue):D19–21.
10. Andrews S: FastQC: A Quality Control Tool for High Throughput Sequence Data. Babraham
Bioinformatics, Babraham Institute, Cambridge, United Kingdom; 2010.
11. Chen Y-C, Liu T, Yu C-H, Chiang T-Y, Hwang C-C: Effects of GC bias in next-generation-
sequencing data on de novo genome assembly. PLOS One 2013, 8(4):e62856.
12. Lightfield J, Fram NR, Ely B: Across bacterial phyla, distantly-related genomes with similar
genomic GC content have similar patterns of amino acid usage. PLoS One 2011, 6(3):e17677.
13. Romiguier J, Ranwez V, Douzery EJ, Galtier N: Contrasting GC-content dynamics across 33
mammalian genomes: relationship with life-history traits and chromosome sizes. Genome
Res 2010, 20(8):1001–1009.
14. FASTX-toolkit [http://hannonlab.cshl.edu/fastx_toolkit/]
15. Bolger AM, Lohse M, Usadel B: Trimmomatic: A flexible trimmer for Illumina sequence data.
Bioinformatics 2014, 30(15):2114–2120.
16. Chen S, Zhou Y, Chen Y, Gu J: fastp: An ultra-fast all-in-one FASTQ preprocessor.
Bioinformatics 2018, 34(17):i884–i890.
 Chapter 2

Mapping of Sequence Reads

to the Reference Genomes

2.1 INTRODUCTION TO SEQUENCE MAPPING

So far, we have already gone through the first two steps of the NGS/HTP data analysis,
namely acquiring the raw data in FASTQ file format and read quality control. Up to this
point, you know that the sequencing raw data must be cleaned from errors and artifacts,
as much as possible, before moving on to the next step of the data analysis. This chapter
discusses the alignment of reads (short or long) to a reference genome of an organism. This
step is crucial for most of the sequencing applications including reference-guided genome
assembly, variant discovery, gene expression (RNA-Seq), epigenetics (ChIP-Seq, Methyl-
Seq), and metagenomics (targeted and shotgun). The reference genome sequence of an
organism is a key element of read alignment or mapping. Scientists have devoted enormous
amount of time and efforts to determine the sequences of many organisms. Complete
genomes of hundreds of organisms have already been sequenced and the list continues to
grow. The sequencing of human genome was completed in 2003 by the National Human
Genome Research Institute (NHGRI), followed by sequencing the genomes of a vari-
ety of model organisms that are used as surrogates in studying the human biology, then
genomes of numerous of organisms, including some extinct organisms like Neanderthals,
were sequenced. The first sequenced genomes of model organisms include the rat, puffer
fish, fruit fly, sea squirt, roundworm, and the bacterium Escherichia coli. The NHGRI has
sequenced numerous species with the aim to provide data for understanding genetic varia-
tions among organisms. Genome sequences are available in sequence databases funded
by governments and supported by institutions. A reference genome sequence of an organ-
ism is a curated sequence that represents the genome of the individuals of that organism.
However, the sequences of the individuals are varied and the reference sequence is only a
sequence that we compare other sequences to. These days, there are reference genomes for
thousands of organisms, including animal, plants, fungi, bacteria, archaea, and viruses,

DOI: 10.1201/9781003355205-2 49
50   ◾    Bioinformatics

together with their gene annotations, which can be used in the process of read alignment/
mapping to act as guides on which new genomes are assembled fast. A reference genome
of an organism is a curated sequence that is built up using the DNA information of several
normal individuals of that organism. The reference genome curation was pioneered by
the Genome Reference Consortium (GRC), which is founded in 2008 as a collaboration of
the National Center for Biotechnology Information (NCBI), the European Bioinformatics
Institute (EBI), the McDonnell Genome Institute (MGI), and the Wellcome Sanger Institute
to maintain and update the human and mouse genome reference assemblies. Now, GRC
maintains the human, mouse, zebrafish, rat, and chicken reference genomes. Reference
genomes of other organisms are curated by specialized institutions including NCBI and
many others, which manually select genome assemblies that are identified as standard or
representative sequences (RefSeq) against which data of the individuals from those organ-
isms can be compared. All eukaryotes have a single reference genome per species, but pro-
karyotes may have multiple reference genome sequences for a species. The NCBI curates
reference genomes from the assemblies categorized as RefSeq on the GenBank database. If
a eukaryotic species has no assemblies in the RefSeq, then the best GenBank assembly for
that species is selected as a representative genome. Viruses as well may have more than one
reference genomes per species. Generally, the update of a reference genome of any species is
a continuous process and a new version, usually called “Build”, may be released whenever
new information emerges. A release of a reference genome may be accompanied by gene
annotations. A well-curated reference genome, like human and other model organisms’
reference genomes, is usually released with annotation information such as gene anno-
tation and variant annotation. Reference genomes are made available at the NCBI web-
site in both FASTA file format and GenBank file format. Several annotation files may be

FIGURE 2.1 Human reference genome on the NCBI Genome page.

 Mapping of Sequence Reads to the Reference Genomes   ◾    51

found including gene annotation in GFF/GTF file format, GenBank format, and tabular
format. The reference transcriptome (whole mRNA of an organism) and proteins may also
be available as shown in Figure 2.1.
For the alignment/mapping of reads produced by sequencing instruments, we may
need to download a reference genome of the species from which the sequencing raw data
are taken. The sequence of the reference genome must be in the FASTA file format. For
example, to download the FASTA file of the human genome, you can copy the link from
“genome” hyperlink on the Genome database web page and on Linux terminal use “wget”
to download the file to the directory of your choice “e.g. refgenome”:

mkdir refgenome
wget \
-O “refgenome/GRCh38.p13_ref.fna.gz” \
https://ftp.ncbi.nlm.nih.gov/genomes/all/GCF/000/001/405/
GCF_000001405.39_GRCh38.p13/GCF_000001405.39_GRCh38.p13_genomic.
fna.gz

This script will create the “refgenome” directory, where it will download the compressed
FASTA sequence of the human reference genome “GRCh38.p13_ref.fna.gz”. The size of
the compressed current FASTA sequence file of the human genome (GRCh38.p13) is only
921M. We can decompress it using the “gunzip” command.

gunzip -d GRCh38.p13_ref.fna.gz

This command will decompress the reference genome file to “GRCh38.p13_ref.fna” and
the file size now is 3.1G. A large file can be displayed using a program for displaying a large
text such as “less” or “cat” Linux commands. The reference sequences are in the FASTA file
format. A file contains several sequences representing the genomic units such as chromo-
somes. Each FASTA sequence entry consists of two parts: a definition line (defline), which
is a single line that begins with “>” symbol, and a sequence, which may span several lines.
Figure 2.2 shows the beginning of the human genome reference sequence. Notice that the
defline includes the GenBank accession of the sequence, species scientific name, genome
unit (chromosome number), and the human genome Build. A chromosome sequence may
begin with multiple ambiguous bases (Ns). In Figure 2.2, we removed several lines of Ns
intentionally to show the DNA nucleobases.
The following Unix/Linux commands are used with the text files as general and here we
can use them with FASTA files to collect some useful information.
To display the FASTA file content page by page, you can use “less” command:

less GRCh38.p13_ref.fna

To count the number of FASTA sequences in the FASTA file, use “grep” command:

grep -c “>” GRCh38.p13_ref.fna

52   ◾    Bioinformatics

FIGURE 2.2 Part of the FASTA sequence of the human reference genome.

To count the total number of bases in the reference file, you can combine “grep”, “wc”, and
“awk” commands as follows:

grep -v “>” GRCh38.p13_ref.fna | wc | awk ‘{print $3-$1}’

If for any reason, you want to split the reference sequences into files, you can use the fol-
lowing script that creates the directory, “chromosomes”, and then it splits the main FASTA
file into several FASTA files:

mkdir chromosomes
cd chromosomes
csplit -s -z ../GRCh38.p13_ref.fna ‘/>/’ ‘{*}’
for i in xx* ; do \
n=$(sed ‘s/>// ; s/ .*// ; 1q’ “$i”) ; \
mv “$i” “$n.fa” ; \
done

The annotation files relevant to a reference genome may also be needed for some of the
steps in the downstream analysis. You can download the annotation file as above. The
annotation file is a description of where genetic element also called a feature such as genes,
introns, and exons are located in the genome sequence, showing the start and end coordi-
nates, and feature name. The annotation files are usually in GFF or GTF file format. The
GFF (General feature format) is a simple tab-delimited text file for describing genomic
features and mapping them to the reference sequence in the FASTA file. The GTF (Gene
Transfer Format) is similar to GFF but it has additional elements. Figure 2.3 shows the
first part of the human annotation file in the GFF format. The content of an annotation file
including the chromosome name or chromosome GenBank accession in the first column,
and features and other annotations are in the other columns.
Both the FASTA reference file and (sometimes) its annotation file are required by the
alignment programs, shortly called aligners, for mapping the reads in the FASTQ files to
 Mapping of Sequence Reads to the Reference Genomes   ◾    53

FIGURE 2.3 Part of the human annotation file in GTF file format.

the reference genome in a process known as read sequence mapping or alignment. In the
read mapping process, the FASTA files may contain millions of read sequences that we
wish to align to a sequence of a reference genome to produce aligned reads in a file format
called SAM, which stands for Sequence Alignment Map format. The aligned reads can also
be stored in the SAM binary form called BAM (Binary Alignment Map format). We will
discuss this file format later in some detail.
In general, sequence mapping or alignment requires three elements: A reference file in
the FASTA format, short-sequence reads in FASTQ files, and an aligner, which is a program
that uses an algorithm to align reads to a reference genome sequence. We have already
discussed how to download the sequence of a reference genome of an organism from the
NCBI Genome database. However, before using a reference genome with any aligner, it
may require indexing with the “samtools faidx” command. You can download and install
Samtools by following the instructions available at “http://www.htslib.org/download/”. On
Ubuntu, you can install it using the following command:

sudo apt-get install samtools

Once you have installed Samtools successfully, you can use that tool to index the reference
genome and other tasks that you will learn later.
You have already downloaded the human reference genome above. If you didn’t do that,
you can download and decompress it using the following commands:

mkdir refgenome
wget \
-O “refgenome/GRCh38.p13_ref.fna.gz” \
https://ftp.ncbi.nlm.nih.gov/genomes/all/GCF/000/001/405/
GCF_000001405.39_GRCh38.p13/GCF_000001405.39_GRCh38.p13_genomic.
fna.gz
cd refgenome
gunzip -d GRCh38.p13_ref.fna.gz
54   ◾    Bioinformatics

Notice that the name of the reference genome or its URL may change in the future. The
above commands will create the directory “refgenome” where it downloads the human
reference genome and decompresses it. Once the reference genome has been downloaded,
you can use “samtools faidx” to index it as follows:

samtools faidx GRCh38.p13_ref.fna

To read more about this command, run “samtools faidx -h”.

Indeed, for the reference sequence to be indexed by “samtools faidx” command, it must
be in FASTA format and well-formatted, which means that the FASTA sequences contained
in the file must have a unique name or ID in the FASTA defline and the sequence lines of
each sequence must be of the same length. Indexing a reference genome with Samtools
enables efficient access to arbitrary regions within the FASTA file of the reference sequence.
The above “samtools faidx” command creates an index file “GRCh38.p13_ref.fna.fai”
for the reference genome with the same name as that of the reference genome but with
“.fai” appended to the file name. For the FASTA file, an fai index file is a text file consist-
ing of lines, each with five TAB-delimited columns, including NAME (name of this refer-
ence sequence), LENGTH (length of sequence), OFFSET (sequence’s first base in bytes),
LINEBASES (the number of bases on each line), and LINEWIDTH (the number of bytes
in each line) as shown in Figure 2.4.
Remember that before you use a reference genome with any aligner, you must index it
with “samtools faidx” as above, and the FASTA file and the index file must be in the same
directory. In some reference genome sequence, the sequence names are labeled by chromo-
somes (e.g., chr1) instead of accession numbers.
In the following, we will discuss the commonly used algorithms for read alignments
and the popular aligners.

FIGURE 2.4 Part of the fai index file of the human reference genome.
 Mapping of Sequence Reads to the Reference Genomes   ◾    55

2.2 READ MAPPING

Sequence alignment is one of the most important tasks in bioinformatics. It has been used
for decades before the recent revolution in genomics which erupted after the discovery
of the high-throughput sequencing technologies. It has been used in numerous bioinfor-
matics applications. The basic alignment algorithms are the global sequence alignment
using Needleman-Wunsch [1, 2] algorithm and the local sequence alignment using Smith–
Waterman (SW) [3] algorithm. Both these two legacy algorithms use dynamic programing,
which was invented by the mathematician Richard Bellman in early 1950s. The dynamic
programming in sequence alignment, which has nothing to do with computer program-
ming, is an algorithm that attempts to find the optimal alignment of two sequences by
dividing the whole alignment into smaller sub-alignments. The global sequence align-
ment is to align two sequences from end to end, and therefore, gaps can be introduced to
stretch the sequences to be equal in length. It is obvious that this type of alignment does
not work for aligning the massive number of short sequences to a genome sequence that
may span millions of bases. The second type is the local sequence alignment, which aligns
two sequences only to the regions of similarity. It is quite clear that this kind of sequence
alignment will work for aligning reads to a reference genome. The local sequence align-
ment in its crude form is exhaustive in a way that if you a have a single sequence and you
want to align it to multiple sequences, the algorithm will exhaust all possible alignments
and then report the results. In most cases, this approach is not practical for aligning mil-
lions of reads and it is computationally expensive with that massive number of genomics
sequences. The Basic Local Alignment Search Tool or BLAST [4] is heuristic version of the
SW local sequence alignment algorithm, which is more time-efficient than the exhaustive
approach because it searches only for the most significant patterns in the sequences and
performs sequence alignment accordingly. BLAST is fast but it may skip some alignments
because of its heuristic nature. Now after all, we can ask the question: Is BLAST suitable for
aligning millions of short-read sequences to a reference genome that may span millions of
bases? To provide a practical answer to this question, assume that BLAST could align a sin-
gle read to the reference genome in a second; if we have 20 million reads, BLAST will align
them to the reference genome in 20 million seconds (231.5 days), if we ignore the possibility
of the power or Internet outage or any other natural disaster. In conclusion, BLAST is not
efficient for read sequence alignment. Considering the large size of the reference genome
sequences and the large number of reads, one of the biggest challenges to perform search-
ing and locating the mapped region with high accuracy is indexing. Efficient indexing algo-
rithms are needed to organize the sequence of the reference genome and the short reads in
memory efficiently and to facilitate fast searching for patterns. Computer scientists play a
key role in developing data structures that allow computer programs to store and process
data effectively with less computational costs. Efficient data structures will help in both
indexing the sequence of a reference genome and fast searching of a read in an indexed
sequence with efficient memory usage. Several data structures for indexing were proposed
and implemented on software packages for read sequence alignment. The most com-
monly used data structures for indexing include suffix tree, suffix array, Burrows–Wheeler
56   ◾    Bioinformatics

transform (BWT), and Full-text Minute-space (FM-index). Aligners implement a variety

of searching algorithms to determine where short reads originated in the indexed refer-
ence genome sequence. Aligning of a read to a genomic location depends on the sequence
similarity. However, a good aligner should expect mismatches due to the sequencing errors
and genetic variation between the reference genome and a sequenced individual. In the
following, we will discuss these commonly used data structures for indexing and popular
searching methods.

2.2.1 Trie
A trie or a prefix tree is a data structure that is used for fast retrieval on large datas-
ets such as looking up sequencing reads. The name was derived by E. Fredkin in 1960
from the phrase “Information Retrieval” by Fredkin (1960) [5]; however, the idea was
first described by the Norwegian mathematician Axel Thue in 1912. A trie is an ordered
tree that can be used to represent a set of strings (reads) over a finite alphabet set, which
is A, C, G, and T for the DNA sequences. It allows reads with common prefixes to use
that prefix and stores only the ends of the reads to indicate different sequences. A trie is
represented by nodes and edges. The root node is empty and then each node in the trie
represents a single character. The maximum number of possible children of a node is four
in case of DNA. The children of a node are ordered alphabetically. The leaf nodes are the
nodes that represent the last characters of reads. A read is represented by the path from
the root node to a leaf node.
The trie data structure, as it is, is not suitable for indexing a genome sequence since
it stores a set of string. However, the generalized idea of the trie is used in the suffix tree
which is widely used to index a reference genome sequence.

2.2.2 Suffix Tree

The suffix tree, as a generalization of the trie data structure, is basically used for pattern
matching and finding substrings in a given string of characters. It is constructed as key-
value pairs (as the python dictionary) where all possible suffixes of the reference sequence
are the keys and positions (indexes) in the sequence as their values. In 1995, Esko Ukkonen
proposed a linear-time algorithm called Ukkonen’s algorithm [6], which constructs a suf-
fix tree in a linear time complexity that the time taken for the indexing increases linearly
with the increase in the number of nodes O(n).
For the sake of simplicity, we will try to show you how to build a suffix tree for a refer-
ence sequence and how mapping of the reads is carried out. Assume that our reference
sequence consists of 10 bases as “CTTGGCTGGA$”, where the positions of the bases
are 0, 1, 2, …, 9 and $ is an empty trailing position. The first step of constructing a suf-
fix tree is by forming suffixes (keys) and indexes (values) pairs from the sequence. We
begin from the last character in the sequence, which is the empty character “$” in the
position 10 in the sequence. Then, we form the suffix “A$”, whose first character is in
position 9 in the sequence. We continue this way until all possible suffixes are created
as shown below:
 Mapping of Sequence Reads to the Reference Genomes   ◾    57

$ 10
A $ 9
GA $ 8
GGA $ 7
TGGA $ 6
CTGGA $ 5
GCTGGA $ 4
GGCTGGA $ 3
TGGCTGGA $ 2
TTGGCTGGA $ 1
CTTGGCTGGA $ 0

Notice that each line includes a suffix (key) and a position (value). Then from the key-value
pairs, we can construct a tree made up of nodes and edges. The positions (values) will be
the nodes and suffixes (keys) will be the edges of the tree. The suffix tree is built as shown
in Figure 2.5, starting from the first suffix on the top and it moves down to make branched
nodes and edges to avoid repeating common characters. This way, we will construct a
­suffix tree with nodes and edges. An entire reference genome can be divided into suffixes
and stored this way with both suffixes and indexed positions in the unbranched nodes so
finding a pattern or a position of a read in the reference genome will be easy.
Once the reference sequence is indexed using the suffix tree, one of several searching
algorithms can be used to find the location where a read maps. For instance, to find “TGG”
in Figure 2.5, we will start searching from the root looking for “T”, and from the next node,
we will look for “GG”; thus, since there are two leaf nodes with the indexes 2 and 6, that
means “TGG” is aligned to the reference sequence in the positions 2 and 6 as shown by the
red color (Figure 2.5).

2.2.3 Suffix Arrays

The suffix array (SA) is similar to the suffix tree for pattern matching and finding sub-
strings (reads) and it can be constructed from the suffix tree. It is basically a sorted array

FIGURE 2.5 Nodes and edges of the suffix tree.

58   ◾    Bioinformatics

FIGURE 2.6 Suffix array.

of all suffixes of a given sequence. There is a variety of algorithms used for constructing
the suffix array implemented by software packages [7]. In its simplest form, a suffix array
is constructed for a sequence or a string. For the string (sequence) S with a length N and
the characters (bases) indexed as 1, 2, 3, …, N, we can construct an array for all suffixes or
substrings of the sequence as S[1…N], S[2...N],…, S[N...N] and then sort the suffixes lexi-
cographically. For instance, for the sequence S = “CTTGGCTGGA$”, we can construct the
arrays of suffixes and sort them as shown in Figure 2.6.
Figure 2.6 shows how the suffix array is constructed from sorted suffixes. The numbers
are the positions in the sequence. The sorting of suffixes lets suffixes beginning with the
same string of characters to appear one after the other and that allows a fast lookup when
we try to find exact matches of a read (substring). For instance, the exact match for the
reads “TGG” can be found by jumping to the sorted suffixes that begin with “TGG” and
we can fast locate the positions 7 and 3 (the coordinate begins from 1 and not 0 as in suf-
fix tree). When a reference genome is indexed using the suffix array, finding a position of
a pattern or a read will have a linear time complexity. A major drawback of using suffix
arrays is that they require a large memory storage depending on the size of the genome
being used. STAR [8] is as an example of the aligners that use suffix arrays.

2.2.4 Burrows–Wheeler Transform

The Burrows–Wheeler Transform or shortly BWT is a data structure algorithm that trans-
forms a string (sequence) into a compressible form that allows fast searching. The BWT is
used by BWA [9], Bowtie [10], and other aligners. BWT algorithm has been used by Unix
and Linux for data compression as bzip2 compression utility.
Let S be a sequence of characters of a sequence or string of a length n. For a genomic
sequence, a sequence is made up of A, C, G, T, or N for an ambiguous base. Let $ also be a spe-
cial single character as a trailing empty end of the sequence S (e.g., S = “CTTGGCTGGA$”).
The Burrows–Wheeler transform of the sequence S or bwt(S) is computed by generating
cyclic rotations of the sequences, as shown in the first column of Figure 2.7. The cyclic
 Mapping of Sequence Reads to the Reference Genomes   ◾    59

FIGURE 2.7 Cyclic rotations and sorted rotation.

rotations then are sorted alphabetically to form a matrix called BWM (Burrows–Wheeler
Matrix) as shown in the second column of Figure 2.7. The last column of characters (in
red) in the BWM is the BWT of the sequence. As shown in Figure 2.8, the BWT of the
sequence S is “AGG$GGTTCTC”. When an aligner uses BWT, it transforms the entire
genome sequence into a BWT. However, computing a BWT using the above naïve approach
is computationally expensive with O(n2 log n) time complexity. Instead, we can use the
suffix array to compute the BWT of a reference genome in O(n) time complexity. This
strategy is utilized by the aligners that use BWT. Constructing a BWT from a suffix array
is straightforward and very simple. To get the ith character of the BWT of the string S,
whose characters are indexed as i=1, 2, …, n, from the suffix array (A), simply use the
following rule:

BWT(S)[i] = S[A[i] - 1] if A[i] > 1 else S[n]

Figure 2.8 shows the indexes i=1, 2, 3, …,11 (first row) of the sequence S (second row) and
suffix array index A (third row), as shown in Figure 2.7. Now to infer the BWT from A,
let us begin from BWT(S)[11]. By applying the rule, BWT(S)[11] = S[A[11] - 1] = S[10] = A.
Likewise, BWT(S)[10] = S[A[10] - 1] = S[9]= G; BWT(S)[6] = S[A[6] - 1]=S[5]=G; but BWT(S)
[1], A[1] is less than 1; hence, BWT(S)[1] = $. For the BWT(S)[i] corresponding to A=9, 5, 8,
4, 7, 3, 2, we will continue using BWT(S)[i] = S[A[i] - 1]. The BWT is shown in Figure 2.9.
The question that comes up is why we need to transform a sequence of a reference
genome into BWT? The BWT serves two purposes; first, BWT groups the characters of
the sequence so that a single character appears many times in a row because the column
is sorted alphabetically and that can be used for sequence compression, which reduces the
memory storage. In this sense, the BWT is compressible and reversible. The second pur-
pose is that BWT is a data structure that can be used for indexing the sequence of a refer-
ence genome to make finding a position of a read fast.
The property of the BWT is that we can reverse it to obtain the BWM by using the last-
to-first column mapping property or simply as LF mapping, where L is the last column
60   ◾    Bioinformatics

FIGURE 2.8 Constructing BWT from Suffix array.

FIGURE 2.9 L and F columns.

in the sorted suffixes that corresponds to the BWT and F is the first column in the sorted
rotations. The L and F columns are separated as shown in Figure 2.9.
We need only the first (F) and last (L) columns to reverse the BWT to the original sorted
suffixes of the string. The BWT can be recovered using the LF mapping.
Since F must begin with “$”, we know that this character is the last one in the original
string. The “$” is preceded by the first character in L, which is “A”; so now we could infer
the last two characters “A$” in the original sequence. To infer the character that comes
before “A”, let us find the first “A” in F, which is in the second row, so the character in the
second row of L which is “G” is the character that comes before “A$”; so now we could
infer the last three characters “GA$” of the original string. Since this “G” is the first “G”
in L, it points to the first “G” in F (5th row); thus, the character that comes before “GA$”
is the character in the 5th row of L, which is “G”. So now the inferred last four characters
are “GGA$”. Since this “G” is the third “G” in L, it points to the third “G” in F, which is in
the 7th row, and the character in the 7th row of L, which is “T”, is the character that comes
before “GGA$” so the inferred last five characters are “TGGA$”. The rank of this “T” is
the first “T” in L which points to the first “T” in F in the 9th row, so the character which
comes before “TGGA$” is “C” in the 9th row of L. Thus, the inferred last 6 characters of
the original string are “CTGGA$”. We can continue using this L-to-F mapping until we
recover the entire sequence “CTTGGCTGGA$” as indicated by the arrows in Figure 2.10.
This reversion process is called backward search mechanism since it begins from the very
end of the string and moves backward. A computer algorithm usually performs this task.
We can recover the original string using the same LF mapping property but in a dif-
ferent way. Given the LF matrix, the string S is reconstructed starting from the end of the
 Mapping of Sequence Reads to the Reference Genomes   ◾    61

FIGURE 2.10 LF mapping to reverse the BWT.

FIGURE 2.11 Using LF mapping to reverse the sorted rotations.

string, $, which is also the first character of F column. The character before $ will be the
first character of L column. In general, the character which comes before can be inferred
with the pairs L and F relationships as shown in Figure 2.10. These relationships are also
illustrated in Figure 2.11 which shows that, first, we recover the first two columns of the
BWM by creating a two-column matrix from columns L and F and sorting it as shown
in Figure 2.11a. The third column is recovered by creating a three-column matrix from
column L and the first two recovered columns and then we sort it (Figure 2.11b). The same
is repeated to recover the fourth column of the sorted rotations as shown in Figure 2.11c.
We will continue like this until we recover all columns and the BWM will be as shown
in Figure 2.11d. The original string is the one ends with the character “$” in the BWM as
indicated in the shaded row in Figure 2.11.
The steps shown in Figures 2.10 and 2.11 are to show how a BWT is reversible. However,
there are several algorithms implemented by software for reversing a BWT using LF
mapping. The BWT is used as data structure for indexing a genome (compressed or
uncompressed) by transforming the entire genome into a BWT. Once a genome has been
transformed, there are several searching algorithms and approaches for finding the posi-
tion of a substring (read) on the reference genome.
62   ◾    Bioinformatics

2.2.5 FM-Index
FM-index (Full-text Minute-space index) performs a backward search mechanism on the
top of BWT to find exact pattern matches on a string with a linear computational time
complexity O(N), regardless of the length of the string. Moreover, it could achieve high
compression ratios that allow the indexing of the whole human genome into less than 2.0G
of memory storage [2]. An FN-index searching is performed by using LF mapping matrix.
Searching for a read (a substring) with FM-index uses the backward matching in which the
LF mapping is used repeatedly until matches for the substring are found. FM-index intro-
duces the rank table and lookup table. A rank table is generated from the BWT (column L)
using the character’s rank, which is defined as the number of times a character occurs pre-
viously in the prefix L[1…k] (i.e., lists the occurrence and order of each unique character).
The rank table acts as a function B(c, k), which is explained as follows: if a character c and
a prefix of length k are given, then we can infer the number of times that character occurs
in the prefix. A lookup table lists the index of the first occurrence of each character c from
the first column (column F) of the sorted matrix. The first occurrence of the character c in
the lookup table can be described as C[c].
Figure 2.12 shows Burrows–Wheeler matrix (a), the rank table (b), and the lookup table
(c). Using the rank table and lookup table, the LF mapping can be described as

LF(i) = C[L[i]] + B(L[i], i)

The original string can be recovered using the above rule and backward mechanisms are
as follows:
The last character in the string is “$” on the first row in column F, preceded by “A” on
the first row in column K. So, the last two characters are “A$”. Now we can use the above
rule to find the character that comes before “A$”. Since “A” is on row 1 in column L, we can
locate that character in column F as LF(1) as follows:
LF(1) = C(A] + B(A, 1) = 1 + 2 =2 : The character is on row 2 in column F. So, “A” is pre-
ceded by “G”, which is on row 2 in column L. The last three characters are “GA$”.
LF(2) = C(G] + B(G, 2) = 4 + 1 =5 : The character is on row 5 in column F. So, “G” is pre-
ceded by “G”, which is on row 5 in column L. The last four characters are “GGA$”.
LF(5) = C(G] + B(G, 5) = 4 + 3 =7 : The character is on row 7 in column F. So, “G” is pre-
ceded by “T”, which is on row 7 in column L. The last four characters are “TGGA$”.
LF(7) = C(T] + B(T, 7) = 8 + 1 =9 : The character is on row 9 in column F. So, “T” is pre-
ceded by “C”, which is on row 9 in column L. The last four characters are “CTGGA$”.
We can continue using that rule with the rank table and lookup table until we recover
the entire spring “CTTGGCTGGA$”.
The FM-index can also be used as a pattern searching technique that operates on the
BWT to map read sequences to locations on the BW-transformed reference genome.
Searching for a pattern is a backward process like recovering the original string; a
­character is ­processed at a time, beginning with the last character of the pattern (read
sequence) [11].
 Mapping of Sequence Reads to the Reference Genomes   ◾    63

FIGURE 2.12 (a) BWT, (b) rank table, and (c) lookup table.

2.3 READ SEQUENCE ALIGNMENT AND ALIGNERS

For read mapping, we usually have millions of reads produced by a high-throughput
instrument and we wish to determine the origin of each of the reads in the sequence of
a reference genome. For most of sequencing applications, the read alignment or mapping
is the slowest and the most computationally expensive step. This is because mapping pro-
grams will attempt to determine the most likely points of origin for each read with respect
to a reference genome. Mapping the reads produced from eukaryotic RNA-Seq requires
extra efforts by aligners. In eukaryote, the coding regions (exons) of the genes are sepa-
rated by non-coding regions (introns). Since only transcriptome or gene transcripts are
targeted in the RNA sequencing, the aligners used for mapping RNA-Seq reads must be
aware of the non-contiguous nature of the exons and the challenge of the detection of the
splicing regions.
Indeed, before performing alignment, we need to download the sequence of the refer-
ence genome of the species studied and then index the reference genome so that the loca-
tions, where reads maps, can be found easily upon the process of searching and alignment.
For most of the aligners, indexing of the reference genome is the first step before per-
forming read mapping. Above, we discussed the most commonly used indexing methods
for storing and organizing the reference genome sequence so it can be easily searched to
determine the locations (coordinates) of aligned reads. There is another challenge faced
by aligners; the reads produced by sequencers may not be exactly aligned to a location in
the reference genome sequence because of base call errors or may not be naturally due to
mutations (substitutions, deletions, or insertions) in the DNA sequences of the individual
64   ◾    Bioinformatics

FIGURE 2.13 The general workflow of the read mapping.

studied. To overcome this challenge, aligners must adopt a strategy to perform the gapped
alignment rather than performing only the exact alignment.
Read mapping is required by the most sequencing applications including reference-based
genome assembly, variant discovery, gene expression, epigenetics, and metagenomics. As
shown in Figure 2.13, the workflow of read mapping/alignment includes downloading the
right FASTA file of the reference genome of the species studied, indexing the sequence of
the reference genome with “samtools index” command, indexing the reference genome
with an aligner, and finally performing mapping of the cleaned reads with the aligner
itself. Remember that, before mapping, the step of quality control must be performed as
discussed in Chapter 1. So, when we move to the step of read mapping, we should have
already cleaned up the reads and fixed most of the failed metrics shown by the QC reports.
Even if we couldn’t fix all errors, we should also be aware of the effect of that in the final
results.
We have also discussed above how to download a reference genome of an organism and
how to index it. To avoid repetition, we will delve into read mapping and generation of
SAM/BAM files without covering the topics that we have already discussed.
Read mapping is the process of finding locations on a reference genome where reads,
contained in FASTQ files, map. The read mapping information are then stored in a
SAM/BAM file format, which is a special file format for storing sequences alignment
 Mapping of Sequence Reads to the Reference Genomes   ◾    65

information. Most aligners are capable of performing both exact matching and inexact
matching, which are essential to find the locations of reads that may have some base call
errors or varied genetically from the reference genome. The different aligners implement
different ­algorithms to perform both kinds of lookups in the indexed reference genome
stored in data structures like suffix tree, suffix array, hashing table, and BWT. While the
exact lookup is straightforward, the inexact matching uses sequence similarity to find the
most likely locations where a read is originated. Although there are different ways to mea-
sure sequence similarity, most aligners used Hamming distance [12] or Levenshtein dis-
tance [13] to score the similarity between a reads and portions of the reference genomes
based on a threshold. Some aligners use the seed-and-extend strategy to extend a seed (an
exact matched substring) across multiple mismatched bases to allow mapping reads with
base call errors or variations. Most aligners employ seed-and-extend strategy on the local
sequence ­alignment using SW algorithm. Seeds are created by making overlapping k-mers
(substrings or words of length k) from the reference genome sequence. Some aligners like
Novoalign [14] and SOAP [15] index k-mers with the trie or hash table data structures for
a quick search.

2.3.1 SAM and BAM File Formats

Almost all read aligning programs (aligners) store alignment information of the reads
mapped to the reference genome in a Sequence Alignment and Map (SAM) file or Binary
Alignment and Map (BAM) file, which is the binary form of SAM. The SAM file is a read-
able plain text file for storing biological sequences mainly aligned to a reference sequence
[16] but it can also contain unmapped reads. It is a TAB-delimited text file consisting of
two main sections: (i) a header section and (ii) an alignment section.
The header section of the SAM file is optional, and when it is present, it must be before
the alignment section. Each line in the header section must start with “@” symbol followed

FIGURE 2.14 A header section of a SAM file.

66   ◾    Bioinformatics

by one of the two-letter header record type codes. A record type code may have two-letter
subtype codes. Table 2.1 lists and describes the two-letter codes of the SAM header section,
and Figure 2.14 shows an example header section. Notice that the SAM file begins with
“@HD VN:1.0 SO:coordinate”, which indicates that the specification of SAM version 1.0
was used and the alignments in the file are sorted by the coordinate. We can also notice
that there are several @SQ header lines, each line is for the reference sequence used for the
alignment. The @SQ header includes the sequence name (SN), which is the chromosome
number, and sequence length (LN). The last two lines in the header section should include
@PG, which describes the program used for the alignment, and @CO, which describes the
command lines.
The alignment section begins after the header section. Each alignment line has 11
­mandatory fields to store the essential alignment information. The alignment section may
have variable number of optional fields, which are used to provide additional and aligner
specific information.
Figure 2.15 shows a partial alignment section of a SAM file. The columns of the align-
ment section are split because they do not fit the page. Table 2.2 lists and describes 11 man-
datory fields of the SAM alignment section.
These 11 mandatory fields are always present in a SAM file. If the information of
any of these mandatory fields is not available, the value of that field will be replaced
with “0” if its data type is integer or “*” if the data type is string. Most field names are
self-explanatory.

TABLE 2.1 The Two-Letter Codes of the Header Section and Their Description
Code Header Code Description
@HD This header codes for metadata, and if it is present, it must be the first line of the SAM file. This
header line may include subtypes: VN for format version, SO for sorting order, GO for
grouping alignment, and SS for sub-sorting order of alignments
@SQ This is for the reference sequence used for aligning the reads. A SAM file may include multiple
@SQ lines for the reference sequences used. The order of the sequences defines the order of
alignment sorting. The two most common sub-type codes used in this header line include SN
for reference sequence name and LN for reference sequence length
@RG This header line is used to identify read group and it is used by some downstream analysis
programs (e.g., GATK) for grouping files based on the study design. Multiple lines can exist in
a SAM file. This line may include the ID for the unique read group identifier, BC for the
barcode sequence identifying the sample, CN for the name of sequencing facility, DS for
description to be used for the read group, DT for the date of sequencing, LB for the
sequencing library, PG for the programs used for processing the read group, PL for the
platform of the sequencing technology used to generate the reads, PM for the platform model,
PU for the platform unit, which is a unique identifier (e.g., flow cell/slide barcode), and SM
for the sample identifier, which is the pool name where a pool is being sequenced
@PG This is the header line for describing the program used to align the reads. It may include ID for
the program unique record identifier, PN for the program name, CL for the command line
used to run the program, PP for the previous @PG-ID, DS for description, and VN for the
program version
@CO This is the header line for a text comment. Multiple @CO lines are allowed
 Mapping of Sequence Reads to the Reference Genomes   ◾    67

TABLE 2.2 Eleven Mandatory Fields of the Alignment Section of the SAM File
Column Field Description
1 QNAME The query sequence name (string)
2 FLAG Bitwise flag (integer)
3 RNAME Reference sequence name (string)
4 POS Mapping position from the leftmost first base (integer)
5 MAPQ Mapping quality (integer)
6 CIGAR CIGAR string (string)
7 RNEXT Reference name of the mate or next read (string)
8 PNEXT Position of the mate or next read (integer)
9 TLEN Sequence length (integer)
10 SEQ The read sequence (string)
11 QUAL ASCII code of the Phred-scaled base (Phred+33) (string)

FIGURE 2.15 An alignment section of a SAM file.

The QNAME field is to show the read sequence name, which is obtained from the FASTQ
file. FLAG is a bitwise integer that describes the alignments (e.g., paired, unaligned, and
duplicate). The description is stored as codes as shown in Table 2.3.
The FLAG field in the SAM file may have one of the decimal values listed in Table 2.3
or the sum of any of those decimal values. For instance, Figure 2.16 shows the FLAG for
the read “SRR062634.6862698” is 99, which means that this FLAG combines 4 conditions:
1+2+32+64=99. This means that the read maps to that position of the reference genome are
described as follows: the read is paired (1), the aligner mapped the two pairs properly (2),
the next sequence (SEQ) is a reverse strand (32), and first read in the pair (64). Instead of
doing mental math to figure out a such FLAG number, we can use “samtools flags” com-
mand as follows (Figure 2.16):

samtools flags 99
samtools flags

The RNAME field shows the reference sequence name of the alignment such as a chromo-
some name (e.g., 1). This field will be filled with “*” for unmapped read.
68   ◾    Bioinformatics

FIGURE 2.16 Using “samtools flags” to convert numeric FLAG representations into textual.

The POS field specifies the leftmost mapping position in the reference genome of the
first base of the aligned read. For the sequence “SRR062634.6862698”, the alignment posi-
tion is 10001 as shown in Figure 2.15. The POS value of the unmapped read is “0”.
The MAPQ field specifies the mapping quality in Phred scoring unit. If the quality score
is not available, the value will be set to 255.
The CIGAR field contains the CIGAR string, which is a sequence of the base length and
the associated operations. It is used to indicate where, in the aligned read, an operation
like match/mismatch, deletion, or insertion took place. Table 2.4 lists the CIGAR opera-
tions that can be associated with a read alignment. For instance, in the SAM file shown in
Figure 2.15, the sequence “SRR062634.6862698” has the CIGAR string “30M70S”, which
means that the first 30 bases of that sequence match (30M) bases in the reference sequence
and 70 bases showing soft clip (70S), which is special mismatch due to error in the reference
sequence.
The RNEXT field contains the reference sequence name of the primary alignment of the
NEXT read in the template. If it is not present, the field will be set to “=” or “*”. The PNEXT
field specifies the position of the primary alignment of the NEXT read in the template. This
field will be set to “0” when the information is unavailable.
The TLEN field specifies the signed observed template length. This field will be positive
for the leftmost read, negative for the rightmost, and undefined for any middle read.
The SEQ field contains the read sequence. This field will be set to “*” if the sequence
was not stored. Otherwise, the sequence must be equal to the sum of lengths of the opera-
tions in CIGAR string. The “=” symbol indicates that the base of the read is identical to
the reference base. The mapped reads are aligned to the forward genomic strand (plus
strand). Therefore, if a read is mapped to the reverse strand (minus strand) of the refer-
ence sequence, the SEQ field will contain the complementary strand of the unmapped plus
strand.
The QUAL field contains the Phred quality scores in ASCII character (Phred+33). The
string in this field is the same as the string in the FASTQ file. This field may be set to “*” if
 Mapping of Sequence Reads to the Reference Genomes   ◾    69

TABLE 2.3 The FLAG Bitwise Decimal and Hexadecimal Numbers and Their Descriptions
Decimal Hexadecimal Description of Read
1 0x1 The read is paired
2 0x2 The aligner mapped the two pairs properly
4 0x4 The read is unmapped
8 0x8 Next segment in the template is unmapped
16 0x10 The sequence in SEQ is a reverse strand (minus strand)
32 0x20 The next sequence (SEQ) is a reverse strand
64 0x40 First read in paired reads
128 0x80 Second read in paired reads
256 0x100 The alignment is secondary
512 0x200 The read fails platform/vendor quality checks
1024 0x400 The read is PCR or optical duplicate (technical sequence)
2048 0x800 The alignment is supplementary

TABLE 2.4 CIGAR Operations and Descriptions

Operation Description
M Alignment match, which can be a sequence match or mismatch
I Insertion to the reference sequence
D Deletion from the reference sequence
N Skipped region from the reference sequence
S Soft clip on the read (present in SEQ)
H Hard clip on the read (not present in SEQ)
P Padding (silent deletion from the padded reference sequence)
= Sequence match
X Sequence mismatch

the quality string is not stored. Otherwise, it must be equal to the length of the sequence
in SEQ.
The alignment section of a SAM file may contain a number of optional fields. Each
optional field is defined by a standard tag accompanied with a data type and a value in the
following format:
TAG:TYPE:VALUE
The TAG is a two-character string. There are several predefined standard tags for SAM
optional fields. The complete list is available at “https://samtools.github.io/hts-specs/
SAMtags.pdf”. The user is allowed to add a new tag.
The TYPE is a single character defining the data type of the field. It can be “A” for the
character data type, “B” for general array, “f” for real number, “H” for hexadecimal array,
“i” for integer, and “Z” for string.
VALUE is the value of the field defined by the tag data type.
Notice that the last four columns in the SAM file shown in Figure 2.16 are for optional
fields identified by the four predefined standard tags: “NH”, “HI”, “AS”, and “NM”. The
“NH” tag shows the number of reported alignments (number of hits) that contain the read
70   ◾    Bioinformatics

sequence in the current record. The “HI” tag shows the index of the query hit. The “AS” tag
shows the alignment score defined by the aligner. The “NM” tag shows the edit distance,
which is defined as the minimal number of single-nucleotide edits (substitutions, inser-
tions, and deletions) needed to transform the read sequence into the aligned segment of
the reference sequence.
For more details about SAM file, read the specification of the Sequence Alignment/Map
file format, which is available at “https://samtools.github.io/hts-specs/”.

2.3.2 Read Aligners

There are several aligners available for mapping reads to a reference genome. However,
next we will discuss only BWA [9], Bowtie [17], and STAR [8] as examples. The use of other
aligners is similar.
When we move to read mapping step, we will have already had cleaned the raw reads
in FASTQ files as discussed in Chapter 1. In the following sections, we will show you
how to map reads contained in FASTQ files to a reference genome. For this purpose, we
will download FASTQ files (run # is SRR769545) from the NCBI SRA database. To avoid
repeating the quality control steps, we will assume that we have cleaned the FASTQ files
following the steps in Chapter 1 and the files are ready for mapping. Our example raw data
are paired-end reads from the 1000 Genomes whole exome sequencing of an individual
from the Great Britain population. We can download the two FASTQ files (forward and
reverse) using the SRA-toolkit. The following commands create the directory “data” where
the two FASTQ files will be downloaded:

mkdir data
cd data
fasterq-dump --verbose SRR769545

The size of each file is 11G; the two files take around 22G of storage space. To save storage
space, we can compress these files using gzip utility, which will reduce each file to only
2.6G. Most aligners accept the gzipped FASTQ files.

gzip SRR769545_1.fastq
gzip SRR769545_2.fastq

The “.gz” will be added to the name of each file to indicate that the two files were com-
pressed with gzip.
We can also run FastQC and display the QC reports as follows:

fastqc SRR769545_1.fastq.gz SRR769545_2.fastq.gz

firefox SRR769545_1_fastqc.html SRR769545_2_fastqc.html

The per base quality reports for the two FASTQ files are shown in Figure 2.17. We can
notice that the reads in the two files have a good quality.
 Mapping of Sequence Reads to the Reference Genomes   ◾    71

FIGURE 2.17 The per base quality reports for the reads in the two FASTQ files.

In Section 2.1, we showed how to download the FASTA file of the reference genome
sequence of an organism and how to index it using “samtools faidx”. So, if you did not do
that, follow the steps in that section to download the human reference and then to index
it. The sequences of reference genomes can also be downloaded from other databases such
as UCSC database. We have also downloaded and compressed example paired-end FASTQ
files for practice. The next step is to show you how to use an aligner (BWA, Bowtie, and
STAR) for read mapping.

2.3.2.1 Burrows–Wheeler Aligner

The Burrows–Wheeler Aligner (BWA) is a sequence aligner that uses BWT and FM-index.
We can install the latest version of the BWA software by following the installation instruc-
tions at “https://github.com/lh3/bwa” or we can use the following commands:

git clone https://github.com/lh3/bwa.git

cd bwa; make

The above will clone the BWA source files into your working directory and then it will
compile it. Once BWA has been installed successfully, you may need to set its path so that
72   ◾    Bioinformatics

you will be able to use it from any directory. While you are in the “bwa” directory run the
“pwd” command to print the absolute path of BWA, copy it, then change to your home
directory, and open “.bashrc” file using “vim” or any text editor of your choice:

cd $HOME
vim .bashrc

Add the following to the end of the “.bashrc” file:

export PATH=”your_path/bwa”:$PATH

Do not forget to replace “your_path” with the path to the “bwa” directory on your com-
puter. Save the “.bashrc” file, exit, and restart the terminal for the change to take effect.
Type “bwa” on the terminal and press the enter key. If the BWA software was installed and
added to the path correctly, you will see the help screen.
BWA has three alignment algorithms: BWA-MEM “bwa mem”, BWA-SW “bwa bwasw”,
and BWA-backtrack “bwa aln/samse/sampe”. Both “bwa mem” and “bwa bwasw” algo-
rithms are used for mapping short and long sequences produced by any of the sequenc-
ing technologies. The “bwa aln/samse/sampe” also called BWA-backtrack is designed for
Illumina short-sequence reads up to 100 bp. Among the three algorithms, “bwa mem” is
the most accurate and the fastest.
Indeed, aligning read sequences to a reference genome with BWA requires indexing the
reference genome using “bwa index” command. We can use this command to index the
human reference genome which was downloaded and indexed with “samtools faidx” above
as follows:

bwa index GRCh38.p13_ref.fna

The indexing will take some time depending on the size of the reference genome and the
memory of your computer. When the “bwa index” command finishes indexing, it will dis-
play the information, including the number of iterations, the elapsed time in second, the
indexed FASTA file name, and the real time and CPU time taken for the indexing process.
The indexing of the human genome may take up to six hours on a desktop computer of
32G RAM.
The BWA indexing process creates five bwa index files with extensions “.amb”, “.ann”,
“.bwt”, “.pac”, and “.sa”. The total storage space for the current human reference genome
and its index files is around 9.4G.
The “.amb” file indexes the locations of the ambiguous (unknown) bases in the FASTA
reference file that are flagged as N or another character but not as A,C,G, or T. The “.ann”
file contains annotation information such as sequence IDs and chromosome numbers.
The “.bwt” is a binary file for the Burrows–Wheeler transformed sequence. The “.pac” is a
binary file for the packed reference sequence. The “.sa” is also a binary file containing the
suffix array index. For mapping read sequences to the reference genome, all these five files
must be together in the same directory.
 Mapping of Sequence Reads to the Reference Genomes   ◾    73

So far, we have two directories: “data” where the FASTQ files were downloaded and
“refgenome” where the FASTA sequence of the reference genome was downloaded and
indexed.
In the next step, we will use any of the alignment algorithms to map the reads to the
human reference genome. The “bwa mem” algorithm can be tried first as recommended
by the BWA.
1-BWA-MEM algorithm
In the following, the “bwa mem” command maps the reads in the FASTQ files to the
human reference genome and saves the output file, which contains the mapped reads in a
new directory called “bwa_sam”. Before running the commands, make sure that the work-
ing directory is just one level out of the two directories.

mkdir sam
bwa mem \
-t 4 \
refgenome/GRCh38.p13_ref.fna \
data/SRR769545_1.fastq.gz \
data/SRR769545_2.fastq.gz \
> sam/SRR769545_mem.sam 2> sam/SRR769545_mem.log

The “bwa mem” command has the capability of mapping reads of a length ranging between
70 bp and 1 Mbp. The BWA-MEM algorithm uses seeding alignments with maximal exact
matches (MEMs) and then it extends seeds with the affine-gap SW algorithm. We can run
“bwa mem” on the command line without any option to learn more about “bwa mem”
command. In the above, we use “-t 4” so that the program can use four parallel threads to
perform read mapping. Then, we provided the path of the FASTA file name of the reference
genome and the two FASTQ file names. The output of the aligned reads will be redirected
to a new file “SRR769545_mem.sam” using the Unix/Linux redirection symbol “>”. The
command execution comments are redirected to the log file “SRR769545_mem.log” using
“2>”, which is used in Unix/Linux to redirect stderr (standard error) to a text file. Both the
output files will be saved in the “sam” directory. The log file contains important informa-
tion about the comments and any standard error that occurs during the mapping process.
Always open the log file if the running fails. The log file will tell you if the failure is due to
the wrong syntax, wrong file path, or any other standard error. Even if the program has
finished running normally, it is a good practice to open the log file and look at the statistics
and comments.

cd sam
less SRR769545_mem.log

You can also display the SAM file produced by “bwa mem” with the following:

less -S SRR769545_mem.sam
74   ◾    Bioinformatics

The size of this SAM file is about 19G. We can convert it to the BAM format using Samtools
to save some storage space. First, you need to change to the “sam” directory and then run
the following:

samtools view \
-uS \
-o SRR769545_mem.bam \
SRR769545_mem.sam

The new BAM file is about 15G. We can then delete the SAM file as follows to save some
storage space:

rm SRR769545_mem.sam

BWA-MEM2 is an optimized BWA-MEM algorithm that has been recently released. This
new version produces alignments identical to BWA-MEM but it is faster and the indexing
occupies less storage space and memory [18]. You can install BWA-MEM2 separately by
following the instructions available at “https://github.com/bwa-mem2/bwa-mem2”.
2-BWA-SW
The BWA-SW [9] algorithm, like BWA-MEM, can also be used for the alignment of
single- and paired-end long reads generated by all platforms. It uses SW local alignment
approach to map reads to a reference genome. BWA-SW has a better sensitivity when
alignments have frequent gaps. However, this algorithm has been depreciated by his devel-
oper since BWA-MEM is restructured for better performance. The following “bwa bwasw”
­performs read alignment as above:

bwa bwasw \
-t 4 \
refgenome/GRCh38.p13_ref.fna \
data/SRR769545_1.fastq.gz \
data/SRR769545_2.fastq.gz \
> sam/SRR769545_bwasw.sam 2> sam/SRR769545_bwasw.log

You can convert this SAM file to BAM file as we did above or you can just delete it to save
some space.
3-BWA-backtrack
The BWA-backtrack algorithm is designed for aligning Illumina short reads of a length
up to 100 bp with sequencing error rates below 2%. It involves two steps: (i) using “bwa
aln” to find the coordinates of the positions, where the short reads align, on the refer-
ence genome, and then (ii) generating alignments with “bwa samse” for single-end reads
or “bwa sampe” for paired-end reads. The base call quality usually deteriorates toward
the end of reads generated by Illumina instruments. This algorithm optionally trims low-
quality bases from the 3′-end of the short reads before alignment. Therefore, it is able to
align more reads with high error rate toward the 3′-ends of the reads.
 Mapping of Sequence Reads to the Reference Genomes   ◾    75

In the following, we will use “bwa aln” to perform the first step of the alignment. Run
“bwa aln” without any option on the command line to learn more about the usage and
options. If the quality of reads at the 3′-end is low, we can use the “-q” option with this
command to specify a quality threshold for read trimming down to 35 bp. Run the follow-
ing commands while you are one step out of the “refgenome” and “data” directories:

bwa aln \
refgenome/GRCh38.p13_ref.fna \
data/SRR769545_1.fastq.gz \
> data/SRR769545_1.sai
bwa aln \
refgenome/GRCh38.p13_ref.fna \
data/SRR769545_2.fastq.gz \
> data/SRR769545_2.sai

Then, we can use “bwa sampe” to generate the SAM file for the alignments.

$ bwa sampe \
refgenome/GRCh38.p13_ref.fna \
data/SRR769545_?.sai \
data/SRR769545_?.fastq.gz \
> sam/SRR769545_aln.sam 2> sam/SRR769545_aln.log

2.3.2.2 Bowtie2
Bowtie2 is an aligner that uses BWT and FM-index as data structures for indexing the
reference genome. It is an ultrafast, memory-efficient short read aligner, and it allows
mapping millions of reads to a reference genome on a typical desktop computer. Bowtie2
is the next generation of the original Bowtie which requires the reads to have equal length
and it does not align reads with gaps. Bowtie2 was developed to overcome those limita-
tions. It performs read mapping in four steps: (i) extraction of seeds from the reads and
their reverse strands, (ii) using FM-index for exact ungapped alignment of the seeds, (iii)
sorting the alignments by scores and identifying the alignment position on the refer-
ence genome from the index, and (iv) extending seeds into full alignments using paral-
lel dynamic programming [16]. Bowtie2 can be installed on Linux with the following
commands:

git clone https://github.com/BenLangmead/bowtie2.git

cd bowtie; make

Then, you need to set Bowtie2 path so that you can run it from any directory by editing
“.bashrc” file from your home directory.

cd #HOME
vim .bashrc
76   ◾    Bioinformatics

Add the following to the end of the file:

export PATH=”your_path/bowtie2”:$PATH

Do not forget to change “your_path” with the right path on your computer. Save the file
and exit. You may need to restart the terminal or run “source .bashrc” to make the change
active. Then, you can enter “bowtie2” on the terminal. If Bowtie2 is installed and its path
was set, help screen will be displayed.
Before read mapping, we need to use “bowtie2-build” command to index the FASTA
sequence of the reference genome. Enter “bowtie2-build” on the command line of the ter-
minal to display the help screen that shows the usage and options. The general syntax is as
follows:

bowtie2-build [options] <reference_in> <ebwt_outfile_base>

The “bowtie2-build” command requires a FASTA file of a reference genome as an input

and a prefix string which is added as a prefix to the file names of the index. The following
command indexes the human genome for Bowtie2. Before running the command, make
sure that the current working directory is a one-level out “refgenome” directory, where we
downloaded the human genome.

bowtie2-build \
--threads 4 \
refgenome/GRCh38.p13_ref.fna \
refgenome/bowtie2

The indexing may take around 25 minutes using four processors on a computer with 32G
of memory. The “bowtie2-build” command generates six index files prefixed with the pre-
fix string provided for the command. Pre-built indexes for some organisms can also be
downloaded from the official Bowtie2 website.
After indexing the reference genome, we can use “bowtie2” command to align the
paired-end reads and to generate SAM file:

bowtie2 -x refgenome/bowtie2 \
-1 data/SRR769545_1.fastq.gz \
-2 data/SRR769545_2.fastq.gz \
-S sam/SRR769545_bowtie2.sam

Instead of “-S” option to generate a SAM file, we can use “-b” option to generate a BAM
file. To learn more about Bowtie2’s options, enter “bowtie2” on the command line of the
Linux terminal.

2.3.2.3 STAR
STAR, which stands for Spliced Transcripts Alignment to a Reference, is a fast read aligner
developed to handle the alignment of massive number of RNA-Seq reads. Its alignment
 Mapping of Sequence Reads to the Reference Genomes   ◾    77

algorithm uses uncompressed suffix array (SA) data structure to perform sequential
­maximum mappable seed search, which is defined by the developer as the longest sub-
string of a read that matches exactly one or more substrings of the reference genome. This
search is achieved by mapping seeds to the reference genome. A read with a splice junction
site will not be mapped continuously. The algorithm will try to align the first unmapped
seed to a donor splice site and then it repeats the search and aligns the unmapped to an
acceptor splice site. The search is performed for forward and reverse direction. This kind
of search will help in the detection of base mismatches and InDels. If a single or multiple
mismatches are found, the matched substrings will act as anchors on the genome to allow
extension. The search is then followed by a seed clustering by proximity for determining
the anchor seeds. Then, the aligned seeds around the anchor seeds within a user-defined
window are stitched together using dynamic programming. STAR is capable of detecting
splices and chimeric transcripts and mapping complete RNA transcripts that are formed
from non-contiguous exons in eukaryotes [8].
The STAR software can be installed by following the installation instructions, which are
available at “https://github.com/alexdobin/STAR”. On Ubuntu, you can install STAR using
the following command:

sudo apt install rna-star

As most of the read aligners, STAR basic workflow includes both index generation and read
alignment. However, for index generation, both a reference genome in the FASTA format
and reference annotation file in GTF format are required. Pre-built indexes for genomes
of some species can be downloaded from the STAR official website. As discussed before,
the reference genomes can be downloaded from databases such as NCBI Assembly, UCSC
genome collection, or any other database. For the aligners discussed before, we down-
loaded the human reference genome from the NCBI Genome database. For STAR, we will
download the human reference genome and its GTF annotation file from the UCSC data-
base. The reason is that UCSC maintains the gene annotation file in GTF format. Use the
following command to create a new directory “ucscref” and then download and decom-
press the human reference genome and GTF annotation file:

mkdir ucscref
wget \
-O “ucscref/hg38.fa.gz” \
https://hgdownload.soe.ucsc.edu/goldenPath/hg38/bigZips/latest/
hg38.fa.gz
wget \
-O “ucscref/hg38.fa.gz” \
https://hgdownload.soe.ucsc.edu/goldenPath/hg38/bigZips/genes/
hg38.ncbiRefSeq.gtf.gz
gzip -d ucscref/hg38.fa.gz
gzip -d ucscref/hg38.ncbiRefSeq.gtf.gz

Then, we will build the index for the reference genome using the “STAR” command.
78   ◾    Bioinformatics

mkdir indexdir
STAR --runThreadN 4 \
--runMode genomeGenerate \
--genomeDir indexdir \
--genomeFastaFiles ucscref/hg38.fa \
--sjdbGTFfile ucscref/hg38.ncbiRefSeq.gtf \
--sjdbOverhang 100

Above, with the “STAR” command, we used “--runThreadN” to specify the number of
threads used for indexing, “--runMode genomeGenerate” to tell the command that
we wish to generate a genome index, “--genomeDir” to specify the directory where the
index files are to be saved, “--genomeFastaFiles” to specify the file path of the reference
genome FASTA file, “--sjdbGTFfile” to specify the file path of the annotation GTF file, and
“--­sjdbOverhang” to specify the length of the genomic read around the annotated junction
to be used in constructing the splice junctions database. For this option, we can provide
read size minus one (n-1) if the read size is equal for all reads; otherwise, we can provide
the maximum size minus one.
The process of indexing may take a long time and may consume much memory and stor-
age space compared to the other aligners. Several files will be generated including binary
genome sequence files, files of the suffix arrays, a text file for the chromosome names or
lengths, splice junctions’ coordinates, and transcripts/genes information. Those files are
for the STAR internal use; however, the chromosome names can be renamed in the chro-
mosome file if needed.
The next step is to use STAR command for aligning the reads. This time we will use
“--runMode alignReads” to tell the program to run read mapping mode, “outSAMtype
BAM Unsorted” to generate an unsorted BAM file, “--readFilesCommand zcat” to tell the
program that the FASTQ files are compressed, “--genomeDir” to specify the index direc-
tory, “--outFileNamePrefix” to specify the prefix for the output files, and “--readFilesIn” to
specify the FASTQ file names. You can also set “outSAMtype BAM SortedByCoordinate”
to generate a BAM file sorted by the alignment coordinates. However, that will exhaust the
memory of a 32G-RAM computer.

STAR --runThreadN 4 \
--runMode alignReads \
--outSAMtype BAM Unsorted \
--readFilesCommand zcat \
--genomeDir indexdir \
--outFileNamePrefix STARoutput/SRR769545 \
--
readFilesIn data/SRR769545_1.fastq.gz data/SRR769545_2.
fastq.gz

STAR alignment mode produces a BAM file, containing read alignment information,
and four text files, three log files with file names “*Log.out”, “*Log.progres.out”, and
“*Log.final.out”, where “*” is for the prefix specified by “--outFileNamePrefix” option,
 Mapping of Sequence Reads to the Reference Genomes   ◾    79

and a tab-delimited file with an extension “*SJ.out.tab”. The “*Log.out” file contains
detailed information about the run. The “*Log.progres.out” file contains statistics for the
alignment progress in 1-minute intervals. The “*Log.final.out” file contains important
summary statistics for the alignment after the alignment is complete. We will discuss
this later in the RNA-Seq chapter. “*Log.final.out” contains information about the splice
junctions.

2.4 MANIPULATING ALIGNMENTS IN SAM/BAM FILES

Before proceeding to the next step of the analysis workflow, we may need to manipulate the
alignments in the SAM/BAM file. There are several programs used to serve that purpose
but Samtools and PICARD are the most commonly used ones. In the following, we will
discuss how to use Samtools to manipulate alignments in SAM/BAM files. PICARD will
be discussed in the RNA-Seq chapter.

2.4.1 Samtools
Samtools is a collection of command-line utilities for the manipulation of alignments in
the SAM/BAM files. These utilities are grouped into indexing, editing, file operations, sta-
tistics, and viewing tools. Samtools can be installed on Linux by running “sudo apt-get
install samtools” or you can follow the installation instructions available at “http://www.
htslib.org/download/”. After installation, you can display the complete list of the Samtools’
utilities by running “samtools” without any option on the command line.
A Samtools utility is executed using the following format:

samtools <command> [options]

where <command> is any Samtools command and [options] is any of the command
options.
The most used commands include “cat”, “index”, “markup”, “rmdup”, “sort”, “mpileup”,
“coverage”, “depth”, “flagstats”, and “view”.
To learn more about the usage and options of any of these commands, you can use the
following format:

samtools <command>

For instance, to display the usage and options of “index” command, run “samtools index”.
In the following, we will discuss the most common uses of Samtools utilities.

2.4.1.1 SAM/BAM Format Conversion

The “samtools view” command can be used to convert SAM file into BAM file and vice
versa. If you did not delete the SAM files, which were produced above and saved in “sam”
directory, you can use any one of them for this practice.
From inside “sam” directory, for instance, you can run the following command to con-
vert “SRR769545_mem.sam” to a BAM file “SRR769545_mem.bam”:
80   ◾    Bioinformatics

samtools view \
-@ 4 \
-uS -o SRR769545_mem.bam SRR769545_mem.sam

The “-@” option specifies the number of threads, the “-u” option is to produce uncom-
pressed BAM output, “-S” is to ignore auto-detection of the input format, and “-o” specifies
the output file.
If you have already run the above command and the old SAM file is still there, you can
delete it with “rm SRR769545_mem.sam” command and then run the following to convert
the BAM file back to a SAM file:

samtools view \
-@ 4 -h \
-o SRR769545_mem.sam \
SRR769545_mem.bam

The “-h” option is to include header in SAM output.

2.4.1.2 Sorting Alignments

Sorted alignments are required for some applications such as variant calling and RNA-Seq.
Samtools can sort the alignments in a BAM file by coordinate order, by read name, or by
a TAG. By default, “samtools sort” will sort the alignment by the coordinate order, unless
“-n” or “-t TAG” option is used. If “-n” option is used, the command will sort the align-
ments by read name. If the “-t TAG” option is used, then the alignments will be sorted by
tag. Refer to SAM and BAM file formats discussed above to learn about the standard TAGs.
In the following, the BAM file will be sorted by coordinate order:

samtools sort \
-@ 4 \
-T mem.tmp.sort \
-o SRR769545_mem_sorted.bam \
SRR769545_mem.bam

This will create a new BAM file with sorted alignments. You can delete the unsorted BAM
file if you need to save some storage space.

2.4.1.3 Indexing BAM File

Before using a BAM file in any of the downstream analysis, it can be indexed to allow fast
random access to the alignments in the file. Thus, the alignment information will be pro-
cessed faster. We can use the “samtools index” command to index the above sorted BAM
file as follows:

samtools index SRR769545_mem_sorted.bam

This will generate the BAM index file “SRR769545_mem_sorted.bam.
bai”.
 Mapping of Sequence Reads to the Reference Genomes   ◾    81

2.4.1.4 Extracting Alignments of a Chromosome

Sometimes, we may need to work with alignments of a specific chromosome or a specific
region of the genome. With the following command, you can split the alignments of a
chromosome 11 in a separate file using “samtools view”:

samtools view SRR769545_mem_sorted.bam NC_000011.10 > chr11_human.

sam

You can use any of the reference sequence names in the RNAME field of the SAM/BAM
file, so you may need to display the content of the file to check how the reference sequences/
chromosomes are named.

2.4.1.5 Filtering and Counting Alignment in SAM/BAM Files

To filter alignments in a SAM/BAM file, we can use “samtools view” with “grep” which is a
Linux command for searching plain-text datasets for lines that match a regular expression.
For instance, to search for the alignments with chimeric reads, which are tagged as “SA:”
an optional SAM/BAM field, we can use the following:

samtools view SRR769545_mem_sorted.bam | grep ‘SA:’ | less -S

The chimeric read is the one that aligns to two distinct portions of the genome with little
or no overlap.
To count the number of chimeric reads, we can use “wc -l” command.

samtools view SRR769545_mem_sorted.bam | grep ‘SA:’ | wc -l

We can also use the option “-c” with “samtools view” to count the number of reads in a
BAM file:

samtools view -c SRR769545_mem_sorted.bam

We can use values in FLAG field of the SAM/BAM file to count the number of reads defined
by a specific FLAG value. For instance, since the unmapped reads will be flagged as “0x4”
in BAM files, we can count all mapped reads by excluding the unmapped from counting
using the “-F” option.

samtools view -c -F 0x4 SRR769545_mem_sorted.bam

To count unmapped reads, use the “-f” option instead of the “-F” option as:

samtools view -c -f 0x4 SRR769545_mem_sorted.bam

We can also use the “samtools view” command together with some Unix/Linux com-
mands and pipe symbol “|” to perform more complex count. For instance, we can count
82   ◾    Bioinformatics

the number of insertions (I) or deletion (D) from the CIGAR strings in a BAM file. Since
the CIGAR field is the sixth column, first, we will use “samtools view -F 0x4 SRR769545_
mem_sorted.bam” to extract the mapped records. Then, we can transfer that output to “cut
-f 6” using the pipe symbol “|” to separate the sixth column. The output is then transferred
to “grep -P” to select only the strings that have either the character “D” or “I” using the
class pattern “[ID]” to match any of the two characters. Then, the output is transferred to
the “tr” command to delete any characters other than “I” and “D”. Finally, the output is
transferred to the “wc -c” command to count the remaining characters:

samtools view \
-F 0x4 SRR769545_mem_sorted.bam \
| cut -f 6 \
| grep -P ‘[ID]’ \
| tr -cd ‘[ID]’ \
| wc -c

To count insertions and deletions separately, use the following, respectively:

samtools view \
-F 0x4 SRR769545_mem_sorted.bam \
| cut -f 6 \
| grep -P ‘I’ \
| tr -cd ‘I’ \
| wc -c
samtools view \
-F 0x4 SRR769545_mem_sorted.bam \
| cut -f 6 \
| grep -P ‘D’ \
| tr -cd ‘D’ \
| wc -c

Refer to Table 2.3 for the different FLAG values and descriptions.

2.4.1.6 Removing Duplicate Reads

Duplicate reads may be produced from the library construction, PCR amplification (PCR
duplicates), or a fault in the sequencing optical sensor (optical duplicates). A large number
of duplicate reads originating from a single fragment may create a bias in some applica-
tions, such as RNA-Seq, in which the count of reads has a biological interpretation. The
“samtools rmdup” command can be used to remove potential duplicate reads from BAM/
SAM files. If multiple reads have identical coordinates, only the read (read pair if paired
end) with the highest mapping quality will be retained. By default, this command works
for paired-end reads. The option “-s” is used if the reads are single end.

samtools rmdup \
SRR769545_mem_sorted.bam \
 Mapping of Sequence Reads to the Reference Genomes   ◾    83

SRR769545_mem_rmdup.bam 2> SRR769545_mem_rmdup.log

The above command removes the duplicate reads from the BAM file (paired end). If we
need paired-end reads to be treated as single end, use “-S” option.

2.4.1.7 Descriptive Statistics

Some Samtools utilities including “flagstat”, “coverage”, and “depth” can provide simple
statistics on a BAM file.

samtools flagstat SRR769545_mem_sorted.bam

samtools coverage SRR769545_mem_sorted.bam > coverage.txt
samtools depth SRR769545_mem_sorted.bam > depth.txt

For other Samtools commands, you can check the Samtools documentation which is avail-
able at “http://www.htslib.org/doc/samtools.html”.

2.5 REFERENCE-GUIDED GENOME ASSEMBLY

The reference-guided genome assembly is the use of a reference genome of an organism
as a guide to assemble a new genome. This kind of assembly is used when a genome is
re-sequenced to obtain better quality genome assembly or for variant discovery and hap-
lotype construction. The reference-guided genome assembly is widely used to sequence
the genome of individuals of the same species as of the reference genome to detect the
genotypes that may associate with certain phenotype or diseases such as cancers, viral and
bacterial variants or strains. It can also be used to assemble the genomes of closely related
species who do not have available reference genomes. Sequences of the whole genome of an
individual are used in the assembly. The workflow is the same as shown in Figure 2.13 until
the point of creating SAM/BAM file. The additional step is that the aligned reads are piled
up to create consensus sequences from the overlapped contiguous aligned reads. These con-
sensus sequences are called contigs. From these contigs, only the different bases (variants)
are used to edit the sequence of the reference genome to create a new genome sequence.
For the following practice, you can use any of the SAM files produced by BWA or
Bowtie2 above or you can run the following commands to download the FASTQ file from
the NCBI SRA database and decompress them, to download the human reference genome
from UCSC database and index it, and then to perform read mapping with Bowtie2 to
produce a SAM file:

mkdir ref_guided_ass
cd ref_guided_ass
mkdir data
fasterq-dump --verbose SRR769545
gzip SRR769545_1.fastq
gzip SRR769545_2.fastq
cd ..
mkdir ref
84   ◾    Bioinformatics

cd ref
wget https://hgdownload.soe.ucsc.edu/goldenPath/hg38/bigZips/hg38.
fa.gz
gunzip -d hg38.fa.gz
samtools faidx hg38.fa
bowtie2-build --thread 4 hg38.fa hg
cd ..
mkdir sam
bowtie2 -x \
ref/hg \
-1 data/SRR769545_1.fastq.gz \
-2 data/SRR769545_2.fastq.gz \
-S sam/SRR769545.sam

This time we have chosen to download the UCSC human reference genomes because the
chromosome names are given instead of accession numbers.
Once the SAM file has been created, we can convert it into a BAM file, sort it, and index
it using Samtools.

samtools view -uS -o sam/SRR769545.bam sam/SRR769545.sam

samtools sort -@ 4 -T bowtie2.tmp.sort \
-o sam/SRR769545_sorted.bam sam/SRR769545.bam
samtools index sam/SRR769545_sorted.bam

Then, we can use bcftools to detect the variants like variant substitutions and write that
information into a binary variant call format (BCF), which is the binary form of vari-
ant call format (VCF). This file includes the difference in genotype between the reference
genome and the genome of the individual sequenced. The variant discovery and VCF file
format will be discussed in detail in Chapter 4. However, they have to be used here to show
how the reference-guided genome assembly is performed. The bcftools can be installed in
Linux as follows:

sudo apt install bcftools

Now, we can use the “bcftools mpileup” command to collapse the pileup of the reads
aligned to the reference genomes in the BAM file and then to write only the variant infor-
mation into a VCF file. In the following, the variant information is written into a BCF file
and then that binary file is converted into a VCF file:

bcftools mpileup -f ref/hg38.fa \

sam/SRR769545_sorted.bam \
| bcftools call -mv -Ob \
-o sam/SRR769545.bcf
bcftools convert -O v \
-o SRR769545.vcf \
SRR769545.bcf
 Mapping of Sequence Reads to the Reference Genomes   ◾    85

The VCF can be compressed using the Linux compression program “bgzip” and then the
compressed file is indexed using tabix program, which is a tool for indexing large bioinfor-
matics text files. The tabix program can be installed on Linux using:

sudo apt install tabix

The following commands compress the VCF file and index it:

bgzip -c SRR769545.vcf > SRR769545.vcf.gz

tabix -p vcf SRR769545.vcf.gz

The final step in the reference-guided genome assembly is to create a consensus sequence
by transferring the sequence reference genome, which was used to create the original BAM
file, to the “bcftools consensus” command that utilizes the indexed variant call file to cre-
ate a new genome sequence for the individual studied.

cat ../ref/hg38.fa \
| bcftools consensus SRR769545.vcf.gz \
> SRR769545_genome.fasta

The “SRR769545_genome.fasta” is the FASTA genome sequence that incorporates variants

genotyped for the individual from whom the whole genome was sequenced. If a region of
the reference genome is uncovered by the reads, this assembly method may skip some vari-
ants; therefore, high sequence coverage is required. Instead of incorporating the uncovered
regions of the reference genome in the new sequence, the bases of those regions can be
masked by Ns. There are several more advanced programs for reference-guided genome
such as RATT [19].
There is a different approach to sequence a genome of an organism from scratch without
using a reference genome. This approach is called de novo assembly which is discussed in
Chapter 3.

2.6 SUMMARY
Except for the sequencing applications that use de novo genome assembly, read mapping
to a reference genome is the most fundamental step in the workflow of the sequencing
data analysis. In the NGS or TGS, the DNA molecules are fragmented into pieces in the
library preparation step and then the DNA libraries are sequenced to produce the raw data
in the form of millions of reads of specific lengths. The lengths of the reads produced by
a high-throughput sequencing instrument vary based on the technology used into short
reads (50–400 bp) or long reads (>400 bp). The quality control step to assess and prepro-
cess the raw reads is essential to reduce the errors in the base calling and the biases that
may arise due to the presence of technical reads. Indeed, read mapping is the most com-
putationally expensive step in the sequencing data analysis workflow. That is because the
alignment program attempts to determine the points of origin for millions or billions of
reads in a reference genome. The alignment requires even more efforts for RNA-Seq read
86   ◾    Bioinformatics

mapping because the aligner should also be able to detect the splice junctions. The process
of alignment is usually complicated by the possible existence of mismatches, which may
be due to base call errors or due to genetic variations in the individual genome. In general,
aligners are required to use a strategy that enables them to perform both an exact search
and an inexact search to allow locating positions of reads with mismatches. Almost all
read aligners perform alignment in two major steps: indexing of the sequence of the refer-
ence genome and finding the most likely locations of the reads in the reference genome.
The FASTA sequence of the reference of an organism can be downloaded from genome
databases such as NCBI Genome and UCSC database. The FASTA genome sequence is
indexed first by the “samtools faidx” command to allow fast processing by the aligners.
The commonly used data structures for genome indexing include BWT, FM-index, suf-
fix arrays, and hash table for their memory efficiency and capability to store a genome
sequence. There are a variety of aligners that use different indexing and lookup algorithms.
We discussed only BWA, Bowtie2, and STAR. However, those are only examples. Before
using an aligner, you may need to know its memory efficiency and whether it is capable to
use short reads, long reads, or both. If you have RNA-Seq reads, you may also need to know
whether that aligner is capable to detect splice junctions or not. Both BWA and Bowtie2 are
general purpose aligners that can be used for all kinds of reads and they can operate well
on a desktop computer with 32GB of RAM or more. STAR is better for RNA-Seq read and
it can also run on a desktop computer but it requires much more memory for both index-
ing and mapping.
Almost all aligners produce SAM/BAM files, which store read mapping information.
A SAM/BAM file consists of a header section and an alignment section. The alignment
section includes nine mandatory columns; each row of the columns contains the mapping
information of a read. The alignment information includes read name, FLAG, reference
sequence name (e.g., chromosome name or accession), position in the reference sequence
(coordinate), mapping quality, CIGAR string, reference name of the mate, position of the
mate, read length, segment of the read sequence, and Phred base quality. FLAG field stores
standard codes that describe the alignment (e.g., unmapped reads, duplicate reads, and
chimeric alignments). The CIGAR string describes the operations that took place on the
reads such as matches, mismatches, insertions, and deletions.
SAM/BAM files can be manipulated by some programs like Samtools and PICARD.
The file manipulation includes format conversion, indexing, sorting, displaying, statistics,
viewing, and filtering.
The SAM/BAM files are used in the downstream data analysis such as reference-guided
genome assembly, variant discovery, gene expression (RNA-Seq data analysis), epigenetics
(ChIP-Seq data analysis), and metagenomics as we will discuss in coming chapters.

REFERENCES
1. Needleman SB, Wunsch CD: A general method applicable to the search for similarities in the
amino acid sequence of two proteins. J Mol Biol 1970, 48(3):443–453.
2. Chacón A, Moure JC, Espinosa A, Hernández P: n-step FM-index for faster pattern matching.
Proc Comput Sci 2013, 18:70–79.
 Mapping of Sequence Reads to the Reference Genomes   ◾    87

3. Smith TF, Waterman MS: Identification of common molecular subsequences. J Mol Biol 1981,
147(1):195–197.
4. Altschul SF, Gish W, Miller W, Myers EW, Lipman DJ: Basic local alignment search tool. J Mol
Biol 1990, 215(3):403–410.
5. Fredkin E: Trie memory. Commun ACM 1960, 3:490–499.
6. Ukkonen E: On-line construction of suffix trees. Algorithmica 1995, 14(3):249–260.
7. Shrestha AMS, Frith MC, Horton P: A bioinformatician’s guide to the forefront of suffix array
construction algorithms. Brief Bioinfor 2014, 15(2):138–154.
8. Dobin A, Davis CA, Schlesinger F, Drenkow J, Zaleski C, Jha S, Batut P, Chaisson M, Gingeras
TR: STAR: ultrafast universal RNA-seq aligner. Bioinformatics 2013, 29(1):15–21.
9. Li H, Durbin R: Fast and accurate short read alignment with Burrows–Wheeler transform.
Bioinformatics 2009, 25(14):1754–1760.
10. Langmead B, Trapnell C, Pop M, Salzberg SL: Ultrafast and memory-efficient alignment of
short DNA sequences to the human genome. Genome Biol 2009, 10(3):R25.
11. Fernandez E, Najjar W, Lonardi S: String Matching in Hardware Using the FM-Index. In:
2011 IEEE 19th Annual International Symposium on Field-Programmable Custom Computing
Machines: 1–3 May 2011 2011. 218–225.
12. Sankoff D, Kruskal J, Nerbonne J: Time Warps, String Edits, and Macromolecules: The Theory
and Practice of Sequence Comparison: Cambridge University Press; 2000.
13. Levin LA: Problems of Information Transmission. In: 1973.
14. Mu JC, Jiang H, Kiani A, Mohiyuddin M, Bani Asadi N, Wong WH: Fast and accurate read
alignment for resequencing. Bioinformatics 2012, 28(18):2366–2373.
15. Li R, Li Y, Kristiansen K, Wang J: SOAP: short oligonucleotide alignment program.
Bioinformatics 2008, 24(5):713–714.
16. Li H, Handsaker B, Wysoker A, Fennell T, Ruan J, Homer N, Marth G, Abecasis G, Durbin R:
The sequence alignment/map format and SAMtools. Bioinformatics 2009, 25(16):2078–2079.
17. Langmead B, Salzberg SL: Fast gapped-read alignment with Bowtie 2. Nature Methods 2012,
9(4):357–359.
18. Vasimuddin M, Misra S, Li H, Aluru S: Efficient Architecture-Aware Acceleration of BWA-
MEM for Multicore Systems. In: 2019 IEEE International Parallel and Distributed Processing
Symposium (IPDPS): 20–24 May 2019 2019. 314–324.
19. Otto TD, Dillon GP, Degrave WS, Berriman M: RATT: Rapid Annotation Transfer Tool.
Nucleic Acids Res 2011, 39(9):e57.
 Chapter 3

De Novo Genome Assembly

3.1 INTRODUCTION TO DE NOVO GENOME ASSEMBLY

In the previous chapter, we discussed reads mapping to the reference genomes of organ-
isms which have available reference genome sequences and we also discussed the refer-
ence-based genome assembly. So, what if there is no reference genome available for an
organism or we need to sequence a genome of an unknown organism. In this case, aligning
reads to a reference genome is not possible and we shall assume no prior knowledge about
the genome of that organism or its length, or composition. Thus, de novo genome assembly
comes into play. It can also be used for a species with a solved reference genome for vari-
ant discovery if we need to avoid any bias created by a prior knowledge. De novo genome
assembly is a strategy to assemble a novel genome from scratch without the aid of a refer-
ence genome sequence. Because of the improvements in cost and quality of DNA sequenc-
ing, de novo genome assembly now is widely used specially in metagenomics for bacterial
and viral genome assembly from environmental and clinical samples.
The de novo genome assembly aims to join reads into a contiguous sequence called
a contig. Multiple contigs are joined together to form a scaffold and multiple scaffolds
can also be linked to form a chromosome. The genome assembly is made of the consen-
sus sequences. Both single-end and paired-end or mate-pair reads can be used in the de
novo assembly, but paired reads are preferred because they provide high-quality align-
ments across DNA regions containing repetitive sequences and produce long contigs by
filling gaps in the consensus sequence. Assembling the entire genome is usually challeng-
ing because of the presence of numerous stretched tandem repeats in the genome. These
repeats create gaps in the assembly. Gap problem can be overcome by deep sequencing,
which is sequencing a genome multiple times to provide sufficient coverage and sequence
depth, which increase the chance for read overlaps. The sequencing coverage is defined as
the average number of reads that align to or cover known reference bases of the genome
and it is estimated as follows [1]:

read length × number of reads

coverage = (3.1)
haploid genome length ( bp )

DOI: 10.1201/9781003355205-3 89
90   ◾    Bioinformatics

FIGURE 3.1 Sequencing coverage and depth.

The coverage describes how often, on average, a reference sequence is covered by bases
from the reads. The sequencing depth is defined as the total number of reads that align to
or cover a specific locus (Figure 3.1).
The de novo assembly strategy has been greatly enhanced by the single-molecule real-
time (SMRT) sequencing (Pacific Biosciences (PacBio)) and nanopore sequencing (Oxford
Nanopore Technologies (ONT)), which produce long reads. In general, the de novo genome
assembly depends on the read length, sequencing depth, and the assembling algorithm.
Based on the sequencing technology used, reads can be classified into short reads, which
are produced by the most sequencing technologies including Illumina, and long reads pro-
duced by PacBio and ONT. In general, either sequence alignment or graphs approach is
used for the de novo assembly. The algorithms used by the de novo genome assemblers can
be classified into (i) greedy approach, (ii) overlap-layout-consensus with Hamiltonian path,
or (iii) de Bruijn graph with Eulerian path. The last two approaches share the idea that a
genome assembly can be represented as a path in graphs.

3.1.1 Greedy Algorithm

The greedy algorithm was the first one used for de novo genome assembly. Like multi-
ple sequence alignment, it depends on the similarity between reads to create a pileup of
aligned sequences, which are collapsed to create contigs from the consensus sequences. The
greedy algorithm uses pairwise alignment to compare all reads to identify the most simi-
lar reads with sufficient overlaps and then it joins them together. The reads with the sig-
nificant overlaps are merged. The algorithm continues by conducting pairwise alignments
on the merged reads to identify the sequences with significant overlaps and merges them.
This process will continue repeatedly until there is no more merging. When there is low
sequencing depth, gaps may be present between the assembled contigs; deep sequencing
and the use of paired-end reads can reduce gaps and allow longer contigs. Greedy assem-
blers include TIGR assembler [2] and PHRAP [3]. For instance, assume that we have the
reads CATTC, ATTCG, GTCAT, TTCGC, and GGCAT. Figure 3.2 shows how the greedy
algorithm works to assemble the consensus sequence GTCATTCGCAT from those reads.

3.1.2 Overlap-Consensus Graphs

The overlap graph method was first used for assembling genomes from sequences produced
by Sanger sequencing method and then it was extended by some assemblers for NGS reads.
 De Novo Genome Assembly   ◾    91

FIGURE 3.2 Greedy assembly.

The overlap-consensus algorithm performs pairwise alignment and then it represents the
reads and the overlaps between the reads with graphs, where contiguous reads are the
vertexes (nodes) and their overlaps (common prefix or suffix) are the edges. Each node (ri )
corresponds to a read and any two reads are connected by an edge e ( r1 , r2 ) where the suffix
of the first read matches the prefix of the second read. The algorithm then attempts to find
the Hamiltonian path of the graph that includes the vertexes (reads) exactly once. Contigs
are then created from the consensus sequences of the overlapped suffixes and prefixes. The
downside of this method is that it requires a large sequencing coverage and the complexity
of the Hamiltonian path increases with the increase of reads. An example for an assembler
using this approach is Celera Assembler [4]. To show you how the approach of the overlap-
consensus graphs works, assume that we have the reads:

CAGAACCTAGC
ATAGCAGAACG
GCTAAGCAGTG
AGTGTCACGAC
TAGCAACTATG
AACGTAGCAGA

Figure 3.3 and 3.4 show the overlap graphs in which the reads are the nodes and the over-
lapped prefixes and suffixes are the edges. The Hamiltonian path includes each node
exactly once to form the consensus sequence.

3.1.3 De Bruijn Graphs

De Bruijn graph [5, 6] approach for de novo assembly was first introduced in 2001 for assem-
bling short reads produced by the NGS technologies and then it was refined in 2008 [7]. In
de Bruijn graphs, each read is broken into overlapping substrings of length k called over-
lapping k-mers. The k-mers then are represented by graphs in which each k-mer is a vertex
(node). Any two nodes of any two k-mers that share a common prefix and suffix of length
92   ◾    Bioinformatics

FIGURE 3.3 Overlap graphs and Hamiltonian path.

FIGURE 3.4 The overlaps and consensus sequence.

k-1 are connected by an edge. The contiguous reads are merged by finding Eulerian path,
which includes every edge exactly once. Most modern assemblers like ABySS and SPAdes
use de Bruijn graphs method because it is less complex than the overlap graphs. The qual-
ity of the de novo assembly is greatly affected by the value of the parameter k (number of
bases). For instance, assume that we have the read TCTACTAAGCTAACATGCAATGCA.
When we implement de Bruijn algorithm in any programming languages, we can obtain
the graphs showing the nodes (k-mers) and the edges as shown in Figure 3.5.
For both overlap and de Bruijn graphs, once we have constructed the graphs, the next
step is to find the relevant paths that are likely to represent the assembly. There should be a
single path for the assembly if there is no error or sequence repeats, which is rare in most
cases.
 De Novo Genome Assembly   ◾    93

FIGURE 3.5 De Bruijn graphs.

3.2 EXAMPLES OF DE NOVO ASSEMBLERS

The above is a brief definition for the three kinds of the algorithms for the de novo genome
assembly. For the algorithms themselves, you may need to refer to an algorithm book.
In the following, we will discuss some example assemblers to show you how the de novo
genome assembly is performed. There are many papers that reported the comparative per-
formance of these assemblers and others. Readers can refer to those papers to see the dif-
ferences found by researchers.

3.2.1 ABySS
ABySS (Assembly By Short Sequences) [8] is a parallel de novo genome assembler devel-
oped to assemble very large data of short reads produced by NGS technologies. It per-
forms assembly in two stages. First, it generates all possible k-mers from the reads, removes
potential errors, and builds contigs using de Bruijn graphs. Second, it uses mate-pair infor-
mation to extend contigs benefiting from contig overlaps and merges the unambiguously
connected graph nodes. ABySS can be used in two modes: bloom filter mode, which uses
hashing, and MPI mode, which uses message passing interface (MPI) to parallelize the de
novo assembly. It is recommended to use the bloom filter mode over the legacy MPI because
it reduces the memory usage to 10 folds. ABySS can be installed following the instructions
available at “https://github.com/bcgsc/abyss”. On Ubuntu, we can install it using “sudo
apt-get install abyss”. Once ABySS has been installed, the “abyss-pe” command can be
94   ◾    Bioinformatics

used for the de novo assembly. For the purpose of the practice, we will use paired-end short
reads produced by Illumina MiSeq for whole genome sequencing of Escherichia coli (str.
K-12). The files (forward and reverse FASTQ files) are available at the NCBI SRA database.
First, using the Linux terminal, create a directory for the exercise and change to it.

mkdir denovo; cd denovo

Inside that directory, create a subdirectory for the FASTQ files.

mkdir fastq; cd fastq

Then, you can download the raw data files into “fastq” directory using SRA toolkits as
follows:

fasterq-dump --threads 4 --verbose ERR1007381

To save some storage space, you can compress the two FASTQ files with GZIP utility as
follows:

gzip ERR1007381_1.fastq
gzip ERR1007381_2.fastq

The compression will reduce the storage of the two files from 11 G to 3 G.
We will use “abyss-pe” command to perform the de novo genome assembly. Change to
the main exercise directory just a single step out of “fastq” directory by using “cd ..”. Then,
run the following command to construct contigs:

abyss-pe \
name=ecoli \
j=4 \
k=25 \
c=360 \
e=2 \
s=200 \
v=-v \
in=’fastq/ERR1007381_1.fastq.gz fastq/ERR1007381_2.fastq.gz’ \
contigs \
2>&1 | tee abyss.log

The following will construct scaffolds from the contigs:

abyss-pe \
name=ecoli \
j=4 \
k=25 \
c=360 \
 De Novo Genome Assembly   ◾    95

e=2 \
s=200 \
v=-v \
in=’fastq/ERR1007381_1.fastq.gz fastq/ERR1007381_2.fastq.gz’ \
scaffolds stats

The “name=ecoli” specifies the prefix string as “ecoli”, “j=4” specifies the number of
­processors to be used, “k=25” specifies the length of the k-mer (substring), “c=255” ­specifies

the minimum mean k-mer coverage of a contig (a high-confidence contig), “e=2” specifies
the minimum erosion k-mer coverage, “s=200” specifies the minimum contig size (bp)
required for building scaffolds, “v=-v” enables verbose, and “in=” specifies the FASTQ file
names as an inputs.
When the execution of the above commands on the Linux terminal is complete, the
assembly statistics will be displayed. Assembling a genome with “abyss-pe” command is
performed in three stages: assembling contigs without paired-end information, aligning
the paired-end reads to an initial assembly, and finally merging contigs joined by paired-
end information. Multiple files will be generated for those three stages. The descriptions of
these files are as follows:

• A file with “*.dot” extension is a graph file in Graphviz format in which graphs (nodes
and edges) are defined in DOT language script. We can visualize this file by drawing
the graphs with the Graphviz program, which can be installed on Linux using “sudo
apt install graphviz”. We can then draw the graphs and save it in the PNG format using:

dot -Tpng ecoli-5.dot -o ecoli-5.png

• A file with “*.hist” extension is a histogram file made of two tab-separated columns:
the fragment size and count. It shows the distribution of the fragment sizes.
• A file with “*.path” extension is an ABySS graph path file, which describes how
sequences should be joined to form new sequences.
• A file with “*.sam.gz” extension is a SAM file that is used by ABySS to describe align-
ments of reads to assemble sequences at different stages of the assembly.
• A file with “*.fa” extension is a FASTA file format for contig and scaffold sequences.
The definition line of the FASTA file consists of three parts: <SEQ_ID> <SEQ_LEN>
<KMERS>, where SEQ_ID is a unique identifier for the sequence assigned by ABySS,
SEQ_LEN is the length of the sequence in bases, and KMERS is the number of
KMERS that mapped to the sequence in the assembly.
• A file with “*.fai” extension is an indexing file for the index of the corresponding
FASTA sequences. It includes five tab-separated columns (name of the sequence,
length of the sequence, offset of the first base in the file, number of bases in each line,
and number of bytes in each line).
96   ◾    Bioinformatics

• A file with “*stats.csv” or “*stats.tab” extension is the statistics of contig/scaffold

contiguity in CSV format. The assembly statistics generated by ABySS are shown in
Figure 3.6 and described in Table 3.1.

The contig/scaffold N50 metric is the most widely used metric for describing the quality
of a genome assembly. A contig/scaffold N50 is calculated by first ordering every contig/
scaffold by length from the longest to the shortest. Next, the lengths of contigs are summed
starting from the longest contig until the sum equals one-half of the total length of all con-
tigs in the assembly. The contig/scaffold N50 of an assembly is the length (bp) of the short-
est contig/scaffold from the sequences that form 50% of the assembly. To compare between
assemblies, the longer the N50 and the smaller the L50, the better the assembly.
In Figure 3.6, the scaffolds file (ecoli-scaffolds.fa) contains 836 sequences, of which
107 sequences are more than 500 bp. The shortest sequence has 584 bp and the longest is
267,586 bp. The N50 is the sequence of length 112,320 bp and L50 (the number of scaffolds
that accounts for more than 50% of the genome assembly) is 15.
Figure 3.7 shows a diagram explaining the major metrics of the genome assembly
(N25=55, N50=70, N75=75, L25=4, L50=6, and L75=7) and how they can be computed. In
the figure, there are eight contigs ranked from the smallest to the largest. The total number
of bases is 445 Mb (100%) and the half number is 222.5 Mb (50%).
You can display both contigs and scaffolds file on a Linux terminal using the “less”
Linux command as:

FIGURE 3.6 Assembly statistics.

TABLE 3.1 Assembly Statistics

Column Description
N The total number of sequences in the FASTA file
n:500 The number of sequences whose lengths are not less than 500 bp
L50 The number of scaffolds that account for more than 50% of the assembly
LG50 The number of scaffolds that account for more than 50% of the genome assembly
NG50 The sequence length of the shortest contig at 50% of the total genome length
Min The size of the smallest sequence
N75 The sequence length of the shortest contig at 75% of the total assembly length
N50 The sequence length of the shortest contig at 50% of the total assembly length
N25 The sequence length of the shortest contig at 25% of the total genome length
E-size The sum of the square of the sequence sizes divided by the assembly size
Max The size of the largest sequence
Sum The sum of the sequence sizes
Name The file name of the assembly
 De Novo Genome Assembly   ◾    97

FIGURE 3.7 Genome assembly metrics.

less ecoli-contigs.fa
less ecoli-scaffolds.fa

Use “awk” command to print the length of the longest scaffold in the scaffold file.

awk ‘{print length}’ ecoli-contigs.fa | sort -n | tail -n1

3.2.2 SPAdes
SPAdes [9] is a de novo genome assembler developed primarily for assembling small
genomes of bacteria. Later, modules were added for assembling small genomes of other
organisms including fungi and viruses. It is not recommended for assembling large mam-
malian genomes. The current SPAdes version works with both Illumina and Ion Torrent
reads, and it can be used for genome hybrid assembly for PacBio, Oxford Nanopore, and
Sanger reads. This assembler can process several paired-end and mate-paired files in the
same time. The program also provides separate modules for metagenomic data, plasmid
assembly from the whole genome sequencing data, plasmid from metagenomic data, tran-
scriptome assembly from RNA-Seq data, biosynthetic gene cluster assembly with paired-
end reads, viral genome assembly from RNA viral data, SARS-CoV-2 assembly, and TruSeq
barcode assembly. The assembling process of SPAdes includes four stages. First, de Bruijn
graphs are built from overlapping k-mers generated from the reads. Second, the k-mers are
adjusted to obtain accurate distance estimates between k-mers using both distance histo-
grams and paths in the assembly graphs. The program then constructs paired de Bruijn
graphs, which is a generalization of the de Bruijn graph that incorporates mate-pair infor-
mation into the graph structure [10]. Finally, contigs are constructed from the graphs.
SPAdes program is made up of modules in Python. The installation instructions are
available at “https://cab.spbu.ru/files/release3.15.4/manual.html”. To install the program
on Linux, use the following steps:
Using the Linux terminal, first download and decompress the source program in a local
directory.

wget https://cab.spbu.ru/files/release3.15.4/SPAdes-3.15.4-Linux.
tar.gz
tar -xzf SPAdes-3.15.4-Linux.tar.gz

Notice that the program name or path may change in the future.
98   ◾    Bioinformatics

After decompression, the program files will be in “SPAdes-3.15.4-Linux/bin” d ­ irectory.

You will need to add the program path to “.bashrc” file to be able to use the SPAdes ­program
files from any directory.

export PATH=”/home/your_path/SPAdes-3.15.4-Linux/bin:$PATH”

Replace “your_path” with the correct path to the directory on your computer. After adding
the line, restart the terminal or use “source ~/.bashrc” for the change to take effect. You can
test the installation by running the following command on the Linux terminal command
line:

spades.py --test

The above command will assemble toy read data that comes with the program. The output
will be saved in “spades_test”, which are the typical SPAdes output files.
Remember that we have downloaded two FASTQ files for E. coli and we used them with
ABySS above. The two files were saved in “denovo/fastq”; run the following command on
the Linux command line while you are in the working directory “denovo”:

python spades.py \
--pe1-1 fastq/ERR1007381_1.fastq.gz \
--pe1-2 fastq/ERR1007381_2.fastq.gz \
--isolate \
-o spades_ecoli_ass

The above code assembles contigs and scaffolds from the paired-end reads in the input
FASTQ files and saves the output files in a new directory “spades_ecoli_ass”. The output
files include intermediate file used for the construction of contigs and scaffolds, log files,
contigs’ file, and scaffolds’ file in FASTA format. The latter represents the assembly.
SPAdes can perform hybrid de novo assembly using any of PacBio continuous long reads
(CLR) or Oxford Nanopore reads as input with Illumina or Ion Torrent reads. You will
need to use “--pacbio” option for the PacBio FASTQ file and “--nanopore” option for the
Oxford Nanopore FASTQ file. In the following example, first, we download four PacBio
SMRT FASTQ files “SRR801646”, “SRR801649”, “SRR801652”, and “SRR801638” for E. coli
K12. Second, we will use these four PacBio files with the Illumina short read files to assem-
ble E. coli genome with SPAdes program.

python spades.py \
--pe1-1 fastq/ERR1007381_1.fastq.gz \
--pe1-2 fastq/ERR1007381_2.fastq.gz \
--pacbio pacbio/SRR801646.fastq.gz \
--pacbio pacbio/SRR801649.fastq.gz \
--pacbio pacbio/SRR801652.fastq.gz \
--pacbio pacbio/SRR801638.fastq.gz \
 De Novo Genome Assembly   ◾    99

--careful \
--isolate \
-o hyb_spades_ecoli_ass

We used “--gzip” with “fastq-dump” command to compress the FASTQ files.

SPADes program has different modes for different applications. With “sapdes.
py” ­command, we can use “--meta” option for metagenomic mode, “--bio” option for
­biosyntheticSPAdes mode, “--corona” option for coronaSPAdes mode, “--rna” option for
transcriptomic mode (RNA-Seq reads), “--plasmid” option for plasmid detection mode,
“--metaviral” option for virus detection mode, “--metaplasmid” option for metagenomic
plasmid detection mode, and “--rnaviral” option for virus assembly mode (from RNA-Seq
reads).
In the following example, we will download Illumine paired-end FASTQ files
“ERR8314890” for the whole genome sequencing of SARS-CoV-2 virus surveillance from
the NCBI SRA database. Then, we will use coronaSPAdes module to assemble SARS-CoV-2
genome. For this exercise, you can create a directory “sarscov2” and change into it as:

makdir sarscov2; cd sarscov2

Then, you can download the FASTQ files from the NCBI SRA database using SRA toolkits
program “fasterq-dump”.

fasterq-dump --verbose ERR8314890

Then, you can run SPAdes program to assemble the SAR-CoV-2 genome using “--corona”
option.

python spades.py \
--pe1-1 ERR8314890_1.fastq \
--pe1-2 ERR8314890_2.fastq \
--corona \
-o sarscov2_genome

The output files including FASTA files of contigs and scaffolds will be saved in the specified
output directory “sarscov2_genome”.
The other SPAdes modes work the same.

3.3 GENOME ASSEMBLY QUALITY ASSESSMENT

After assembling a genome using any of the de novo genome assemblers, the next step is
to assess the quality of assembly to have an idea about how the assembly is good. Genome
assessment metrics provide important information on how the assembly is reliable or not.
There are two approaches for the quality assessment of an assembly. The first one is a sta-
tistical approach that depends on statistical metrics for measuring the quality of a genome
100   ◾    Bioinformatics

assembly. An example of programs which use the statistical approach is QUAST. The sec-
ond approach depends on evolutionary relatedness to similar organisms to estimate the
number of genes in the genomes and gene completeness. An example of the programs that
use this approach is BUSCO.

3.3.1 Statistical Assessment for Genome Assembly

The statistical assessment of a genome assembly can be performed in two ways: reference-
guided approach and non-reference approach. One reason for using the de novo genome
assembly is the unavailability of a reference genome for the species studied. However, we
know that the de novo assembly can also be used in some applications to assemble a genome
of an organism that has a reference genome. When an organism has no reference genome,
the de novo genome assembly can be assessed without a reference genome. Some assem-
blers like ABySS can generate some useful statistics for assessing the assembled genome.
Table 3.1 includes the statistical metrics for a genome assembly. Not all assemblers have
modules generating that assessment metrics. We can use a program like QUAST (QUality
ASsessment Tool) [11] for this purpose. The current version of QUAST can be downloaded,
decompressed, and installed as follows:

wget https://downloads.sourceforge.net/project/quast/quast-
5.0.2.tar.gz
tar -xzf quast-5.0.2.tar.gz
cd quast-5.0.2
sudo ./setup.py install_full

This will install QUAST and set the path so that you will be able to run it from any direc-
tory. Notice that the above file path or name may change in a future version. QUAST can
be used to assess genome assemblies (contigs or scaffolds) generated by de novo assemblers.
It performs reference-guided assessment, which requires a reference genome sequence for
the species studied, or non-reference assessment, in which no reference sequence is used.
This is usually the case when the organism is unknown. In the following, we will assess the
de novo E. coli genomes assembled with ABySS and SPAdes above without using a refer-
ence sequence. You can copy the FASTA scaffold files created by these two programs into a
separate directory (e.g., “qc”) with new names using Linux command line as follows:

mkdir qc
cp abyss_ecoli_ass/ecoli-scaffolds.fa qc/abyss_ecoli_ass.fasta
cp spades_ecoli_ass/scaffolds.fasta qc/spades_ecoli_ass.fasta
cp hyb_spades_ecoli_ass/scaffolds.fasta qc/spades_hyb_ecoli_ass.
fasta

With the above commands, we created a directory “qc” and we copied the scaffolds’ files
created by ABySS and SPAdes into it.
The next step, change into the new directory “cd qc” and run the following QUAST
command to perform quality assessment for the three genome assemblies.
 De Novo Genome Assembly   ◾    101

quast.py \
-o quast_Ecoli_ass \
-t 4 \
abyss_ecoli_ass.fasta \
spades_ecoli_ass.fasta \
spades_hyb_ecoli_ass.fasta

If the following error is displayed:

“/home/…/jsontemplate.py” line 583, in <module> ‘html’: cgi.escape,
You may need to open that file “jsontemplate.py” using a text editor of your choice. In
line 50 of that file change “import cgi” to “import html” and in line 583 change “‘html’: cgi.
escape” to “‘html’: html.escape” and rerun the above command.
The above QUAST command saves the output files in “quast_Ecoli_ass” directory. The
assessment report is generated in different formats including pdf, html, text, and TSV. You
can open any of these reports with the right application. We can open “report.html” on
Firefox browser by running the following on the Linux terminal:

cd quast_Ecoli_ass
firefox report.html

This will display a statistics table, which includes assembly assessment statistics (Figure 3.8),
and three interactive plots: cumulative length, Nx, and GC content (two are shown in

FIGURE 3.8 QUAST assembly assessment report.

102   ◾    Bioinformatics

FIGURE 3.9 QUAST assembly assessment plots.

FIGURE 3.10 Icarus contig browser.

Figure 3.9). Use the mouse to display these three plots and also move the mouse to obtain
immediate information about the scaffold.
The assembly metrics shown in Figure 3.6 are the ones used for assembly assessment.
The important metrics are the total length of the assembly, N50, N75, L50, and L75. Figure
3.8 shows the metrics for three assemblies in three columns. The two assemblies generated
by SPAdes (non-hybrid and hybrid) are better than the assembly generated by ABySS. The
largest N50 value, the lowest L50 value, and longest assembly are indicative measures of
better assembly. The total lengths of the three assemblies are 4,609,549b (4.609549Mb),
4,686,651b (4.686651Mb), and 4,623,212b (4.623212Mb), respectively, which are close to
the length of the reference genome of E. coli str. K-12, which is 4.641650Mb. Notice also
that SPAdes assemblies have the largest N50 and N75 and the lowest L50 and L75.
The QUAST report also includes Icarus contig viewer, which is a genome viewer based
on QUAST for the assessment and analysis of genomic draft assemblies. Click “View in
Icarus contig browser” on the top right to display the Icarus QUAST contig browser.
In Figure 3.10, the Icarus visualizer shows contig size viewer, on which contigs are ordered
from the longest to the shortest, highlights N50, N75 (NG50, NG75) and long contigs larger
than a user-specified threshold. To learn more about Icarus viewer and QUAST output
reports, refer to the QUAST manual at “http://cab.cc.spbu.ru/quast/manual.html#sec3.4”.
If you need to assess the assembly using a reference as a guide, you can download the
FASTA file of a reference genome (Genome) with its annotation file (GFF) from a database
 De Novo Genome Assembly   ◾    103

(e.g., NCBI genome) for the organism, and use “-r” and “-g” options as follows (you may
need to decompress GFF file) (Figures 3.11 and 3.12):

quast.py \
-o ref_quast_Ecoli_ass \
-t 4 \
-r ecolref/GCF_000005845.2_ASM584v2_genomic.fna.gz \
-g ecolref/GCF_000005845.2_ASM584v2_genomic.gff \
abyss_ecoli_ass.fasta \
spades_ecoli_ass.fasta \
spades_hyb_ecoli_ass.fasta

3.3.2 Evolutionary Assessment for De Novo Genome Assembly

Rather than contig or scaffold length distributions such as N50 and L50, the evolutionary
assessment for a genome assembly is based on the completeness of a genome on the evo-
lutionary informed expectation of genes inferred from closely related orthologous groups
of sequences. They assess the completeness of a genome assembly in terms of gene content.
BUSCO (Benchmarking Universal Single-Copy Orthologs) [12, 13] is an evolutionary-
based quality assessment program that uses information of known genes from a database

FIGURE 3.11 QUAST assembly assessment report (reference-guided assessment).

104   ◾    Bioinformatics

FIGURE 3.12 Icarus contig browser displaying de novo assemblies aligned to a reference genome.

of orthologs (OrthoDB) to measure genome assembly and annotation completeness. The

quality of a genome assembly is described by some metrics including C for complete, D
for duplicate, F for fragmented, M for missing, and n for the number of genes used for the
assessment. The genes recovered from the de novo assembly are reported as complete (C)
when their lengths are within two standard deviations of the mean length of genes on the
ortholog database. Multiple copies of complete genes are reported as duplicate (D), which
is an indication of inaccuracy in the assembly of haplotypes. Incomplete or partially recov-
ered genes are reported as fragmented (F) and the unrecovered genes are reported missing
(M). The number of gene used (n) reflects the confidence of the assessment results.
BUSCO uses a number of third-party software packages that must be installed for the
program to run properly. The BUSCO dependencies include Python 3.x, BioPython, pan-
das, tBLASTn 2.2+, Augustus 3.2, Prodigal, Metaeuk, HMMER3.1+, SEPP, and R + ggplot2
for the plotting companion script. Some of these packages are needed in some cases. For
the complete installation instructions, visit the BUSCO website at “https://busco.ezlab.org/
busco_userguide.html”. You can run the following commands on the Linux command
line to install some BUSCO third-party dependencies and BUSCO software on Ubuntu:

sudo apt update && sudo apt upgrade

pip install biopython
pip install pandas
sudo apt-get install ncbi-blast+
sudo apt install augustus augustus-data augustus-doc
sudo apt install prodigal
sudo apt install hmmer

BUSCO software can be cloned and installed by running the following commands:

git clone https://gitlab.com/ezlab/busco.git

cd busco/
python3 setup.py install –user
 De Novo Genome Assembly   ◾    105

If the installation is successful, you can run the following to display the help:

busco –help

BUSCO databases include ortholog databases for several clades of organisms. Before using
BUSCO, you may need to identify the database to use for the assessment. The database list
can be displayed using the following command:

busco --list-datasets

Now, we can use BUSCO to assess the three E. coli assemblies (one generated with ABySS
and two generated by SPAdes). We can save the output of each assessment in a separate
directory.

busco \
-i abyss_ecoli_ass.fasta \
-o abyss_ecoli_ass.out \
-l bacteria \
-m genome

busco \
-i spades_ecoli_ass.fasta \
-o spades_ecoli_ass.out \
-l bacteria \
-m genome

busco \
-i spades_hyb_ecoli_ass.fasta \
-o spades_hyb_ecoli_ass.out \
-l bacteria \
-m genome

The BUSCO assessment output for each assembly will be saved in a separate directory:
“abyss_ecoli_ass.out”, “spades_ecoli_ass.out”, and “spades_hyb_ecoli_ass.out”. Each of
these directories includes an assessment report as a text file and JSON file, in addition to
subdirectories for the predicted genes and used ortholog database.
Comparing between the three assemblies based on BUSCO assessment metrics (Figures
3.13–3.15), the two assemblies generated by SPAdes are better than the one generated by
ABySS. A total number of 4085 genomes and 124 genes were used to extract informed
expected information. The E. coli assembled by SPAdes shows 100% completeness (C:100%),
no duplicate (D:0.0%), no fragments (F:0.0%), no missing gene (M:0.0%) out of the 124
genes, whereas the BUSCO assessment report for the assembly generated by SPAdes shows
C:98.4% [S:98.4%, D:0.0%], F:1.6%, M:0.0%, n:124, which indicates 98.4% of completeness
(122 genes are recovered), 1.6% of fragment (2 partially recovered genes).
Combining both statistical and evolutionary assessment for the de novo assembly will
provide a good idea about the quality of the de novo assembled genome.
106   ◾    Bioinformatics

FIGURE 3.13 BUSCO assessment for the E. coli assembly generated by ABySS (Illumina reads).

FIGURE 3.14 BUSCO assessment for the E. coli assembly generated by SPAdes (illumine reads).

FIGURE 3.15 BUSCO assessment for the E. coli assembly generated by SPAdes (Illumine reads).

3.4 SUMMARY
Genome assembly is the process of putting nucleotide reads generated by sequencers into
the correct order as in the genome of the organism. Despite the development of sequencing
technologies and assembling algorithms, the assembled genomes will remain approximate
 De Novo Genome Assembly   ◾    107

for the actual genomes. This is because it is hard to avoid errors in sequencing and also
genomes of many organisms including repetitive sequences. But the assembly accuracy
can be increased by deep sequencing, paired-end sequencing, and the use of long reads
produced by PacBio and Oxford Nanopore Technology. The de novo genome assembly
has been widely used for different kinds of organisms, specially in metagenomics for the
assembly of bacterial, fungal, and viral genomes.
We can use either greedy alignment approach or graph-based approaches for the de novo
genome assembly. The greedy method works as multiple sequence alignment by perform-
ing pairwise alignment and merging reads to build up contigs. The graph theory approach
can be either overlap-consensus graphs or de Bruijn graphs. In the overlap graphs, reads
are represented as nodes and the overlapping regions of the reads as edges. Contigs are
built by finding the Hamiltonian path which includes each node once. On the other hand,
de Bruijn graphs form k-mers from the reads and the k-mers then are represented as nodes.
Contigs are formed by including edges using Eulerian path. De Bruijn graphs are more
efficient than overlap graphs. The assemblers that use Bruijn graphs include ABySS for
small and large genomes and SPAdes for bacterial, fungal, and viral small genomes. SPAdes
program has several modules such as metagenomic module, viral assembly module, and
SARS-CoV2. SPAdes can also be used to assemble a genome from hybrid reads such as
Illumina reads and PacBio reads or Oxford Nanopore reads.
After assembling, a genome can be assessed using both statistical method and evolu-
tionary method. The statistical method generates the number, lengths, and distributions of
contigs. The assembly with few but long contigs is an indicator of the good quality. Metrics
used to describe statistical quality are N25, N50, N75, L25, L50, and L75. We can com-
pare the performance of assemblers using these metrics. The evolutionary assessment of an
assembly relies on the genomes of the evolutionarily related species to identify the number
of genes in the assembled genome. The completeness of the genome assembly depends on
the complete identified genes. For statistical quality assessment, we can use QUAST, and
for evolutionary assessment, we can use BUSCO.

REFERENCES
1. Lander ES, Waterman MS: Genomic mapping by fingerprinting random clones: a mathemati-
cal analysis. Genomics 1988, 2(3):231–239.
2. Pop M, Kosack D: Using the TIGR assembler in shotgun sequencing projects. Methods Mol
Biol 2004, 255:279–294.
3. de la Bastide M, McCombie WR: Assembling genomic DNA sequences with PHRAP. Curr
Protoc Bioinform 2007, Chapter 11:Unit11. 14.
4. Myers EW, Sutton GG, Delcher AL, Dew IM, Fasulo DP, Flanigan MJ, Kravitz SA, Mobarry
CM, Reinert KH, Remington KA et al: A whole-genome assembly of Drosophila. Science
2000, 287(5461):2196–2204.
5. Pevzner PA, Tang H: Fragment assembly with double-barreled data. Bioinformatics 2001,
17(suppl_1):S225–S233.
6. Pevzner PA, Tang H, Waterman MS: An Eulerian path approach to DNA fragment assembly.
Proc Natl Acad Sci U S A 2001, 98(17):9748–9753.
7. Chaisson MJ, Pevzner PA: Short read fragment assembly of bacterial genomes. Genome Res
2008, 18(2):324–330.
108   ◾    Bioinformatics

8. Simpson JT, Wong K, Jackman SD, Schein JE, Jones SJ, Birol I: ABySS: a parallel assembler for
short read sequence data. Genome Res 2009, 19(6):1117–1123.
9. Bankevich A, Nurk S, Antipov D, Gurevich AA, Dvorkin M, Kulikov AS, Lesin VM,
Nikolenko SI, Pham S, Prjibelski AD et al: SPAdes: a new genome assembly algorithm and its
applications to single-cell sequencing. J Comput Biol 2012, 19(5):455–477.
10. Medvedev P, Pham S, Chaisson M, Tesler G, Pevzner P: Paired de bruijn graphs: a novel
approach for incorporating mate pair information into genome assemblers. J Comput Biol
2011, 18(11):1625–1634.
11. Gurevich A, Saveliev V, Vyahhi N, Tesler G: QUAST: quality assessment tool for genome
assemblies. Bioinformatics 2013, 29(8):1072–1075.
12. Manni M, Berkeley MR, Seppey M, Simão FA, Zdobnov EM: BUSCO update: novel and
streamlined workflows along with broader and deeper phylogenetic coverage for scoring of
eukaryotic, prokaryotic, and viral genomes. Mol Biol Evol 2021, 38(10): 4647–4654.
13. Simão FA, Waterhouse RM, Ioannidis P, Kriventseva EV, Zdobnov EM: BUSCO: assessing
genome assembly and annotation completeness with single-copy orthologs. Bioinformatics
2015, 31(19):3210–3212.
 Chapter 4

Variant Discovery

4.1 INTRODUCTION TO GENETIC VARIATIONS

Classical genetics began with the works of Gregor Mendel in the 19th century as a field
of biology for studying hereditary in species based on morphology and visible results of
reproductive acts. Very recently, genetics has been greatly instrumented by bioinformatics
that emerged as a logical consequence of the modern progress in molecular biology, infor-
mation technology, and the urgent need for handling genome-scale data. Bioinformatics
has become a key in each step of genetic analysis and it provides standardized framework
for variant discovery and description, which are the essence of the modern genetics.
The general workflow of the research on genetic variation focuses on the analysis and
identification of genetic variants associated with specific phenotypes or populations.
Mutations are the key source of genetic variation within a species. A mutation is any
change in the nucleotide sequences of the genome of a living organism. The basic muta-
tions include base substitution, insertion, and deletion. However, translocation (exchange
of segments between chromosome) and copy number variation (CNV), which is multiple
copies of DNA segment, are also source of genetic variations. These mutations can occur
in parts of the genome. Base substitution is the most frequent one, and it has been thought
that it was a random process and it could occur anywhere in a genome with equal prob-
ability, but now such thoughts began to change after genomic studies in several organisms
showed that mutations are less frequent in important regions of the genome [1]. Therefore,
most genetic studies focus on the mutations that affect genes.
The single base substitution is called point mutation and it can be either transition
(when a purine is substituted with another purine or a pyrimidine is substituted with
another pyrimidine) or transversion (when a purine is substituted with a pyrimidine or
a pyrimidine replaces a purine). The base substitution may have a silent consequence if it
occurs in the third position of a codon and the result is a synonymous codon that codes
for the same amino acid and does not alter the protein sequence. The substitution will have
missense consequence if it results in a nonsynonymous codon that codes for a different
amino acid and alters the protein polypeptide sequence. The consequence of a substitution,

DOI: 10.1201/9781003355205-4 109

110   ◾    Bioinformatics

in this case, can either lead to no change in the structure and function of the protein
(­conservative mutation) or it may lead to a deleterious consequence if the change alters
the protein s­ tructure and function. The base substitution may also have a nonsense con-
sequence if it results in a stop codon that truncates the translated protein leading to an
incomplete and nonfunctional protein.
A deletion mutation is the removal of a single pair of nucleotides or more from a gene
that may result in a frameshift and a garbled message and nonfunctional product. Deletion
may have deleterious consequence or not depending on the part it alters and its impact
on the protein sequence. The insertion mutation is the insertion of additional base pairs
and it may lead to frameshifts depending on whether or not multiples of three base pairs
are inserted. Mutations may include combinations of insertions and deletions leading to a
variety of outcomes.
In general, a gene variant is a permanent change in the nucleotide sequence of a gene
that can be either germline variants, which occur in eggs and sperms of parents and pass
to offspring, or somatic variants, which are present only in specific cells and are generally
not hereditary.
In terms of sequence change, variants can be classified into single-nucleotide variant
(SNV), insertion–deletion (InDel), or structural variation (SV). The SNV is a base substi-
tution of a single nucleotide for another. It is known as single-nucleotide polymorphism
(SNP) if its allelic frequency in a population is more than 1%. InDel refers to insertion and/
or deletion of nucleotides into genomic DNA and it includes events less than 1000 nucleo-
tides in length. InDels are implicated as the driving mechanism underlying many diseases.
The SV involves change in more than 50 base pairs in a sequence of a gene; the change may
include rearrangement of part of the genome, a deletion, duplication, insertion, inversion,
translocation, or a combination of these. A CNV is a duplication or deletion that changes
the number of copies of a particular DNA segment within the genome. SVs have been
implicated in a number of health conditions.
In this chapter, we will learn about the major steps in the process of variant identifica-
tion and analysis, including variant representation, variant calling workflow, and variant
annotation. The process by which we identify variants from sequence data (reads) is called
variant calling, which is the central topic of this chapter.

4.1.1 VCF File Format

Since a variant is a change in a specific location in a genome, in bioinformatics, this
requires a format that can describe the type of a mutation and its position relative to the
genome coordinates. Thus, the variant call format (VCF) file [2] was developed to hold
the information of a large number of variants and also to hold genotype information of
multiple samples in the same position. The VCF file, as shown in Figure 4.1, consists of (i)
a metadata section for the meta-information and (ii) a data section for variant data. The
VCF file has become the standard file for storing variant information for almost all variant
calling programs.
Each line in the metadata section of a VCF file begins with “##”. The metadata lines
describe the format and content of a VCF file. This can include information about the
 Variant Discovery   ◾    111

sequencing, variant calling software that created the file, or the reference genome used for
determining variants. The first few lines of metadata section describe file format, file date,
source program, and the reference used. The metadata section also declares and describes
the fields provided at both the site-level (INFO) and sample-level (FORMAT) in the data
lines of the data section. For example, the following metadata lines describe the ID, data
type, and description of some fields that can be found in the INFO and FORMAT columns
in the data section:

##INFO=<ID=NS,Number=1,Type=Integer,Description=”Number of Samples
Data”>
##INFO=<ID=DP,Number=1,Type=Integer,Description=”Total Depth”>
##INFO=<ID=AF,Number=A,Type=Float,Description=”Allele Frequency”>
##FORMAT=<ID=GT,Number=1,Type=String,Description=”Genotype”>
##FORMAT=<ID=GQ,Number=1,Type=Integer,Description=”Genotype
Quality”>
##FORMAT=<ID=DP,Number=1,Type=Integer,Description=”Read Depth”>

The data section begins with a tab-delimited single header line that has eight mandatory
fields representing columns for each data line (Table 4.1). The column headers of the data
section are as follows:

#CHROM POS ID REF ALT QUAL FILTER INFO

Only if there is genotype data, then a FORMAT column is declared and followed by unique
sample names. All of these column names must be separated by tabs as well. Each line in
the data section represents a position in the genome. The data corresponds to the columns
specified in the header and must be separated by tabs and ended with a new line. Below are
the columns and their expected values. In all cases, MISSING values should be represented
by a dot (“.”).
As shown in Figure 4.1, the variants are in chromosome 20 on the reference genome
NCBI36 (hg18). The figure shows five positions whose coordinates are 14370, 18330,

TABLE 4.1 VCF File Columns

Column # Column Description
1 #CHROM A chromosome identifier (e.g., 11, chr11, X or chrX)
2 POS A reference position (sorted numerically in ascending order by chromosome)
3 ID Variant IDs separated by semicolons (no whitespaces allowed)
4 REF A reference base (A, C, G, or T). Insertions are represented by a dot
5 ALT A comma-separated alternate base(s) (A, C, G, or T). Deletions are represented by
a dot
6 QUAL A quality score in a log scale (Phred quality score)
7 FILTER This indicates which filters failed (semicolon-separated), PASS or MISSING
8 INFO A site-level information in semicolon-separated name-value format
9 FORMAT A sample-level field name declarations separated by semicolons
112   ◾    Bioinformatics

FIGURE 4.1 VCF file showing metadata and data sections.

1110696, 1230237, and 1234567. The first two variants are single point substitutions. The
third position has two alternate alleles (G and T) that replaced the ref nucleotide (A). The
fourth variant is a deletion of a single nucleotide T since the alt allele is missing (“.”). In the
fifth position, there are two alternative alleles, the first is a deletion of two nucleotides (T
and C) and the second is an insertion of a single nucleotide T.
The QUAL column holds the quality level of the data at each position. The FILTER column
designates what filters can be applied; the keywords in this column can be used to filter the
variants as we will discuss later. The second row (position 17330) does not pass the thresh-
old for the quality of more than 10 Phred quality score.
The INFO column includes position-level information for that data row and can be
thought as aggregate data that includes all of the sample-level information specified.
The FORMAT column specifies the sample-level fields to expect under each sample.
Each row has the same format fields (GT, GQ, DP, and HQ) except for the last row which
does not have HQ. Each of these fields is described in the metadata section. GT (Genotype)
indicates which alleles separated by / are unphased or | phased, GQ is the Genotype
Quality which is a single integer, DP is the Read Depth which is a single integer, and HQ is
the Haplotype Quality, and it has two integers separated by a comma.
This VCF file has three samples identified by their names (NA00001, NA00002, and
NA00003) in columns 10 through 12.
Genetic variants discovered by researchers are submitted, usually in VCF files, to data-
bases that archive information of the genetic differences with other related information.
Researchers submit data to these databases, which collect, organize, and publicly docu-
ment the evidence supporting links between genetic variants and diseases or conditions.
The variants are usually submitted with their assertions, which are informed assessments
of the association or lack of association between a disease or condition and a genetic vari-
ant based on the current state of knowledge. The variant databases include dbSNP (for
human variants of lesser than 50 base pairs), dbVar (for human variants of greater than 50
base pairs), and European Variation Archive (for variants of all species).
Variant submitted to a database is given a unique identifier that can be used in finding
that variant in the database and the related information because they are unambiguous,
 Variant Discovery   ◾    113

unique, and stable, unlike descriptive names, which can be used differently by different
people. For example, the NCBI dbSNP assigns an ID with “rs” prefix to the accepted human
variants with asserted positions mapped to a reference sequence as reference variants
(RefSNP) and also it assigns an ID with “ss” prefix for a variant submitted with flanking
sequence. Figure 4.1 shows two dbSNP IDs for reference variants in the VCF ID columns.
Variants may have identifiers from multiple databases. You will see these different types of
identifiers used throughout the literature and in other databases. Different types of identi-
fiers are used for short variants and structural variants.

4.1.2. Variant Calling and Analysis

Variant calling is the process by which we can identify variants on sequence data. The
sequence data are usually stored in FASTQ files obtained from whole genome, whole
exome sequencing, or targeted gene sequencing. The reads in the FASTQ files are assessed
for quality and then preprocessed to ensure that final reads are of high quality. The reads
are then aligned to a reference genome and the read alignment information are stored in
BAM files. The BAM files are then used as an input for variant calling programs for vari-
ant identification and analysis. The identified variants are written in a VCF file. A single
VCF file can hold thousands of variants and genotypes of multiple samples. The genetic
studies usually focus on the germline variant calling, where the reference genome used for
mapping the reads is standard for the species of interest; that will allow us to identify geno-
types. The somatic variant calling is used to study diseases like cancer. In somatic variant
calling, the reference is a related tissue from the same individual (e.g., healthy tissue in the
case of cancer). Here, we expect to see genetic mosaicism between cells or presence of more
than one genetic line as a result of genetic mutations.
A variant calling workflow begins with raw sequencing data for multiple samples or
individuals and ends with a single VCF file containing only the genomic positions where at
least one individual in the population has a variant due to mutations. After variants have
been called, they can be analyzed in different ways. For example, we may wish to deter-
mine which genes are affected by the variants, what consequences they have on them, and
the phenotypes associated with them. Thus, variants that have been called can be anno-
tated with their consequences and can also be associated to certain phenotypes and the
results can be interpreted to answer some research questions.
Before digging into the steps of variant calling and analysis, it is better to distin-
guish between the types of genetic variation studies. There are several types of genetic
­variation studies but generally they can be classified into (i) Genome-wide association
­studies (GWASs) [3], (ii) studies on consequences of variants [4], and (iii) Population
genetics [5].
The GWASs involve genotyping a sample of individuals at common variants across the
genome using a genome-wide survey for variants. Variants associated with a phenotype
will be found at a higher frequency. This kind of studies are carried out on individuals to
identify variants and their associated phenotypes as variants causing the phenotype will be
at higher frequency in the affected individual than in the control. The phenotype–genotype
associations must be supported by statistical evidence based on the population studied.
114   ◾    Bioinformatics

The studies on the consequences of variants focus on understanding the molecu-

lar mechanisms and pathways that link a genotype to a phenotype. This kind of studies
interpret the consequences of the variants on the protein function. Simple base substitu-
tion such as missense, stop gained, and stop lost variants can alter the translated protein
sequence, causing functional consequences. Moreover, functional consequences due to
structural variants are usually defined by the physiological phenotypes observed. These
can be complex descriptions which are described using general phenotypic traits rather
than specific biochemical effects caused by the variant. Clinical functional consequences
are represented by a simple controlled vocabulary that defines the relative pathogenicity
of a variant, such as benign, likely benign, uncertain significance, likely pathogenic, or
pathogenic.
The studies of population genetics are the studies of variation within populations of
individuals and the forces that shape it. This usually involves studying changes in frequen-
cies of genetic variation in populations over space and time. Some of the major forces that
shape variation in natural populations are mutations, selection, migration, and random
genetic drift. When a new mutation occurs, it may be beneficial to the organism, deleteri-
ous (harmful) to the organism, or it can be neutral (have no effect on the fitness of the
organism). Indeed, beneficial and deleterious mutations are subject to natural selection,
typically leading to increases and decreases in their allele frequency, respectively. Allele
frequencies are also influenced by the random genetic drift. This process explains the fluc-
tuation in allele frequencies from one generation to another.

4.2 VARIANT CALLING PROGRAMS

There are several programs for variant calling using different variant calling algorithms.
The most commonly used variant calling programs are categorized into two groups: con-
sensus-based callers like BCFTools mpileup and haplotype-based callers like FreeBayes
and GATK HaplotypeCaller. In the following, we will discuss these two types of variant
callers with some examples. We will assume that the FASTQ files used in the exercise are
preprocessed and clean as explained in Chapter 1.

4.2.1 Consensus-Based Variant Callers

The consensus-based variant callers depend on the pileup of the aligned reads covering a
position on the reference sequence to call the variants (SNVs or InDels). The read align-
ment information is in SAM/BAM file. We can then check the pileup of all bases of the
reads covering a reference base position. In most cases, the bases covering that position
will be the same as the base of the reference sequence, but in the case of variants, the bases
will be different from the reference base. The consensus sequence is created by collapsing
bases on all position and choosing the most frequent bases. In some positions, there may be
differences between the sequence of the reference genome and consensus sequence. These
differences can also be due to errors; however, when there is a sufficient sequencing depth,
that will provide sufficient confidence to call the variants. Figure 4.2 shows a diagram for
reads aligned to the sequence of a reference genome and a consensus sequence formed by
 Variant Discovery   ◾    115

FIGURE 4.2 Read alignment and a consensus sequence.

collapsing bases in each position and identifying the most frequent base. In the figure, you
can notice how depth is important in determining the consensus base.

4.2.1.1 BCF Tools Variant Calling Pipeline

In the following, we will use BCFtools as a consensus variant caller to pileup and call vari-
ants. BCFtools [6] is a set of tools for the manipulation of variant in VCF files and its binary
form BCF. To download the latest BCFtools version, follow the installation instructions
available at “http://www.htslib.org/”. On Linux and Linux like platforms, you can also use
the following commands to install it:

sudo apt-get update -y

sudo apt-get install -y bcftools
Run the following commands on the Linux terminal to learn about
the usage and options of the “bcfools” and “bcftools mpileup”:
bcftools --help
bcftools mpileup

The general form for the variant calling is as follows:

116   ◾    Bioinformatics

bcftools mpileup -Ou \

-f reference_file \
-b bam_list_file \
| bcftools call -vmO z \
-o sarscov2.vcf.gz

The “bcftools mpileup” is the command used for the pileup. The “-f” option specifies the
reference file used by the aligner to produce the BAM file, “-b” option specifies the BAM
file from which we wish to call variants. The above form allows both pileup and variant
calling.
In the following, we will discuss an example of variant calling pipeline using “bcftools”.
We will use SARS-CoV-2 raw sequence data from the NCBI SRA database to demonstrate
variant calling.
SARS-CoV-2 is the virus that causes COVID-19 which affected millions of people
around the world and caused thousands of deaths. The virus mutates in a short period of
time, producing new strains. The following are the NCBI SRA run unique identifiers of raw
data generated by whole genome sequencing of SARS-CoV-2:

SRR16989486
SRR16537313
SRR16537315
SRR16537317

In the following, we wish to identify variants (SNVs and InDels) from short reads of those
four samples and report that in a VCF file.
The programs used with this example include SRA toolkits, wget, gzip, samtools, bwa,
and bcftool on Linux.
The workflow for variant discovery includes: (i) acquiring the raw data (FASTQ files), (ii)
quality control if required, (iii) downloading and indexing the reference genome, (iv) using
an aligner to map reads in FASTQ files to a reference genome to generate a BAM file, (v)
sorting the aligned reads in BAM files, (vi) removal of duplicate reads from the BAM files,
and (vii) performing variant calling and generation of a single VCF file for genotyping of all
samples. Any new generated BAM files must be indexed before proceeding to the next step.
In the following, we will discuss each of these steps. First, we need to open the Linux ter-
minal, create a directory called “sarscov2” for this project, and make it the current working
directory.

mkdir sarscov2
cd sarscov2

4.2.1.1.1 Acquiring the raw data

The raw data files are stored in the NCBI SRA database with above SRA run IDs. The
FASTQ files are in sequence read archive (SRA) format that requires SRA toolkits to
be downloaded and extracted. For that purpose, we can use “fasterq-dump” with the
­following syntax:
 Variant Discovery   ◾    117

fasterq-dump --progress --outdir fastq id

The “--progress” option is to display the downloading progress, “--outdir fastq” specifies
the directory where FASTQ files are downloaded, and “id” is replaced by any of the above
SRA run IDs.
The above “fasterq-dump” form is suitable for a single run, but what if we have multiple
run IDs as above, or in some cases, we may have tens of IDs to download and running that
command for each ID would be tedious. In such case, bash “while loop” would come in
handy. First, we need to store the above run IDs in the file “ids.txt”, each run ID in a line,
and save the file in the current directory and then run the following bash script, which cre-
ates the subdirectory “fastq” and then uses “while loop” to loop over each run ID in the text
file and use it as an argument for the “fasterq-dump” command as follows:

mkdir fastq
while read id;
do
fasterq-dump --progress --outdir fastq “$id”
done < ids.txt

The above script creates the directory “fastq” and downloads the FASTQ files into it. There
are two FASTQ files for each sample since the reads are paired end (forward and reverse

FIGURE 4.3 Using fasterq-dump to download FASTQ files from the NCBI SRA database.
118   ◾    Bioinformatics

FASTQ files). When the FASTQ files have been downloaded successfully as shown in
Figure 4.3, we can use “ls fastq” to display the files in the new created directory.

4.2.1.1.2 Downloading and indexing the reference genome sequence

The reads in the FASTQ files must be aligned to a reference genome of the organism stud-
ied. Therefore, the latest FASTA file of the reference genome sequence is downloaded from
a genome database into a local drive. The NCBI Genome database “https://www.ncbi.nlm.
nih.gov/genome/” is one of the databases that curates reference genome sequences. We can
use the database query box to search for the latest reference genome of “SARS-CoV-2” and
copy the link to the FASTA sequence of the reference genome. The following bash script
creates the “ref” subdirectory, downloads the compressed FASTA sequence of the latest
SARS-CoV-2 reference genome into that subdirectory, and decompresses the FASTA file.
Notice that the URL of the reference sequence may change if a new version is available.
Therefore, visit the reference genome page for the latest sequence. When you use the fol-
lowing script, make sure that “wget” and the URL are in the same line and that there is no
whitespace in the URL. After downloading the reference sequence, use “ls” to make sure
the file has been downloaded and decompressed.

mkdir ref
cd ref
wget https://ftp.ncbi.nlm.nih.gov/genomes/all/GCF/009/858/895/
GCF_009858895.2_ASM985889v3/GCF_009858895.2_ASM985889v3_genomic.
fna.gz
f=$(ls *.*)
gzip -d ${f}

4.2.1.1.3 Indexing the FASTA file of the reference genome

As discussed in Chapter 2, we need to index the FASTA sequence of the reference genome
with both “samtools faidx” and the aligner used for mapping. In this example, we will use
“bwa” aligner; therefore, we will use “bwa index” for indexing as well. The following bash
script uses “samtools” and bwa” to index the reference genome:

f=$(ls *.*)
samtools faidx ${f}
bwa index ${f}
cd ..

When you display the content of the “ref” subdirectory, you may see the following files if
you follow the above steps successfully:

GCF_009858895.2_ASM985889v3_genomic.fna
GCF_009858895.2_ASM985889v3_genomic.fna.amb
GCF_009858895.2_ASM985889v3_genomic.fna.ann
GCF_009858895.2_ASM985889v3_genomic.fna.bwt
 Variant Discovery   ◾    119

GCF_009858895.2_ASM985889v3_genomic.fna.fai
GCF_009858895.2_ASM985889v3_genomic.fna.pac
GCF_009858895.2_ASM985889v3_genomic.fna.sa

4.2.1.1.4 Aligning short reads of FASTQ files to the reference genome

Since we have downloaded the FASTQ files and the sequence of the reference genome, it
is time to map the reads to the reference genome and to create a SAM file for each sample.
Remember that since the reads are paired end, there are two files (forward and reverse) for
each run. This time we will use the following syntax of “bwa mem” command:

bwa mem -M -t 4 \
-R “@RG\tID:….\tSM:…..” \
ReferenceSequence \
file1.fastq file2.fastq > output.sam 2> output.log

The “-M” option is to mark shorter split hits as secondary, “-t” option specifies the num-
ber of threads to speed up the mapping, and “-R” to add a header line. For variant calling,
we may need to add or create a read group (RG) header to hold the sample information.
As input, we will provide the FASTA reference sequence and the forward FASTQ file and
reverse FASTQ file in the case of paired-end reads. The output can be directed to a SAM
file.
The above form is suitable for a single run. However, we may have multiple samples and
mapping the reads of a single-sample files at a time may be tedious. Instead, we can use a
bash script with a loop to align all runs. The following script creates the subdirectory “sam”
and then aligns reads of each sample to the reference genome and outputs a SAM file and a
log file. When you run the script, make sure that the main project directory is the working
directory.

mkdir sam
cd fastq
for i in $(ls *.fastq | rev | cut -c 9- | rev | uniq);
do
bwa mem -M -t 4 \
-R “@RG\tID:${i}\tSM:${i}” \
../ref/GCF_009858895.2_ASM985889v3_genomic.fna \
${i}_1.fastq ${i}_2.fastq > \
../sam/${i}.sam 2> ../sam/${i}.log;
done
cd ..

To make sure that the read mapping process has been completed successfully, you can dis-
play the content of the “sam” directory with “ls sam”. You can also display the content of
any of the SAM files by using “less -S SamFileName”.
120   ◾    Bioinformatics

Notice that we added a RG header “@GR” with “ID” and “SM” to each SAM file to pro-
vide the sample information that will be used later by the variant caller to create a genotype
column for each sample.

4.2.1.1.5 Converting SAM into BAM format

The “bwa mem” aligner created a SAM file. We can convert the SAM files into a BAM for-
mat by running the following script. Make sure that you are running the script from the
main project directory.

mkdir bam
cd sam
for i in $(ls *.sam | rev | cut -c 5- | rev);
do
samtools view -uS -o ../bam/${i}.bam ${i}.sam
done
cd ..

The above script will create the subdirectory “bam” and convert the SAM files into BAM
files and stores them in the “bam” subdirectory.

4.2.1.1.6 Sorting and indexing BAM files

The BAM files must be sorted and indexed before the next step of the analysis. The follow-
ing bash script creates a new subdirectory “sortedbam”, sorts the bam files, and saves the
new BAM files with sorted alignment in the new subdirectory:

mkdir sortedbam
cd bam
for i in $(ls *.bam);
do
samtools sort -T ../sortedbam/tmp.sort -o ../sortedbam/${i} ${i}
done
cd ..
cd sortedbam
for i in $(ls *.bam);
do
samtools index ${i}
done
cd ..

4.2.1.1.7 Marking duplicates

The identical reads mapped to the reference genome may bias variant calling. The dupli-
cate reads are created by the PCR amplification during sequencing. Usually, using small
amount of DNA for library preparation may lead to more duplication. Some sequencing
protocols use Unique Molecular Identifiers (UMIs) to distinguish between the artifact and
natural duplicates. Marking or removing duplicate aligned reads is a common best practice
 Variant Discovery   ◾    121

in whole genome sequencing. However, several studies have shown that retaining PCR and
Illumina clustering duplicates does not cause significant artifacts as long as the library
complexity is sufficient. The duplicate alignments can be removed with “samtools rmdup”.
The following script creates a new subdirectory “dupRemBam” and then removes duplicate
alignments from the BAM files and stores the new BAM file in the new subdirectory:

mkdir dupRemBam
cd sortedbam
for i in $(ls *.bam);
do
samtools rmdup ${i} ../dupRemBam/${i} 2> ../dupRemBam/${i}.log
done;
cd ..

4.2.1.1.8 Varian calling with bcftools

It is time to use the bcftools for variant calling. The bcftools is a consensus variant caller as
discussed above. It takes a single or multiple BAM files as input and outputs a compressed
VCF file. The following script creates the subdirectory “variants” in the main project direc-
tory and then it uses “bcftools mpileup” for performing pileup and “bcftools call” to call
the variants for the piled-up reads:

mkdir variants
cd dupRemBam
ls *.bam | rev | rev > bam_list.txt
bcftools mpileup -Ou \
-f ../ref/GCF_009858895.2_ASM985889v3_genomic.fna \
-b bam_list.txt \
| bcftools call -vmO z \
-o ../variants/sarscov2.vcf.gz
cd ..

The VCF file that contains the variants (SNPs and InDels) will be saved in “variants” sub-
directory. For further analysis, the VCF file can be indexed using “tabix”, which is a generic
indexer for TAB-delimited genome position files.

tabix variants/sarscov2.vcf.gz

We can view the VCF file with “bcftool view” as follows:

bcftools view variants/sarscov2.vcf.gz | less -S

All steps from downloading the raw data to variant calling can be included in a single bash
file that can be executed on the command-line prompt of the Linux terminal. First, create
a new directory, make that directory the current working directory, and then create a file
with the run ids “ids.txt” as described above. Then, create a bash file “pipeline_bcftools.
122   ◾    Bioinformatics

sh” and copy and save the following script on it and then execute the bash file as “bash
pipeline_bcftools.sh”:

#!/bin/bash
#Sars-Cov2 variant calling
#-------------------------
#1- download fastq files from the NCBI SRA database
mkdir fastq
while read f;
do
fasterq-dump --progress --outdir fastq “$f”
done < ids.txt
#2- download and extract the reference genome
mkdir ref
cd ref
wget https://ftp.ncbi.nlm.nih.gov/genomes/all/GCF/009/858/895/
GCF_009858895.2_ASM985889v3/GCF_009858895.2_ASM985889v3_genomic.
fna.gz
#Extract the compressed the reference FASTA file
f=$(ls *.*)
gzip -d ${f}
#3- Index the reference FASTA file using samtools and bwa
f=$(ls *.*)
samtools faidx ${f}
bwa index ${f}
cd ..
#4- Align the fastq reads (multiple samples) to the reference
genome
mkdir sam
cd fastq
for i in $(ls *.fastq | rev | cut -c 9- | rev | uniq);
do
bwa mem -M -t 4 \
-R “@RG\tID:${i}\tSM:${i}” \
../ref/GCF_009858895.2_ASM985889v3_genomic.fna \
${i}_1.fastq ${i}_2.fastq > \
../sam/${i}.sam 2> ../sam/${i}.log;
done
cd ..
#5- convert SAM files into BAM files
mkdir bam
cd sam
for i in $(ls *.sam | rev | cut -c 5- | rev);
do
samtools view -uS -o ../bam/${i}.bam ${i}.sam
done
cd ..
 Variant Discovery   ◾    123

#6- Sorting and indexing BAM files

#Sorting BAM files
mkdir sortedbam
cd bam
for i in $(ls *.bam);
do
samtools sort -T ../sortedbam/tmp.sort -o ../sortedbam/${i}
${i}
done
cd ..
#Indexing the sorted BAM files
cd sortedbam
for i in $(ls *.bam);
do
samtools index ${i}
done
cd ..
#7- Marking/removing duplicate alignments
mkdir dupRemBam
cd sortedbam
for i in $(ls *.bam);
do
samtools rmdup ${i} ../dupRemBam/${i} 2> ../dupRemBam/${i}.log
done;
cd ..
#8- Alignment pileup and variant calling using bcftools
mkdir variants
cd dupRemBam
for i in $(ls *.bam);
do
samtools index ${i}
done;
ls *.bam | rev | rev > bam_list.txt
bcftools mpileup -Ou \
-f ../ref/GCF_009858895.2_ASM985889v3_genomic.fna \
-b bam_list.txt \
| bcftools call -vmO z \
-o ../variants/sarscov2.vcf.gz
cd ..
#9- Index the VCF file
tabix variants/sarscov2.vcf.gz
#to view contents
bcftools view variants/sarscov2.vcf.gz | less -S

The resulted VCF file, which contains the genotypes of all samples (Figure 4.4), can be
further analyzed as we will discuss at the end of this chapter. However, before analysis, the
identified variants must be filtered.
124   ◾    Bioinformatics

FIGURE 4.4 VCF file displayed in a spreadsheet.

4.2.1.1.9 Filtering variants

Once variants have been called and identified, it is important to filter out low-quality vari-
ants before annotation and interpretation. The traditional variant filtering performs filter-
ing by checking QUAL field and some annotations listed in INFO and FORMAT fields of
a VCF file such as MQ (Mapping quality), DP (read depth), FS (FisherStrand), and SOR
(StrandOddsRatio) and choosing threshold values to filter out the variants that do not meet
the criteria. The base call quality (QUAL) is the most used filter parameter. The traditional
VCF filtering is performed with bcftools which uses either “bcftools filter” or “bcftools
view” with options “-i” and “-e” to filter-in or filter-out variants. The FILTER field of the
VCF file is empty “.” before forcing any filter as shown in Figure 4.4. Once a filter has been
applied, a variant in the VCF file either passes or fails. If it passes, PASS will be added to
FILTER field and any other variant that does not meet the criterion will be filtered out leav-
ing only the ones with PASS (see Figure 4.5). The following command filters in the variants
with Phred quality scores greater than 60:

bcftools filter -O z \
-o filtered_sarscov2.vcf.gz \
-i ‘%QUAL>60’ sarscov2.vcf.gz

The same results can be obtained with the following command, which filters out variants
with quality score less than 60:

bcftools view -O z \
-o filtered_sarscov2.vcf.gz \
-e ‘QUAL<=60’ sarscov2.vcf.gz

However, we can also filter variants based on other criteria using the statistics or annota-
tions in INFO or FORMAT fields. For example, you may decide to filter out variants with
depth less than 300; thus, you can use the following command:
 Variant Discovery   ◾    125

FIGURE 4.5 VCF file containing filtered variants.

bcftools filter -O z \
-o filtered2_sarscov2.vcf.gz \
-i ‘DP>300’ filtered_sarscov2.vcf.gz

You can open the filtered VCF file to notice the changes and that the filter command will
be added to the VCF header.
Usually, you can implement different filters on the variants in a VCF file to achieve accu-
rate and reliable results.
Another way to filter variants is to use “bcftools isec” with truth variants in a VCF file
as input together with your raw VCF file to create intersections, unions, and complements
of the VCF files.

bcftools isec -c both -p isec truth.vcf.gz input.vcf.gz

Refer to bcftools help for more details.

4.2.2 Haplotype-Based Variant Callers

The haplotype-based variant calling programs usually use Bayesian probabilistic model
to predict variants on aligned reads based on a haplotype structure of variants rather
than only sequence alignment. The haplotype is a set of genetic variants that are inherited
together. The haplotype-based variant detection depends on the physical phasing, which
is the process of inferring haplotype structure based on genotypic data using the Bayesian
approach. The prediction is based on relating the probability of a specific genotype given a
set of reads to the likelihood of sequencing errors in the reads and the prior likelihood of
specific genotypes. The Bayesian haplotype-based approach allows modeling multiallelic
loci that enables direct detection of a longer, multi-base alleles from sequence alignment.
Using reads aligned to a reference sequence (BAM) as input, haplotype-based algorithm
first attempts to identify the active regions of variations on the reads aligned to the refer-
ence genome. The identification of the active regions is carried out with dynamic sliding
windows of certain size along the reference sequence. The number of events (mismatches,
InDels, and soft clips) is counted in each window. When the number of events in that
126   ◾    Bioinformatics

window passes a threshold, then that window will be identified as an active region. A mea-
sure like entropy may be used to measure the activity on the region.
The haplotypes are constructed from the reassembled reads following the identifica-
tion of the active regions. The de Bruijn-like graph is used to reassemble the active region
and to identify the possible haplotypes present in the alignments. Once the haplotypes are
determined, the original alignment of the reads will be ignored and the candidate haplo-
types are realigned to the haplotypes of the reference genome using the Smith-Waterman
local alignment. The pairwise alignment is also performed using Pairwise Hidden Markov
Model (PairHMM) which generates a likelihood matrix of haplotypes given. These likeli-
hoods are then marginalized to obtain the likelihoods of alleles for each potentially variant
site given the read data. The genotype or the most likely pair of alleles is then determined
for each position. For a given genotype (Gi) on a subset of overlapped reads (Ri), the variant
callers then use Bayesian statistics to evaluate the posterior probability of the hypothetic
phenotype (Gi) as follows:

P( Ri | Gi ) × P (Gi )
P (Gi |Ri ) = (4.1)
P ( Ri )
where the posterior probability P (Gi |Ri ) is the probability of the phenotype (Gi) given that
subset of reads (Ri), P (Gi ) is the prior probability that we expect to observe the genotype
based on previous observations, P ( Ri ) is the probability of the subset of the reads being
true (the probability of observing the evidence), and P( Ri | Gi ) is the probability of reads
given the genotype. The Bayesian variant caller writes the above formula as:

P( Ri | Gi ) × P (Gi )
P (Gi |Ri ) = (4.2)
∑ P( R | Rk ) × P (Gk )
k

We can ignore the denominator because it is the same for all genotypes. Thus

P (Gi |Ri ) ∝ P( Ri | Gi ) × P (Gi ) (4.3)

The variant callers use a flat prior probability that can be changed by the users if the
probabilities of the genotypes are known based on previous observations. The important
probability in the above formula is P( Ri | Gi ), which can also be described in terms of the
likelihood of the hypothesis of the genotype (Gi ) given the reads (Ri ):

 L ( R j | H1 ) L ( R j | H 2 ) 
P ( Ri |Gi ) = L (Gi |Ri ) = ∏
j

 2
+
2


(4.4)

where L ( R j | H1 ) and ( R j | H 2 ) are the haplotype likelihoods.

The likelihoods of all possible genotypes are calculated based on the alleles that were
observed at the site, considering every possible combination of alleles. Then, the most likely
 Variant Discovery   ◾    127

genotype is called and assigned to the sample, and positions with putative variants (substi-
tutions or InDels) are written into the VCF file.
In the following, we will perform variant calling with both FreeBayes and GATK, which
are examples of haplotype-based variant callers.

4.2.2.1 FreeBayes Variant Calling Pipeline

FreeBayes [7] is a haplotype-based and Bayesian variant detector that is used to detect small
variants such as single and multiple-nucleotide polymorphisms and InDels. On Linux, we
can install FreeBayes as follows:

sudo apt update

sudo apt install freebayes

Use the following command to read more about FreeBayes usage and options:

freebayes –help

For variant calling, we will use the following form:

freebayes \
-f ../ref/GCF_009858895.2_ASM985889v3_genomic.fna \
-C 5 \
-L bam_list.txt \
-v ../variants/sarscov2.vcf

The “-f” option specifies the reference file, “-C” specifies the minimum number of observa-
tions supporting an alternate allele within a single individual in order to evaluate the posi-
tion, “-L” will pass the name of the text file that contains the names of the BAM files (each
file name in a line), “-v” will pass the VCF file name.
We will use FreeBayes in the above example to identify variants in the SARS-CoV-2
samples. We will follow the same steps we did for “bcftools” above. First, we will create a
project directory and store the run IDs in a file “ids.txt” as above. Then, we will save the
following script in a file “pipeline_freebayes.sh” and execute it as “bash pipeline_freebayes.
sh”:

#!/bin/bash
#Sars-Cov2 variant calling
#-------------------------
#1- download fastq files from the NCBI SRA database
mkdir fastq
while read f;
do
fasterq-dump --progress --outdir fastq “$f”
done < ids.txt
128   ◾    Bioinformatics

#2- download and extract the reference genome

mkdir ref
cd ref
wget https://ftp.ncbi.nlm.nih.gov/genomes/all/GCF/009/858/895/
GCF_009858895.2_ASM985889v3/GCF_009858895.2_ASM985889v3_genomic.
fna.gz
#Extract the compressed the reference FASTA file
f=$(ls *.*)
gzip -d ${f}
#3- Index the reference FASTA file using samtools and bwa
f=$(ls *.*)
samtools faidx ${f}
bwa index ${f}
cd ..
#4- Align the fastq reads (multiple samples) to the reference
genome
mkdir sam
cd fastq
for i in $(ls *.fastq | rev | cut -c 9- | rev | uniq);
do
bwa mem -M -t 4 \
-R “@RG\tID:${i}\tSM:${i}” \
../ref/GCF_009858895.2_ASM985889v3_genomic.fna \
${i}_1.fastq ${i}_2.fastq > \
../sam/${i}.sam 2> ../sam/${i}.log;
done
cd ..
#5- convert SAM files into BAM files
mkdir bam
cd sam
for i in $(ls *.sam | rev | cut -c 5- | rev);
do
samtools view -uS -o ../bam/${i}.bam ${i}.sam
done
cd ..
#6- Sorting and indexing BAM files
#Sorting BAM files
mkdir sortedbam
cd bam
for i in $(ls *.bam);
do
samtools sort -T ../sortedbam/tmp.sort -o ../sortedbam/${i}
${i}
done
cd ..
#Indexing the sorted BAM files
cd sortedbam
 Variant Discovery   ◾    129

cd ..
#7- Marking/removing duplicate alignments
mkdir dupRemBam
cd sortedbam
for i in $(ls *.bam);
do
samtools rmdup ${i} ../dupRemBam/${i} 2> ../dupRemBam/${i}.log
done;
cd ..
#8- Alignment pileup and variant calling using bcftools
mkdir variants
cd dupRemBam
for i in $(ls *.bam);
do
samtools index ${i}
done;
ls *.bam | rev | rev > bam_list.txt
freebayes \
-f ../ref/GCF_009858895.2_ASM985889v3_genomic.fna \
-C 5 \
-L bam_list.txt \
-v ../variants/sarscov2.vcf
cd ..
#9- Compress and index the VCF file
bgzip variants/sarscov2.vcf
tabix variants/sarscov2.vcf.gz
#to view contents
bcftools view variants/sarscov2.vcf.gz | less -S

We may notice that there is a little difference between the VCF file generated by bcftools
and the one generated by FreeBayes. The reason is that each program uses a different algo-
rithm for variant calling as discussed above.

4.2.2.2 GATK Variant Calling Pipeline

Genome Analysis Toolkit (GATK) [8] is a collection of command-line tools for variant dis-
covery and analysis of the sequence data acquired from the high-throughput sequencing
instruments. For installing the latest release of GATK, follow the installation instructions
available at the GATK website “https://gatk.broadinstitute.org/”. GATK uses the haplo-
type approach as FreeBayes. However, it uses advanced workflow for variant calling. The
GATK tools can be used individually or together as a complete pipeline for variant discov-
ery. GATK also provides complete workflows called GATK Best Practices for variant call-
ing tailored for specific cases. In the following, we will discuss the most commonly used
variant calling pipeline for GATK best practice. The following workflow assumes that the
acquired raw data in FASTQ format and that the quality control has been performed on the
data. The general GATK workflow consists of the following major steps:
130   ◾    Bioinformatics

1. Acquiring the raw data in FASTQ files.

2. Downloading the sequence of the reference genome, indexing (with samtools and
bwa) the reference sequence, and creating a dictionary (with Picard) for the reference
sequence.
3. Mapping the raw reads in FASTQ files to a reference genome using any of the aligner
(e.g., BWA) to generate aligned reads in SAM files.
4. Converting SAM files to BAM files using samtools.
5. Sorting and indexing the alignment in the BAM files using samtools.
6. Extracting the chromosome or the interval of interest, otherwise, skipping this step
if you are interested in the variants of the whole genome.
7. Removing or marking duplicate alignments in the BAM files using Picard and index-
ing the resulted BAM with samtools.
8. Adding RG header to the BAM files using Picard and indexing the resulted BAM files
with samtools.
9. Applying Base Quality Score Recalibration (BQSR) for detecting systematic errors
made by the sequencer.
10. Identification of the genotype at all positions per sample. This step involves
“HaplotypeCaller”, which takes a BAM file as input and outputs a GVCF file format
for each sample.
11. The next step in the GATK workflow is to consolidate the GVCF files of the multiple
samples and the variants are called across all samples in the cohort forming a single
VCF file. This step involves GenomicsDBImport and GenotypeGVCFs.
12. Separation of SNPs and InDels in separate VCF files using GATK SelectVariants.

4.2.2.2.1 Acquiring raw data

The raw sequence data usually is in FASTQ format. This time, we will use human raw data
from 1000 Genome project to demonstrate variant calling using GATK4 best practice. For
simplicity, instead of the whole genome, we will extract only the reads of chromosome
21 of 13 individuals including males and females from Africa, Asia, and Europe. Table
4.2 shows the NCBI SRA data from which the FASTQ files for chromosome 21 will be
extracted and sample information.
To keep the files organized, we can create a directory for this project “vcfgatk”, and
inside the directory, we will create a text file “ids.txt” that contains only the above ID col-
umn without a column header and each ID in a line. Then, while you are inside the project
directory, you can save the following bash script in a file “download.sh” and execute it as
“bash download.sh”:
 Variant Discovery   ◾    131

TABLE 4.2 The NCBI SRA Run IDs and Individual

Information
Run ID Country Population Gender
ERR1019055 China East Asia Female
ERR1019056 China East Asia Male
ERR1019057 China East Asia Female
ERR1019081 Pakistan South Asia Male
ERR1025616 Pakistan South Asia Male
ERR1019044 Kenya Africa Female
ERR1025600 Kenya Africa Female
ERR1025621 Nigeria Africa Male
ERR1025640 Senegal Africa Male
ERR1019034 Russia Europe Male
ERR1019045 France Europe Male
ERR1019068 Italy Europe Male
ERR1025614 Bulgaria Europe Male

mkdir fastq
cd fastq
while read id;
do
sam-dump \
--verbose \
--fastq \
--aligned-region chr21 \
--output-file ${id}_chr21.fastq \
${id}; \
done < ../ids.txt
cd ..

The download may take a long time depending on the speed of the Internet connection and
computer memory and processing units. The above script will create the “fastq” directory
and save the FASTQ files of chromosome 21 of 13 human individuals.

4.2.2.2.2 The reference genome

The FASTA sequence of the reference genome is required for reads mapping. We can
download it from a reliable database such as the NCBI Genome database. However, for
GATK pipeline, the sequence of the human genome can be downloaded from GATK
resource bundle, which is a collection of standard files prepared to be used with GATK.
The resource bundle is hosted on a Google Cloud bucket and can be accessed with a google
account using the following address:
https://console.cloud.google.com/storage/browser/genomics-public-data/resources/
broad/hg38/v0/
132   ◾    Bioinformatics

With the FASTA sequence, we can also download the sequence dictionary file. Once we
have downloaded the FASTA sequence of the reference genome, we can index the sequence
with both samtools and bwa. The following script first creates the directory “refgenome”,
and then, in that directory, it downloads the sequence of the current human genome and
its dictionary file from GATK resource bundle and indexes the FASTA:

mkdir refgenome
cd refgenome
wget https://storage.googleapis.com/genomics-public-data/
resources/broad/hg38/v0/Homo_sapiens_assembly38.fasta
wget https://storage.googleapis.com/genomics-public-data/
resources/broad/hg38/v0/Homo_sapiens_assembly38.dict
#or create a fasta dict for reference
f=$(ls *.fasta)
samtools faidx ${f}
bwa index ${f}
cd ..

Only human sequences are available in the GATK resource bundle. However, if you down-
load the sequence of a reference genome from a different database, you may need to create
a dictionary for that sequence using Picard software as follows:

java -jar ~/software/picard.jar \

CreateSequenceDictionary \
R=Reference.fasta \
O=Reference.dict

4.2.2.2.3 Mapping reads to the reference genome

The second step is to align the raw reads to the reference genome that we have downloaded
and indexed in the previous step. We can use the aligner of our choice. The most com-
monly used one is “bwa mem”. The following script creates the directory “sam” and saves
the SAM files that contain the read mapping information into that directory. There will be
a SAM file for each individual.

mkdir sam
cd fastq
ref=$(ls ../refgenome/*.fasta)
for i in $(ls *.fastq | rev | cut -c 13- | rev);
do
bwa mem -M -t 4 \
-R “@RG\tID:${i}\tSM:${i}” \
${ref} \
${i}_chr21.fastq > \
../sam/${i}_chr21.sam 2> ../sam/${i}_chr21.log;
done
cd ..
 Variant Discovery   ◾    133

4.2.2.2.4 Converting SAM files into BAM files

We must convert SAM files into BAM files to save storage space. Moreover, BAM files can
be manipulated faster. The following script creates the directory “bam” and converts the
SAM files into BAM files:

mkdir bam
cd sam
for i in $(ls *.sam | rev | cut -c 5- | rev);
do
samtools view -uS -o ../bam/${i}.bam ${i}.sam
done
cd ..

4.2.2.2.5 Sorting and indexing alignments in the BAM files

The alignments in BAM files are to be sorted by chromosomes in the reference genome to
be used in the downstream analysis. The following bash script uses samtools to sort and
index the BAM files and stores them in a new directory called “sortedbam”:

mkdir sortedbam
cd bam
for i in $(ls *.bam);
do
samtools sort -T ../sortedbam/tmp.sort -o ../sortedbam/${i} ${i}
samtools index ../sortedbam/${i}
done
cd ..

4.2.2.2.6 Extracting a chromosome or an interval

Most of the time, we may be interested in the identification of variants on the whole genome.
However, sometimes the study may focus on a specific chromosome or an interval of the
genome. In case the target is the variants of the whole genome, you can skip this step. You
should remember that identifying variants from whole genome requires large memory and
storage space. Therefore, for demonstration and the sake of simplicity, we will focus only
on chromosome 21 of human genome. In the following, we will create a directory “chr21”
and use samtools to extract the alignments of chromosome 21 and store the BAM files and
sort them:

mkdir chr21
cd sortedbam
for i in $(ls *.bam|rev|cut -c 5-|rev);
do
samtools view -b ${i}.bam chr21 > ../chr21/${i}.bam
samtools index ../chr21/${i}.bam
done
cd ..
134   ◾    Bioinformatics

4.2.2.2.7 Mark the duplicate reads

GATK4 best practice recommends removing or marking the duplicate reads mapped to the
reference genome to avoid biases in variant calling. In the above examples, we used sam-
tools to mark the duplicate reads. With GATK4 best practice, we can use Picard to mark
the duplicate. Picard is a set of command-line tools for manipulating sequencing data such
as SAM/BAM and VCF. To install Picard, follow the instructions at “https://broadinsti-
tute.github.io/picard/”.
The Picard MarkDuplicates function works by comparing reads in the 5′ positions and
collecting the duplicates. The function identifies the primary and duplicate reads. The
duplicates are marked with the hexadecimal value of 0x0400. For more details, check out
Picard manual.
The following script creates a directory called “dedup”, marks the duplicate alignments
in BAM files and indexes the resulted BAM files, and stores them in the new directory:

mkdir dedup
cd chr21
for i in $(ls *.bam|rev|cut -c 5-|rev);
do
java -Xmx7g -jar ~/software/picard.jar MarkDuplicates \
-INPUT ${i}.bam \
-OUTPUT ../dedup/${i}.dedup.bam \
-METRICS_FILE ../dedup/${i}.dedup_metrics.txt
samtools index ../dedup/${i}.dedup.bam
done
cd ..

4.2.2.2.8 Adding read grouping header to each BAM file

RGs are identified in the BAM file by a number of tags that are defined according to a specific
format. We can use the following samtools command to display the RG tags on a BAM file:

samtools view -H filename.bam | grep “^@RG”

GATK4 RGs’ tags allow us to differentiate samples and technical features that are asso-
ciated with artifacts. Such information is used to reduce the effects of artifacts during the
duplicate marking and base recalibration steps. GATK requires several RG fields to be pres-
ent in input files and will fail with errors if this requirement is not met. The RG fields are
usually set when the reads are aligned to a reference genome by the aligner. However, if the
BAM files do not include the RG fields, we can use the “AddOrReplaceReadGroups” Picard
function to add or modify the RG fields as we will do next. AddOrReplaceReadGroups uses
the following arguments to add or modify the RG fields on a BAM file:
RGID: This adds group read ID, which must be unique across the samples.
RGPU: This argument is optional for GATK, and it holds the flow cell barcode, lane ID,
and sample barcode separated by a period “.”.
RGSM: This holds the sample name that will be used later as a header of sample column
in the VCF file.
 Variant Discovery   ◾    135

RGLP: This argument sets the RG field which holds the technology name used for
sequencing. The value can be ILLUMINA, SOLID, LS454, HELICOS, or PACBIO.
RGLB: This sets the RG field which holds the DNA preparation library identifier. The
“MarkDuplicates” function of GATK uses this field to determine which RGs contain
duplicates.
The following script creates the directory “RG” and uses Picard to add the RG to each
BAM file. We use the run ID as RG and sample number.

mkdir RG
cd dedup
for i in $(ls *.bam|rev|cut -c 5-|rev);
do
java -jar ~/software/picard.jar AddOrReplaceReadGroups \
I=${i}.bam \
O=../RG/${i}.RG.bam \
RGID=${i} \
RGLB=lib RGPL=ILLUMINA \
SORT_ORDER=coordinate \
RGPU=bar1 RGSM=${i}
samtools index ../RG/${i}.RG.bam
done
cd ..

4.2.2.2.9 Building a model for the BQSR

We already know that the raw data may have systematic errors that may affect reporting
of the base calling quality score. Such error may lead to overestimate or underestimate
the reported quality score. The quality of variant calling basically depends on the qual-
ity scores of the base calling that will also affect the read alignment. To minimize the
effect of the systematic errors on variant calling, a BQSR is implemented by GATK4
best practice. The BQSR is a machine learning-based method that uses training data to
model the empirically observed errors and adjust the quality scores of the aligned reads
using that model. The adjusted scores are then used by the variant caller to take deci-
sion about a variant calling. The BQSR is achieved in two steps: (i) using a set of known
variants as a training dataset for building the recalibration table (with BaseRecalibrator
GATK4 function) and (ii) adjusting the base quality scores (with ApplyBQSR GATK4
function). The first step of recalibration process generates a table indicating which sites
of the BAM file need adjustment of quality score. The second step of the recalibra-
tion process applies recalibration or adjusting the quality scores. The BQSR generates a
new BAM file with recalibrated quality scores that variant calling process can rely on.
Moreover, the known variants are used to mark the bases at the sites of real variation
to avoid being ignored as artifacts. The model training requires high-quality variant
datasets (SNPs and InDels) in VCF files downloaded from a reliable source such as
NCBI database. Human variant VCF files can also be downloaded from GATK resource
bundle as mentioned above.
136   ◾    Bioinformatics

The following bash script downloads a standard human SNPs and InDels VCF files with
their index files in the “refvcf” subdirectory and then it performs the two steps of BQSR.
Notice that for BaseRecalibrator tool, the known variant files are provided in “--known-
sites” option. The outputs of the base recalibration are stored in “applyBQSR” subdirectory.

mkdir refvcf
cd refvcf
wget https://storage.googleapis.com/genomics-public-data/
resources/broad/hg38/v0/1000G_phase1.snps.high_confidence.hg38.
vcf.gz
wget https://storage.googleapis.com/genomics-public-data/
resources/broad/hg38/v0/1000G_phase1.snps.high_confidence.hg38.
vcf.gz.tbi
wget https://storage.googleapis.com/genomics-public-data/
resources/broad/hg38/v0/Homo_sapiens_assembly38.known_indels.vcf.
gz
wget https://storage.googleapis.com/genomics-public-data/
resources/broad/hg38/v0/Homo_sapiens_assembly38.known_indels.vcf.
gz.tbi
cd ..
##B- Build the BQSR model
#------------------------
mkdir BQSR
cd RG
ref=$(ls ../refgenome/*.fasta)
for i in $(ls *.bam|rev|cut -c 5-|rev);
do
~/software/gatk-4.2.3.0/gatk --java-options \
-Xmx4g BaseRecalibrator \
-I ${i}.bam \
-R ${ref} \
--known-sites ../refvcf/1000G_phase1.snps.high_confidence.
hg38.vcf.gz \
--known-sites ../refvcf/Homo_sapiens_assembly38.known_
indels.vcf.gz \
-O ../BQSR/${i}.table
done
cd ..
##C- Apply the model to adjust the base quality scores
#----------------------------------------------------
mkdir applyBQSR
cd RG
ref=$(ls ../refgenome/*.fasta)
for i in $(ls *.bam|rev|cut -c 5-|rev);
do
~/software/gatk-4.2.3.0/gatk \
--java-options \
 Variant Discovery   ◾    137

-Xmx4g ApplyBQSR \
-I ${i}.bam \
-R ${ref} \
--bqsr-recal-file ../BQSR/${i}.table \
-O ../applyBQSR/${i}.bqsr.bam
done
cd ..

4.2.2.2.10 Variant calling

After BQSR of BAM files, we can perform variant calling on each sample using the
“HaplotypeCaller” GATK4 function, which identifies the variation active regions, con-
structs possible haplotypes using de Bruijn-like graph, and then uses Bayesian theory to
call variants as described above. HaplotypeCaller will generate a Genome Variant Call
Format (gVCF) file for each sample. The gVCF stores sequencing information for both
variant and non-variant sites on a genome sequence. It can hold representation of geno-
type, annotation, and other information across all sites in the genome in a compact format.
Storing sample variant in gVCF format will make consolidation of variants across samples
easy.
The following script uses HaplotypeCaller to generate gVCF file for each sample. Notice
that since we are targeting only chromosome 21, we will use “-L chr21” option to restrict
variant calling to that chromosome. Also notice that chromosome label may be different
(e.g., 21, chromosome21); therefore, view the BAM file to check the right chromosome
names. The GATK4 GenotypeGVCFs tool is used to generate gVCFs.

mkdir gvcf
cd applyBQSR
ref=$(ls ../refgenome/*.fasta)
for i in $(ls *.bam|rev|cut -c 5-|rev);
do
~/software/gatk-4.2.3.0/gatk \
--java-options \
-Xmx10g HaplotypeCaller \
-I ${i}.bam \
-R ${ref} \
-L chr21 \
-ERC GVCF \
-O ../gvcf/${i}.g.vcf.gz
done
cd ..

4.2.2.2.11 Consolidating variants across samples

The above script used HaplotypeCaller to generate gVCF file for each sample. The next step
is to use GenomicsDBImport tool to import single-sample gVCFs into GenomicsDB and to
use GenotypeGVCFs tool to consolidate variants across the sample in a single VCF file. For
GenomicsDBImport, the input gVCF file is passed through “-V” option. For multiple gVCF
138   ◾    Bioinformatics

FIGURE 4.6 Cohort sample map file.

files, we can use “-V” for each one. Instead of using “-V” option several times for multiple
gVCF files, the sample information can be saved in a text file called cohort sample map file.
The file then can be passed in “--sample-name-map” option. The cohort sample map file is
a plain text file that contains two tab-separated columns; the first column is for the sample
IDs and the second column is for the names of the gVCF files. Each sample ID is mapped
to a sample file name as shown in Figure 4.6.
The cohort sample map file can be created manually by the user. However, we can also
use bash script to create it. The following script creates a cohort sample map file for our 13
samples and the file will be as shown in Figure 4.6:

cd gvcf
#a- make file name and absolute path
find “$PWD”/*_chr21.dedup.RG.bqsr.g.vcf.gz -type f -printf ‘%f
%h/%f\n’ > ../tmp.txt
#b- remove _1/2.fastq
awk ‘{ gsub(/_chr21.dedup.RG.bqsr.g.vcf.gz/,”,”, $1); print } ‘
../tmp.txt > ../tmp2.txt
rm ../tmp.txt
#remove space
cat ../tmp2.txt | sed -r ‘s/\s+//g’ > ../tmp3.txt
rm ../tmp2.txt
#replace comma with tab
sed -e ‘s/\,\+/\t/g’ ../tmp3.txt > ../cohort.sample_map
rm ../tmp3.txt

Once we have created the cohort sample map file, we can run GenomicsDBImport tool
to import gVCF sample files and GenotypeGVCFs tool to consolidate the variants of 13
samples in a single VCF file.

#create a database
ref=$(ls ../refgenome/*.fasta)
~/software/gatk-4.2.3.0/gatk \
--java-options -Xmx10g \
GenomicsDBImport \
 Variant Discovery   ◾    139

-R ${ref} \
--genomicsdb-workspace-path ../gvcf21db \
--batch-size 50 \
--sample-name-map ../cohort.sample_map \
-L chr21 \
--tmp-dir ../tmp \
--reader-threads 4
cd ..
mkdir vcf
ref=$(ls refgenome/*.fasta)
~/software/gatk-4.2.3.0/gatk \
--java-options -Xmx10g \
GenotypeGVCFs \
-R ${ref} \
-V gendb://gvcf21db \
-O vcf/allsamples_chr21.vcf

4.2.2.2.12 Variant separation

The output of the previous step is a single VCF file containing both the identified SNPs and
InDels. It is good practice to separate each kind of variants in a VCF file because each type
of variants may be used for certain analysis. The “SelectVariants” GATK tool is used to
select a subset of variants based on a specific criterion. We can use “-select-type” option to
select a specific variant type. The valid types are INDEL, SNP, MIXED, MNP, SYMBOLIC,
and NO_VARIATION. The following script stores SNPs and InDels in separate CVF files:

#SNPS
~/software/gatk-4.2.3.0/gatk \
--java-options \
-Xmx10g SelectVariants \
-V vcf/allsamples_chr21.vcf \
-select-type SNP \
-O vcf/allsamplesSNP_chr21.vcf
#INDEL
~/software/gatk-4.2.3.0/gatk \
--java-options \
-Xmx4g SelectVariants \
-V vcf/allsamples_chr21.vcf \
-select-type INDEL \
-O vcf/allsamplesIndels_chr21.vcf

4.2.2.2.13 Variant filtering

Rather than the traditional variant filtering discussed with the previous variant callers,
GATK4 uses an advanced filtering approach called Variant Quality Score Recalibration
(VQSR) to filter variants called in the previous step. This approach is similar to the BQSR dis-
cussed above as it uses machine learning to model a variant profile in a high-quality training
140   ◾    Bioinformatics

dataset from validated data resources, such as 1000 Genomes, OMNI, and ­hapmap, and
then it uses the model to filter out the putative artifacts from the called variants. The
­application of the model results in assigning a log-odds ratio score (VQSLOD) for each
variant that measures how likely that variant is real based on the data used in the training.
The VQSLOD is added to the INFO field of the variant. The variants are then filtered based
on a threshold. SNPs and InDels are recalibrated separately. The variant calibration and
filtering are performed in two steps:
(i) Building of the recalibration model:
The recalibration model is built using VariantRecalibrator tool. The input file for this
tool is the variants to be recalibrated “-V” and the known training dataset “--resource”.
The latter must be downloaded from a reliable source such as GATK resource bundle. The
fitted model is used to estimate the relationship between the probability that whether a
variant is true or artifact and continuous covariates that include QD (quality depth), MQ
(Mapping quality), and FS (FisherStrand). The VQSLOD is estimated based on Gaussian
mixture model whether a variant is true versus being false. Each variant in the input VCF
file is assigned a VQSLOD in INFO field of the VCF file and the variants are ranked by
VQSLOD. A tranche sensitivity threshold can be provided in “-tranche” as a percentage.
Several thresholds can be set. The output of this step is a recalibrated VCF file and other
files including tranches, which will be used by ApplyVQSR, and plot files.

cd refvcf
wget https://storage.googleapis.com/genomics-public-data/
resources/broad/hg38/v0/Homo_sapiens_assembly38.dbsnp138.vcf
wget https://storage.googleapis.com/genomics-public-data/
resources/broad/hg38/v0/Homo_sapiens_assembly38.dbsnp138.vcf.idx
cd ..
mkdir VQSR
cd vcf
~/software/gatk-4.2.3.0/gatk --java-options \
-Xmx10g VariantRecalibrator \
-R ../refgenome/Homo_sapiens_assembly38.fasta \
-V allsamplesSNP_chr21.vcf \
--trust-all-polymorphic \
-tranche 100.0 \
-tranche 99.95 \
-tranche 99.90 \
-tranche 99.85 \
-tranche 99.80 \
-tranche 99.00 \
-tranche 98.00 \
-tranche 97.00 \
-tranche 90.00 \
--max-gaussians 6 \
--resource:1000G,known=false,training=true,truth=true,prior=10.0
\
 Variant Discovery   ◾    141

../refvcf/1000G_phase1.snps.high_confidence.hg38.vcf.gz \
--resource:dbsnp,known=false,training=true,truth=false,prior=2.0
\
../refvcf/Homo_sapiens_assembly38.dbsnp138.vcf \
-an QD -an MQ -an MQRankSum \
-an ReadPosRankSum -an FS -an SOR \
-mode SNP \
-O ../VQSR/tranches.out \
--tranches-file ../VQSR/vqsrplots.R
cd ..

(ii) Applying recalibration rules:

Next is applying the model on the variants using ApplyVQSR by tranche sensitivity
thresholds to filter variants and adding PASS to FILTER field of the variants that have
VQSLOD above the threshold while the variants with VQSLOD below the threshold are
labeled with tranche name.

#Apply VQSR
mkdir filteredVCF
~/software/gatk-4.2.3.0/gatk \
--java-options -Xmx4G ApplyVQSR \
-V vcf/allsamplesSNP_chr21.vcf \
-O filteredVCF/humanSNP.vcf \
--truth-sensitivity-filter-level 99.7 \
--tranches-file VQSR/vqsrplots.R \
--recal-file VQSR/tranches.out \
-mode SNP \
--create-output-variant-index true

We can follow the same steps for filtering InDels.

The downside of the VQSR approach is that it is not applicable for all organisms, because
high-quality variant datasets are available only for human and some model organisms. As
an alternative to the VQSR, we can use hard filtering with VariantFiltration tool, which
enforces a hard filter for one or more annotations such as QD (quality depth), QUAL
(Phred quality score), SOR (StrandOddsRatio), FS (FisherStrand), MQ (Mapping quality),
and MQRankSum. For these filters, use the thresholds recommended by GATK4. PASS
will be added to FILTER field on the VCF file if a variant passes the filtering; otherwise, the
name of the filter will be added instead. GATK4 recommends some filters for both SNPs
and InDels. The following scripts are for hard filtering SNPs and InDels.

#The hard filtering GATK best practice for SNP

~/software/gatk-4.2.3.0/gatk \
--java-options \
-Xmx10G VariantFiltration \
-V vcf/allsamplesSNP_chr21.vcf \
-filter “QD<2.0” \
142   ◾    Bioinformatics

FIGURE 4.7 Visualizing InDels and SNPs with IGV.

--filter-name “QD2” \
-filter “QUAL<30.0” \
--filter-name “QUAL30” \
-filter “SOR>3.0” \
--filter-name “SOR3” \
-filter “FS>60.0” \
--filter-name “FS60” \
-filter “MQ<40.0” \
--filter-name “MQ40” \
-filter “MQRankSum<-12.5” \
--filter-name “MQRankSum-12.5” \
-filter “ReadPostRankSum<-8.0” \
--filter-name “ReadPostRankSum-8” \
-O filteredVCF/hardfilteredVCF.vcf

#The hard filtering GATK best prectice for INDEL

~/software/gatk-4.2.3.0/gatk \
--java-options \
-Xmx10G VariantFiltration \
-V vcf/allsamplesIndels_chr21.vcf \
-filter “QD<2.0” \
-- filter-name “QD2” \
-filter “QUAL<30.0” \
--filter-name “QUAL30” \
-filter “FS>200.0” \
--filter-name “FS200” \
-filter “ReadPostRankSum<-20.0” \
 Variant Discovery   ◾    143

--filter-name “ReadPostRankSum-20” \
-O filteredVCF/hardfilteredIndels.vcf

4.3 VISUALIZING VARIANTS

The variants in a VCF file can be visualized on a variant viewer such as IGV (Integrated
Genomics Viewer), which is an open-source program for all platforms. It can be down-
loaded from “https://software.broadinstitute.org/software/igv/download” and installed on
a local computer. Figure 4.7 shows the allele fractions and genotypes for each of the InDels
and SNPs of the samples. The dark blue color indicates heterozygous genotype, cyan indi-
cates homozygous genotype, and gray indicates the same genotype as the reference. Refer
to the documentation of the IGV to read more about this.

4.4 VARIANT ANNOTATION AND PRIORITIZATION

The variant calling using any of the variant callers, such as bcftools, FreeBayes, or GATK,
and variant filtering is followed by variant annotation and prioritization. Variant annota-
tion involves adding information and knowledge to high-confidence variants in an effort
to enhance assessment of variants that are likely to impact functions. Following the work-
flow of variant calling, we will obtain high-quality variants in a single VCF file including
the genotypes of all samples. Since variant discovery usually involves the whole genome
or whole exome of an individual or multiple individuals of a species, thousands of vari-
ants may be discovered. However, we are usually interested in the variants that affect the
function or have associations with diseases or other important phenotypes. Variants may
impact functions in different ways depending on the type of the variants. Variants can be
everywhere on the genome sequence, but the most deleterious and damaging are the ones
that have effect on the function of a gene. A variant may suppress, inhibit, or activate a
gene if it affects the gene regulatory region. This kind of effects are usually seen in cancer
cell in which mutations may lead to the hyperactivity of proto-oncogenes, which accelerate
cell growth and division or inactivation of tumor suppressor genes. Variants which affect
the coding regions of a gene may cause an impact depending on the consequence of the
change. A single-nucleotide variant (SNV) that forms a stop codon will cause a truncated
protein that does not function normally. On the other hand, SNV in a stop codon may lead
to a stop loss that results in a longer protein. A variant in a splicing region may also alter
the sequence and function of the protein. More often, SNVs affect the coding regions of a
gene producing new amino acids that change the characteristic and function of the trans-
lated protein. These SNVs are known as missense SNVs and they are the easiest predictable
variants. But an SNV can also be synonymous in the sense that it changes the codon but it
does not change the amino acid. Although this kind of variant does not change the protein
sequence, it may still have a biological consequence. Insertion or deletion of a single or
multiple nucleotides in the coding region may lead to the frameshift, and hence, the pro-
tein will be translated incorrectly from that point.
The most deleterious variants are stop-gain, frameshift, and splicing region variants
since they lead to loss of function. However, before we decide on a variant effect, we should
144   ◾    Bioinformatics

FIGURE 4.8 Variant effect on gene regions.

TABLE 4.3 Gene Regions and Variant Effect

Region Variant Effect
(1) Regulatory region including transcription Deleterious variants
factor (TF) binding site
(2) Upstream gene region Intergenic variants/upstream gene variant
(3) 5′ UTR region 5′ UTR variant
(4) Transcription start site (TSS) Start retained or start lost variants
(5) Exon region Exonic variants include missense, nonsense (stop gained),
frameshift, inframe insertion or deletion
(6) Splice donor region Splice-site variant (exon loss, intron inclusion, altered
protein-coding sequence)
(7) Splice acceptor region Splice-site variant (exon loss, intron inclusion, altered
protein-coding sequence)
(8) Intron region Intronic variant
(9) Transcription termination site (TTS) Stop lost, stop retained, incomplete terminal codon
(10) 3′ UTR region 3′ UTR variant
(11) Downstream gene region Downstream gene variant

remember that the consequence may also be beneficial in some cases. For instance, it has
been reported that truncating variants in CARD9, IL23R, and RNF186 proteins may protect
against Crohn’s disease and ulcerative colitis and also truncating variants in ANGPTL4,
APOC3, PCSK9, and LPA proteins may protect against coronary heart disease [9, 10]. The
impact of variants on a protein is also measured by the number of isoforms produced by
the affected gene, the percentage of the protein affected, and moreover, we should put into
consideration that a frameshift may be bypassed by splicing or its impact may be avoided
by another frameshift. In the attempt to annotate genetic variants, the SNVs effect can be
predicted with high accuracy followed by small InDels (1–50 bp) and then medium InDels
(50–100 bp). It is also easy to predict the effect of missense SNV. The variant annotation
tools have several approaches to predict the effect of missense variant. For instance, they
can take into consideration the physicochemical properties of amino acids, whether the
variant is in a conserved region or not, or does it affect the three-dimensional structure of
the protein. A variant in a region conserved across the species or in a region of a secondary
or tertiary structure is more likely to be deleterious. Some tools use homology modeling
to simulate the structure of the new protein to predict the effect of variants and other tools
use machine learning utilizing multiple features to annotate the variants with the right
information and consequences. Figure 4.8 and Table 4.3 show a segment of a eukaryotic
genomic gene and possible variant annotations in each region.
In a typical genome-wide variant study, thousands of variants may be discovered. The
significance of these variants varies based on the type, location, and possible consequence.
 Variant Discovery   ◾    145

Covering all variants is daunting and hence prioritizing these variants with potential
­associations to phenotypes of interest is usually the key target of the variant annotation.
During the last decade, numerous GWASs were conducted to identify genomic variants
associated with many complex diseases and traits. However, most studies were focused
on human and some model organisms. Databases for variant mapping and genotype–
phenotype association were developed to serve as rich resources for variant annotation.
Examples of these databases include NHLBI, which contains health information col-
lected by NHLBI’s epidemiological cohorts and clinical trials, dbGaP, which is an NCBI
database of Genotypes and Phenotypes association, the Exome Aggregation Consortium
(ExAC), which includes sequencing data from a variety of large-scale sequencing projects,
Catalogue of Somatic Mutations in Cancer (COSMIC), etc. Numerous similar databases
were developed for specific diseases such as cancer, autoimmune diseases, and Alzheimer’s
disease. Prioritizing genetic variants relevant to the human diseases is the top. Guidelines
have been developed for investigating variants and their association with human diseases
so that such knowledge can be used for diagnosis in a clinical setting. Indeed, after acquir-
ing high-confidence variants, the next step is to annotate and interpret these variants using
either prior knowledge or functional prediction based on the impact of the variant on the
translated protein. The studies on genetic variants are usually interested in the variants
that are associated with diseases, traits, or have an effect on functions of protein. There are
a variety of consequences that can be caused by variants. A variant may be pathogenic or
implicated with healthy conditions, or may be a damaging variant that alters the normal
function of a gene, or may be deleterious variant that reduces the quality of the affected
individuals. Hence, variant annotation must be conducted after filtering variants as dis-
cussed above to avoid misinterpretation, false positive, and false negative. Generally, we
can define variant annotation as the process of assigning functional or phenotype infor-
mation to genetic variants such as SNPs, InDels, or copy number variants. Based on this
definition, perhaps, the most significant variants are the ones on the coding region of the
genome. This is because mutations on coding region may have a direct impact on the pro-
tein and may be implicated in a disease. The variants on non-coding region of the genome
may also have impact but the challenge is that it is difficult to establish a testable hypoth-
esis. Therefore, statistical methods were developed for variant prioritization by incorpo-
rating diverse functional evidence, so that variants with small effect sizes but possessing
functional features may be prioritized over variants with similar effect sizes but less likely
to be functional.
There are numerous variant annotation tools that attempt to associate variants to knowl-
edge-based information and generate reports. The most commonly used tools include SIFT,
SnpEff, Annovar, and VEP, which we will use to annotate the variants.

4.4.1 SIFT
The SIFT [11], which stands for Sorting Intolerant from Tolerant, was first introduced
in 2001 as an online variant annotation tool that annotates coding region of genes with
the missense variant effects on the translated protein. SIFT relies on the assumption that
substitutions in conserved regions are more likely to be deleterious if the missense SNV
146   ◾    Bioinformatics

resulted in a nonsynonymous codon that is translated into an amino acid with d ­ ifferent
physicochemical properties. For instance, if a hydrophobic amino acid is replaced by
another hydrophobic amino acid, SIFT will predict that change is tolerated; however, if it is
substituted with a polar amino acid, the variant will be predicted as deleterious. SIFT algo-
rithm avails of the NCBI PSI-BLAST as it uses the translated protein as a query sequence
against a database of protein sequences. The search hit sequences are aligned using mul-
tiple sequence alignment (MSA) and the probabilities of all possible substitutions at each
position are computed forming position-specific scoring matrix (PSSM), where each entry
in the matrix represents the probability of observing an amino acid in that column of the
alignment. The probabilities are normalized based on the consensus amino acids. Then,
position with normalized probability ranges between 0 and 1. SIFT predicts that a SNV
with a probability between 0.0 and 0.05 on that position is deleterious and will affect the
function of the protein and a probability greater than 0.05 (>0.05) can be tolerated. SIFT
also measures conservation of the sequence using the median sequence conservation,
which ranges from 0 to log2(20) or from 0 to 4.32, where median sequence conservation
of 4.32 indicates that all sequences in the alignment are identical to each other, and hence,
any variant in this region will be predicted as damaging. SIFT also reports the number of
sequences at the variant position. The latest version of SIFT is SIFT 4G (SIFT for genomes),
which is faster and enables practical computations on reference genomes using precom-
puted databases and also it provides SIFT prediction for more organisms. Hundreds of
databases for different organisms are available.
Use the following steps to annotate variants using SIFT 4G on Linux terminal:
First, create a directory with the name of your choice or “sift4g” and change into it.

mkdir sift4g
cd sift4g

Open “https://sift.bii.a-star.edu.sg/sift4g/public/”. You will see databases of tens of organ-

isms. Scroll down to the database of your interest, open its folder, and download the appro-
priate database build into your working directory. Since we have variants called above
from human samples, we can download the latest human build GRCh38.78 by copying the
link and using “wget” command and then unzip it using “unzip” or follow the instructions.

wget https://sift.bii.a-star.edu.sg/sift4g/public/Homo_sapiens/
GRCh38.78.zip
unzip GRCh38.78.zip

Each chromosome will have three files: a compressed file with “gz” file extension, a region
file with “.region” file extension, and a chromosome statistics file with “.txt” file extension.
Download SIFT 4G Annotator Java executable file (.jar) in a directory or in your work-
ing directory:

wget https://github.com/paulineng/SIFT4G_Annotator/raw/master/
SIFT4G_Annotator.jar
 Variant Discovery   ◾    147

Assuming that your VCF file is in the current directory, you can run the following on the
command line:

java -jar SIFT4G_Annotator.jar \

-c -i humanSNP.vcf \
-d GRCh38.78 \
-r output -t

The option “-jar” for the SIFT 4G path, “-c” is essential for the command line, “-i” for the
input path, “-d” for the database path, “-r” specifies the output folder, and “-t” is to extract
annotations for multiple transcripts.
The above command generates two files: an Excel annotation file with “.xls” file exten-
sion and a VCF file.
Instead of the command line, you can also use SIFT 4G with a graphic user interface
(GUI) by running the following command:

java -jar SIFT4G_Annotator.jar

This will open SIFT interface where you can browse to select the VCF file and the database
and click Start to annotate the variants. The links of the two output files are given at the
bottom of the GUI (Figure 4.9).
We can open the Excel file to check its content (Figure 4.10). We can notice that some
annotations have been added to the variants, such as Ensemble transcript ID, gene ID,
gene name, region, variant type (synonymous or nonsynonymous), SIFT score, median

FIGURE 4.9 SIFT 4G annotator.

148   ◾    Bioinformatics

FIGURE 4.10 SIFT 4G annotation file.

conservation score, number of sequences, dbSNP accession, and SIFT prediction whether
the variant is tolerated or deleterious.

4.4.2 SnpEff
SnpEff [12] is another variant annotation tool that categorizes the coding effects of variants
based on their genomic locations such as introns, untranslated region (UTR), upstream,
downstream, and splicing site. SnpEff predicts a variety of variant effects including syn-
onymous or nonsynonymous substitution, start-gain codon, start-loss codon, stop-gain
codon, stop-loss codon, or frameshifts.
In general, SnpEff consists of two main components: (i) database builds and (ii) vari-
ant effect calculation. The SnpEff database builds are usually distributed with SnpEff and
there are around hundreds of databases available. A database build is a gzip-compressed
serialized object that is formed of the genome FASTA sequence and an annotation file (in
GTF or GFF format). These database files can be acquired from database resources such as
ENSEMBL and UCSC. The variant effect calculation is performed after building the data-
base. It begins with building a data structure which is a hash table interval trees indexed
by chromosome. The data structure indexes intervals and makes their search efficient. The
SnpEff program uses the VCF file as input and finds the intersections with the annotated
database. The intersecting genomic regions are then identified and the variant effect is
calculated from exonic region only. Simply, SnpEff will take information from the pro-
vided annotation database and populate the input VCF file by adding annotation into the
INFO field name, ANN. Data fields are encoded separated by pipe sign “|”; and the order
of fields is written in the VCF header. As examples, variants may be categorized by SnpEff
as SNP (single-nucleotide polymorphism), Ins (insertion), Del (deletion), MNP (multiple-
nucleotide polymorphism), or MIXED (multiple-nucleotide and InDel). The impacts of
variants are classified into high, moderate, low, or modifier based on the affected region. A
variant will have a high impact when it is disruptive and likely to cause protein truncation,
loss of function, or triggering nonsense mediated decay. The variants with high impact
are frameshift and stop-gain variants. The non-disruptive variants such as missense SNV
and inframe deletion that might change protein effectiveness only are moderate impact
 Variant Discovery   ◾    149

variants. The low-impact variants are the synonymous ones that do not change protein
behavior; however, they may still have some effects. Variants on introns (intron-variants)
and variants in the downstream region of a gene are annotated as MODIFIER since they
affect non-coding regions, where prediction of the impact is difficult or there is no substan-
tial evidence of any effect.
To use snpEff, we should install snpEff software and download the database of interest.
Installation of SnpEff software requires Java V1.8 or later installed on your computer. The
installation instructions are available at “https://pcingola.github.io/SnpEff/download/” or
“http://pcingola.github.io/SnpEff/”. First, download the compressed folder using “wget”
Linux command and then decompress it with “unzip” command. You can also install the
binary.

wget https://snpeff.blob.core.windows.net/versions/snpEff_latest_
core.zip
unzip snpEff_latest_core.zip

This will create the directory “snpEff”, which contains two Java executable files with “.jar”
extension (snpEff.jar and snpSift.jar), a snpEff configuration file (snpEff.config), a license
file, and four subdirectories (“examples”, “exec”, “galaxy”, and “script”). By default, the
database will be stored in a subdirectory “data” in the “snpEff” directory but we can change
it by opening “snpEff.config” and editing “data.dir = ./data/” to the path that we need.
The database can be downloaded automatically and this is recommended but you can
also install it manually. For example, if you need to download the human database manu-
ally, you can run the following:

java -jar snpEff.jar download GRCh38.

Figure 4.11 shows the directory structure of the snpEff. If you installed the binary files, you
can use any of the executables without the “java” command as follows:

snpEff download GRCh38.

When you use snpEff, you must provide the right file path. For instance, if you are just one
step out of the “snpEff” directory, then you can run:

FIGURE 4.11 snpEff directory structure.

150   ◾    Bioinformatics

java -jar snpEff/snpEff.jar download GRCh38.99

To list all available SnpEff database, run the following command:

java -jar snpEff/snpEff.jar databases

The “snpeff” is the current directory that we created to store the snpEff software, the VCF
file, and the output. We will copy our VCF file to the root directory “snpeff”, while the
snpEff executable file and database are in “snpEff”. After copying our VCF file “humanSNP.
vcf” into the working directory, you can annotate it using the following command:

java -Xmx8g -jar snpEff/snpEff.jar GRCh38.99 humanSNP.vcf >

mySNPanot.vcf

This command will produce three files: a VCF file (mySNPanot.vcf), gene file (snpEff_
genes.txt), and summary file in html format (snpEff_summary.html). SnpEff adds func-
tional annotations in the ANN keyword in the INFO field of the VCF output file. Figure
4.12 shows the VCF output file, which is modified to show ANN under INFO field. The
INFO field may include the effect of the variant (stop loss, stop gain, etc.), effect impact
on gene (High, Moderate, Low, or Modifier), or functional class of the variant (nonsense,
missense, frameshift, etc.).
Moreover, we can view the summary on the html file to have a general idea about the
type and regions and effects of the variants. If you have “firefox” installed, you can display
the summary on the html file using the “firefox” command or you can open it with an
Internet browser.

firefox snpEff_summary.html

Figure 4.13 shows the summary of the annotation using SnpEff and variant rate details.
Remember that the VCF file contains the variants of the human chromosome 21 only.
Figure 4.14 shows the number of variant effects by impact and by functional class. Only
68 SNVs (0.009%) have high impact. The remaining variants are SNV with moderate
impact (0.149%), SNV with low impact (0.149), and modifier (99.575%).

FIGURE 4.12 A VCF annotated with SnpEff.

 Variant Discovery   ◾    151

FIGURE 4.13 SnpEff annotation summary.

FIGURE 4.14 Variant effect summary.

Figure 4.15 shows the number of effects by type and region. Figure 4.16 shows a bar
chart showing the percentage of effects by region.

4.3.3 ANNOVAR
ANNOVAR [13] is one of the most commonly used annotation tools that is used to anno-
tate SNVs and InDels with different types of annotations including functional conse-
quence on genes, inferring cytogenetic bands, functional importance, impacts of variants
in conserved regions, and identifying variants reported in the NCBI dbSNP and the 1000
152   ◾    Bioinformatics

FIGURE 4.15 Number of variant effects by type and region.

FIGURE 4.16 A bar chart shows percentage of variants by region.

Genomes Project. In addition to variant annotation with respect to genes, ANNOVAR

has the ability to perform annotation based on genomic region and to compare variants
to existing variation databases. In general, the types of annotations with ANNOVAR can
be grouped into the following: (i) gene-based annotation which identifies the effects of
variants on the proteins, (ii) region-based annotation identifies the affected region (e.g.,
conserved region), and (iii) filter-based annotation identifies variants based on a specific
database such as dbSNP, ExAC, 1000 Genome Project, and gnomAD. The filter-based
annotation may also generate scores including SIFT, PolyPhen, LRT, MutationTaster,
MutationAssessor, FATHMM, MetaSVM, and MetaLR.
ANNOVAR uses annotation databases to perform the above types of annotation.
The annotation databases are built with the organism annotation file in GFF3 format.
 Variant Discovery   ◾    153

TABLE 4.4 ANOVAR Script Files

ANNOVAR Program Description
annotate_variation.pl The core ANNOVAR program for annotation and database download
coding_change.pl To calculate the mutated sequence and make inference
convert2annovar.pl To convert genotype-calling file format into ANNOVAR input format
retrieve_seq_from_fasta.pl To retrieve genomic nucleotide, cDNA sequences, or translated amino acid
sequences from FASTA file
table_annovar.pl To generate a tab-delimited output file with annotation columns
variants_reduction.pl For prioritizing causal variants

Ready-built human databases are available at ANNOVAR server, UCSC Genome Browser
website, or third parties and can be downloaded using “annotate_variation.pl” program.
For the non-human species which have no available annotation databases, a database can
be built from the FASTA sequence of the reference genome and the GFF/GTF annota-
tion file of that species. Those two files can be downloaded from databases such as NCBI
Genome database or UCSC database.
As shown in Table 4.4, ANNOVAR consists of six Perl files that can be used as com-
mand-line programs on any computer with Perl installed. The download instructions are
available at “https://annovar.openbioinformatics.org/en/latest/user-guide/download/”.
You may be asked to register with your school email. The download link will be emailed to
you, and then you can download the compressed file onto your computer and decompress
it with “tar xvf” command. If you are using Linux, you can add ANNOVAR to the path by
adding the following line to the end of “.bashrc” file:
Export PATH=”YOURPATH/annovar:$PATH”

4.3.3.1 Annotation Databases

For variant annotation, ANNOVAR uses annotation databases of an organism to be down-
loaded in a directory. Databases can be downloaded from UCSC Genome Browser, 1000
genome project or ANNOVAR website, or from a third-party URL. You can use “anno-
tate_variation.pl” to annotate, download a database, or list the available databases for a
specific build. The general syntax is as follows:

annotate_variation.pl \
[arguments] \
<query-file|table-name> \
<database-location>

For the complete list of argument run:

annotate_variation.pl -h

To list the available annotation databases for the hg19 build of the human reference genome,
you can run the following command:
154   ◾    Bioinformatics

annotate_variation.pl -webfrom annovar -downdb avdblist -buildver

hg19 dblist

The list of the databases of the hg19 build will be saved in a file “hg19_avdblist.txt” in
“dblist” subdirectory. You can open that file and study the available annotation databases.
The “-webfrom” argument specifies the source of database; the values can be “ucsc”,
“annovar”, or a URL where the database is available.
The “-downdb” argument is the command to download an annotation database from
the source specified by “-webfrom”.
The “-buildver” argument specifies the genome build version (e.g., hg19).
For example, the following command downloads some annotation databases for hg19
build and they will be stored in “humandb” directory:
To download the gene-based annotation database for human variants:

annotate_variation.pl \
-buildver hg19 \
-downdb \
-webfrom annovar refGene humandb/

To download region-based annotation (cytogenetic bands) database of the human genome:

annotate_variation.pl \
-buildver hg19 \
-downdb cytoBand humandb/

To download filter-based ExAC (Exome Aggregation Consortium) annotation database:

annotate_variation.pl \
-buildver hg19 \
-downdb \
-webfrom annovar exac03 humandb/

To download the filter-based NCBI dbSNP annotation database (version 147):

annotate_variation.pl \
-buildver hg19 \
-downdb \
-webfrom annovar avsnp147 humandb/

To download the filter-based dbNSFP annotation database, which was developed for func-
tional prediction and annotation of all potential nonsynonymous SNVs in the human
genome:

annotate_variation.pl \
-buildver hg19 \
 Variant Discovery   ◾    155

-downdb \
-webfrom annovar dbnsfp30a humandb/

The database files are downloaded into the specified directory “humandb”. Save the anno-
tation databases of each organism in a separate file.
Not all non-human organisms have annotation databases. In this case, you can build an
annotation database for any organism by yourself. The following steps show how to build a
gene-based annotation database. As an example, we will build an annotation database for
SARS-CoV-2 and we will use it later to annotate the variants called in a previous example.
The following are the steps to build SARS-CoV-2 gene-based annotation database:
1. Download the reference genome sequence of the organism in FASTA format and the
sequence annotation file in GFF/GTF format. For SARS-CoV-2, we can download both
files from the NCBI Genome database at
https://www.ncbi.nlm.nih.gov/genome/86693?genome_assembly_id=757732
Use the following commands to create a directory “sarscov2db” and download the ref-
erence FASTA file and GFF file into it:

mkdir sarscov2db
cd sarscov2db
wget https://ftp.ncbi.nlm.nih.gov/genomes/all/GCF/009/858/895/
GCF_009858895.2_ASM985889v3/GCF_009858895.2_ASM985889v3_genomic.
fna.gz
wget https://ftp.ncbi.nlm.nih.gov/genomes/all/GCF/009/858/895/
GCF_009858895.2_ASM985889v3/GCF_009858895.2_ASM985889v3_genomic.
gff.gz

Then, decompress the two files with “gunzip” command:

gunzip GCF_009858895.2_ASM985889v3_genomic.fna.gz
gunzip GCF_009858895.2_ASM985889v3_genomic.gff.gz

2. Use the “gff3ToGenePred” tool to convert the GFF file to GenePred file, which is a file
format used to specify the gene track annotations for an imported genome. For GFT for-
mat, use “gtfToGenePred” to convert it into GenePred file. Both “gff3ToGenePred” and
“gtfToGenePred” are ones of the UCSC Genome Browser application binaries built for
standalone command-line use on Linux and UNIX platforms. They can be downloaded by
choosing the right platform at “http://hgdownload.soe.ucsc.edu/admin/exe/”. For the sake
of simplicity, we can download “gff3ToGenePred” in the same “sarscov2db” directory and
use “chmod” to allow it to run as a program:

wget http://hgdownload.soe.ucsc.edu/admin/exe/linux.x86_64/
gff3ToGenePred
chmod +x gff3ToGenePred

If you wish to download all UCSC Genome Browser binaries, run the following:
156   ◾    Bioinformatics

mkdir ucsc
cd ucsc
rsync -aP rsync://hgdownload.soe.ucsc.edu/genome/admin/exe/linux.
x86_64/ ./

This will download all binaries in “ucsc” directory.

After downloading the right tool, you can use it to convert the annotation GFF file into
GenePred file:

gff3ToGenePred \
GCF_009858895.2_ASM985889v3_genomic.gff \
SARSCOV2_refGene.txt

This will create “SARSCOV2_refGene.txt”, which is a GenePred file.

3. Use ANNOVAR “retrieve_seq_from_fasta.pl” script to generate a transcript FASTA file
from the reference sequence.

retrieve_seq_from_fasta.pl \
--format refGene \
--seqfile GCF_009858895.2_ASM985889v3_genomic.fna \
SARSCOV2_refGene.txt \
--out SARSCOV2_refGeneMrna.fa

This will create a transcript FASTA file “SARSCOV2_refGeneMrna.fa”.

Thus, the gene-based database for SARS-CoV-2 variants is ready to use. We will use it
in a later example.

4.3.3.2 ANNOVAR Input Files

The “annotate_variation.pl” script is the core ANNOVAR program for variant annotation.
The raw variants (SNVs or InDels) must be in an ANNOVAR input file for annotate_varia-
tion.pl. The ANNOVAR input file is a plain text file that contains space- or tab-delimited
five columns for chromosome, start position, end position, the reference nucleotides, and
the observed nucleotides. Additional columns can be added. You can open the example
ANNOVAR input file “ex1.avinput” in the “example” directory to have an idea about how
it looks. Use “less -S ex1.avinput” to display the file.
Since variants come in different variant calling file formats, “convert2annovar.pl” script
can be used to convert those files to the ANNOVAR input file format. The variant calling
files that can be converted by that program include VCF format, samtools genotype-calling
pileup format, Illumina export format from GenomeStudio, SOLiD GFF genotype-calling
format, and complete genomics variant format. To learn about the use and options of “con-
vert2annovar.pl” script, run the following:

convert2annovar.pl -h
 Variant Discovery   ◾    157

FIGURE 4.17 The directory tree of the ANNOVAR.

FIGURE 4.18 ANNOVAR input file.

The VCF file format is the standard format for variant calling. The VCF file can be con-
verted into ANNOVAR input file by using “-format vcf4” argument.
Figure 4.17 shows the directory tree which includes “annovar” directory that contains
the ANNOVAR scripts and subdirectories and the “input” directory that includes the VCF
files (sarscov2.vcf and humanSNP.vcf) from the previous SARS-CoV-2 and human variant
calling examples. We copied them to this directory for simplicity. The following command
will convert “humanSNP.vcf” file into ANNOVAR input format “humanSNP.avinput”:

convert2annovar.pl \
-format vcf4 input/humanSNP.vcf \
> input/humanSNP.avinput

Figure 4.18 shows the ANNOVAR input file, which includes the first five essential columns
and additional three columns.
For converting other variant calling file formats, run “convert2annovar.pl -h”. This
command is also used with “-dbSNP” option to add the dbSNP accessions.
Variant annotation with ANNOVAR:
The “annotate_variation.pl” script is the core program for ANNOVAR annotation. It
requires ANNOVAR input file. However, “table_annovar.pl” script is also used for annota-
tion and it takes a VCF file as input.

./annotate_variation.pl \
-out ../output/humanSNPannot \
158   ◾    Bioinformatics

-build hg19 \
../input/humanSNP.avinput \
humandb/

The above command generates three files with the “humanSNPannot” prefix as shown in
Figure 4.19. The “humanSNPannot.variant_function” file contains annotation for all vari-
ants, by adding two columns to the beginning of each input line (Figure 4.20).
The first column is annotated with the affected part of the gene. The parts can be exonic,
splicing, nrRNA, UTR5, UTR3, intronic, upstream, downstream, or intergenic region. The
second added column annotates the gene name or names.
The second annotation output file, “humanSNPannot.exonic_variant_function”, con-
tains the amino acid changes as a result of the exonic variant (Figure 4.21).

FIGURE 4.19 Variant annotation files.

FIGURE 4.20 The ANNOVAR variant annotation file.

FIGURE 4.21 Exonic variant function file.

 Variant Discovery   ◾    159

The first column includes the line number in the original input file. The second column
shows the functional consequences of the variant. The possible consequences are nonsyn-
onymous SNV, synonymous SNV, frameshift insertion, frameshift deletion, nonframeshift
insertion, nonframeshift deletion, frameshift block substitution, or nonframeshift block
substitution. The third column includes the gene name, the transcript identifier, and the
sequence change in the corresponding transcript.
The “annotate_variation.pl” tool has numerous arguments. Use “annotate_variation.pl
-h” to display the complete list of arguments.
ANNOVAR provides “table_annovar.pl” script as an easy way to annotate variants in
a VCF file as an input. No need to convert VCF file into ANNOVAR input file. It takes a
VCF file as an input and generates a tab-delimited output file with many columns, each
represents one set of annotations. It also generates a new output VCF file with the INFO
field filled with annotation information.

./table_annovar.pl ../input/humanSNP.vcf humandb/ \

-buildver hg19 \
-out ../output/humanSNP2 \
-remove \
-protocol refGene,cytoBand,exac03,avsnp147,dbnsfp30a \
-operation g,r,f,f,f \
-nastring . \
-vcfinpu

The “-remove” option removes all temporary files. The “-protocol” option is comma-delim-
ited string that specifies an annotation protocol. These strings typically represent data-
base names in ANNOVAR. The “-operation” option tells ANNOVAR which operations
to use for each of the protocols, where “g” means gene-based, “gx” means gene-based
with cross-reference annotation (from -xref argument), “r” means region-based, and
“f ” means filter-based. The above ANNOVAR command generated three output files:
“humanSNP2.avinput”, “humanSNP2.hg19_multianno.txt”, and “humanSNP2.hg19_
multianno.vcf ”. The first one is an ANNOVAR input file. The second one is the annota-
tion file with annotation columns, and the last one is a VCF file with annotation added
to INFO fields. Open each of these files and study their contents.
We can also try the annotation database that we created for SARS-CoV-2. We can anno-
tate “sarscov2.vcf”, which was generated from a previous variant calling example. You can
copy it to the “input” directory for easy use. Thus, we can annotate it using the following
script:

./table_annovar.pl ../input/sarscov2.vcf sarscov2db/ \

-buildver SARSCOV2 \
-out ../output/cov2SNP \
-remove \
-protocol refGene \
160   ◾    Bioinformatics

FIGURE 4.22 All-variant annotation file of SARS-COV-2.

-operation g \
-nastring . \
-vcfinpu

The all-variant annotations of SARS-CoV-2 are shown as in Figure 4.22.

4.5 SUMMARY
The high-throughput sequencing makes variant discovery much easier than the use of
traditional methods like microarrays. Raw data obtained from sequencing technology is
used for the detection of variants including base substitutions, insertions, deletions, and
structural variants. Variants can be in any region of the genome; however, only variants
that affect functions of the genes are studied. The consequences of variants depend on the
affected regions and they may be deleterious implicating in healthy conditions and disease
like cancers or may lead to the appearance of a new strain in bacteria and viruses that
is more infectious and lethal like the recent variants of SARS-CoV2 or more antibiotic-
resistant strain of bacteria. This why the variant discovery using sequencing data gained
importance and it is widely used in genetics, medical diagnosis, and drug discovery.
Sequencing depth, paired-end sequencing, and the use of long reads make variant
detection more accurate and allow detection of large-scale variants like structural vari-
ants, insertions, and deletions.
The variant calling pipelines use SAM/BAM files of whole genome, whole transcrip-
tome, or targeted gene sequences to discover the bases in the samples that are different
from the bases on the same locations on the reference genome. Variant calling programs
use two approaches for variant calling. The first approach is used by bcftools and it is based
on consensus sequence which is formed by collapsing the piled-up aligned reads. The sec-
ond approach is used by the recent variant callers like GATK. This approach is based on
haplotypes of the variants that are more likely to be inherited together. GATK 4 is the most
commonly used program for variant calling. It uses an advanced workflow pipeline called
GATK best practice pipeline which leads to the detection of accurate variants.
After variant identification with a variant calling program and filtering, variants can
be annotated by assigning functional information to variants using annotation programs.
 Variant Discovery   ◾    161

REFERENCES
1. Monroe JG, Srikant T, Carbonell-Bejerano P, Becker C, Lensink M, Exposito-Alonso M, Klein
M, Hildebrandt J, Neumann M, Kliebenstein D et al: Mutation bias reflects natural selection
in Arabidopsis thaliana. Nature 2022, 602(7895): 101–105.
2. Danecek P, Auton A, Abecasis G, Albers CA, Banks E, DePristo MA, Handsaker RE, Lunter
G, Marth GT, Sherry ST et al: The variant call format and VCFtools. Bioinformatics 2011,
27(15):2156–2158.
3. Tam V, Patel N, Turcotte M, Bossé Y, Paré G, Meyre D: Benefits and limitations of genome-
wide association studies. Nat Rev Genetics 2019, 20(8):467–484.
4. Martincorena I, Campbell PJ: Somatic mutation in cancer and normal cells. Science (New
York, NY) 2015, 349(6255):1483–1489.
5. Population genetics [https://www.nature.com/subjects/population-genetics]
6. Danecek P, Bonfield JK, Liddle J, Marshall J, Ohan V, Pollard MO, Whitwham A, Keane T,
McCarthy SA, Davies RM et al: Twelve years of SAMtools and BCFtools. Gigascience 2021,
10(2).
7. Pegueroles C, Mixão V, Carreté L, Molina M, Gabaldón T: HaploTypo: a variant-calling pipe-
line for phased genomes. Bioinformatics 2020, 36(8): 2569–2571.
8. Yun T, Li H, Chang PC, Lin MF, Carroll A, McLean CY: Accurate, scalable cohort variant
calls using DeepVariant and GLnexus. Bioinformatics 2021, 36(24): 5582–5589.
9. Rivas MA, Graham D, Sulem P, Stevens C, Desch AN, Goyette P, Gudbjartsson D, Jonsdottir
I, Thorsteinsdottir U, Degenhardt F et al: A protein-truncating R179X variant in RNF186
confers protection against ulcerative colitis. Nat Commun 2016, 7:12342.
10. Stitziel NO, Stirrups KE, Masca NG, Erdmann J, Ferrario PG, König IR, Weeke PE, Webb TR,
Auer PL, Schick UM et al: Coding Variation in ANGPTL4, LPL, and SVEP1 and the Risk of
Coronary Disease. N Engl J Med 2016, 374(12):1134–1144.
11. Ng PC, Henikoff S: SIFT: Predicting amino acid changes that affect protein function. Nucleic
Acids Res 2003, 31(13):3812–3814.
12. Cingolani P, Platts A, Wang le L, Coon M, Nguyen T, Wang L, Land SJ, Lu X, Ruden DM:
A program for annotating and predicting the effects of single nucleotide polymorphisms,
SnpEff: SNPs in the genome of Drosophila melanogaster strain w1118; iso-2; iso-3. Fly (Austin)
2012, 6(2):80–92.
13. Yang H, Wang K: Genomic variant annotation and prioritization with ANNOVAR and wAN-
NOVAR. Nat Protoc 2015, 10(10):1556–1566.
 Chapter 5

RNA-Seq Data Analysis

5.1 INTRODUCTION TO RNA-SEQ

RNA sequencing or shortly RNA-Seq is a high-throughput sequencing technique to
­examine gene expression or profiling in biological samples. Rather than the whole genome,
RNA-Seq focuses only on transcriptomes to tell us which one of the genes in the samples
are turned on or off and to what extent. The transcriptome is the set of all messenger RNA
transcribed from genes in cells. Genes represent a small portion of the whole genome of
an organism. For instance, genes are only around 3% of the human genome. RNA, in gen-
eral, can either be translated into proteins or play functional role such as rRNA for protein
synthesis, tRNA for amino acid transfer, and microRNA that plays a role in gene regula-
tion. Out of whole transcriptome, protein-coding RNA (mRNA) is around 2%–4%, while
the greatest percentage of RNA molecules is the other types of RNA. The gene expres-
sion studies mainly focus on mRNA only because it is translated into proteins. Proteins
are required for the structure, function, and regulation of the cells forming body’s tissues
and organs. Dysfunctions of proteins implicate in most of the diseases and healthy condi-
tions. By studying the mRNA, we can find out which genes are active in a particular cell
type or under a certain condition, giving us a clue about the function of the cells and the
biological activities under that condition. Moreover, we can compare the gene activities
in different types of cells or in different conditions and study how the patterns of gene
expression change over time or in response to different stimuli. We can use such informa-
tion to understand biological activities, normal functions of the cells, effects of pathogens
or therapy, or response to any factors. In short, examining mRNA of cells under a specific
condition provides a snapshot or a checkpoint of the cellular activities at that moment.
Therefore, it is mostly studied to investigate how a certain disease like cancer may affect the
gene expression or to study the cellular response to a treatment or a condition.
A gene is a DNA sequence that consists of several functional components; each plays
a different role in the process of gene transcription into RNA. In general, the functional
gene regions can be divided into two main units: the promoter region and the coding
region. The promoter region controls the transcription of the mRNA, whereas coding

DOI: 10.1201/9781003355205-5 163

164   ◾    Bioinformatics

region consists of codons that are translated into amino acids. The major difference
between prokaryotic and eukaryotic genes is that the coding regions of a eukaryotic gene
(also known as exons) are found between the non-coding regions (known as introns)
that make gene expression in eukaryotes more complex. In eukaryotes, once mRNA
is transcribed, the introns are removed by splicing and the final mRNA transcript will
form an open reading frame (ORF) that can be translated into a polypeptide. A single
eukaryotic gene may code for several genes by the so-called alternative splicing produc-
ing multiple proteins called isoforms. In eukaryotic mRNA, introns are spliced out at
specific splicing sites found at the 5′ and 3′ ends of introns. Most commonly, the intron
sequence that is spliced out begins with the dinucleotide GT (donor splice site) at its 5′
end and ends with AG (acceptor splice site) at its 3′ end [1]. Very rare splice site may be
found at AT of 3′-end and AC at the 3′-end of an intron and also at GC (5′-end) and AG
(3′ end) of an intron [2]. On the other hand, a prokaryotic gene consists of a continuous
sequence of codons that makes gene expression in prokaryotes simpler. A typical func-
tional promoter region of a eukaryotic gene consists of the transcription start site (TSS,
where Polymerase II binds), a core promoter called TATA box (around 40 nucleobases
upstream from the transcription start site), upstream promoters (different in sequence
from a gene to another), and the enhancer (may be found thousands of nucleobases
upstream or downstream). TATA box is consistent among eukaryotic species; it is the
site where transcription factors (TATA box-binding proteins or TBP) bind to form a
complex that is necessary for t­ ranscription to take place. The upstream promoters can
bind to different types of proteins to activate the gene. The enhancer binds to special
types of proteins to bend the DNA sequence so the enhancer that can bind to the protein
complex on the promoter region.
The prokaryotic genes are organized into blocks called operons. Genes in a single operon
are regulated and expressed together. An operon contains all genes that encode proteins
which carry out together a specific function. For instance, the bacterial genes required for
lactose as energy source are found next to each other in the lactose operon (lac operon).
The gene structure and regulation of prokaryotic genes is different from that of eukaryotic
ones. There is no intron in prokaryotic genes and an operon, which contains several genes,
has a single promoter controlling the gene expression of all genes in that operon. There are
three types of proteins (repressors, activators, and inducers) that regulate expression of
genes in an operon.
In general, the normally functioning cells of a living organism must be able to control
gene transcription by turning on and off genes based on the need of the cells for protein
synthesis and other functions. The process of turning on a gene to be transcribed to syn-
thesize functional gene products is known as the gene expression. The activity of a gene is
measured by the amount of its transcript (mRNA). Laboratory techniques like northern
blot, serial analysis of gene expression (SAGE), quantitative PCR (qPCR), and DNA micro-
array have been used to study gene expression. However, these techniques can be replaced
these days by the high-throughput RNA sequencing, which provide more information
than any of the other techniques.
 RNA-Seq Data Analysis   ◾    165

5.2 RNA-SEQ APPLICATIONS

The RNA-Seq emerged as an alternative to microarrays for gene expression research. It
uses the high-throughput sequencing of RNA to examine the quantity of RNA in biologi-
cal samples. The primary application of the RNA-Seq is the gene profiling, which is the
determination of the patterns of genes expressed at the level of transcription under specific
condition or in a specific type of cells to give an overall picture of biological activities.
Simply, it focuses on the transcriptome to tell us which of the genes in the genome of an
organism are turned on or off and to what extent. Compared to microarrays, RNA-Seq
provides an unbiased information about transcripts and it can detect a larger number of
differentially expressed genes. In addition to gene expression profiling, RNA-Seq is also
used in a variety of applications, including complex differential gene expression, single-
cell RNA-Seq, small RNA profiling, variants identification and allele-specific expression,
detection of alternative splicing patterns, system biology, and fusion gene detection.
The differential gene expression is the most important application of RNA-Seq, and it
is the application on which we will focus in detail in this chapter. In differential expres-
sion analysis, we will compare transcriptomes across different conditions of interest. These
conditions can be different samples, development stages, treatments, disease conditions,
etc. The sequencing reads produced from the samples are counted to measure the gene
expression levels. The read counts then are modeled for the statistical significance that
helps us to conclude whether the difference between samples is significant. The differential
analysis can also be followed by annotation using functional annotation database such as
Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) pathways.
RNA-Seq is also used for the single-cell RNA analysis in which the RNA is extracted
from a single cell or a single type of cells. Single-cell RNA sequencing provides transcrip-
tional profiling of individual cells that enable researchers to study the genes expressed
in the single-cell level and how expression level varies across thousands of cells within
a heterogeneous sample. Most of living cells cannot be cultivated in vitro. Therefore, the
single-cell RNA-seq may lead to the discovery of novel species, pathway, or regulatory
processes of biotechnological or medical relevance. The workflow of single-cell RNA-Seq
generally involves single-cell isolation, cDNA library preparation, RNA-Seq, and RNA-Seq
data analysis [3].
The RNA-Seq is also used to study the small non-coding RNA (sRNA-Seq), which is a
type of non-coding RNA that includes microRNA, small interfering RNA (siRNA), and
P-element-induced wimpy testis (Piwi)-interacting RNA, small nucleolar RNA (snoRNA),
and small nuclear RNA (snRNA). The small RNA-Seq data analysis is known to be chal-
lenging due to the short length of the sequence and non-unique genomic origin.
The RNA-Seq can also be used for variant detection, allele-specific expression, and
expression quantitative trait loci (eQTL). The polymorphisms in coding region (exons) are
better to be detected using transcriptomic sequencing. However, the variant calling pipe-
line must be applied. But the basic RNA-Seq pipeline can be used to quantify the allele-
specific expression (ASE) that is affected by single-nucleotide polymorphisms (SNPs). The
ASE analysis will show which one of two alleles is highly transcribed into mRNA and
166   ◾    Bioinformatics

which one is lowly transcribed or even not transcribed at all. The RNA-Seq count data is
used as an alternative to microarray data in eQTL analysis. QTL analysis is a statistical
method that links phenotypic data (trait measurements) and genotypic data (markers usu-
ally SNPs) in an attempt to explain the genetic basis of variation in complex traits. On the
other hand, eQTL analysis links markers (genotype) with gene expression levels measured
in a large number of individuals and the data is modeled using generalized linear models.
RNA-Seq is a powerful tool for detecting alternative splice patterns, which are important
to understand development of human diseases. Paired-end sequencing enables sequence
information from both ends and help in detecting splicing patterns without requirement
for previous knowledge of transcript annotations. The single-molecule, real-time (SMRT)
sequencing is the core technology powering long-read sequencing that allows examination
of splicing patterns and transcript connectivity in a genome-scale manner by generating
full-length transcript sequences.
RNA-Seq is also used for fusion gene detection. A fusion gene is a gene made by join-
ing two different genes. It is usually created when a gene from one chromosome moves to
another chromosome. The fusion gene is transcribed into mRNA that will be translated
into fusion protein. The fusion proteins implicate usually in some types of cancer includ-
ing leukemia; soft tissue sarcoma; cancers of the prostate, breast, lung, bladder, colon,
and rectum; and CNS tumors. Paired-end RNA-Seq data are usually used for fusion gene
detection [4].
Other kinds of RNA-Seq applications include integration of RNA-Seq data analysis with
other technologies.
The library preparation of the mRNA is similar to that of DNA. However, mRNA must
be separated from other types of RNA by enrichment technique which uses either PCR
amplification or the depletion of the other types of RNA. The RNA must be converted
into complementary DNA (cDNA) by reverse transcription before library preparation. As
DNA library, the cDNA library preparation involves fragmentation and adaptor ligation
to each end of the fragments. The cDNA fragments then are sequenced with the sequenc-
ing machine and the sequencing can either be single end (forward strand only) or paired
end (forward and reverse strands). The sequencing generates sequence data in a form of
reads in FASTQ files. Those reads are the sequenced fragments of the expressed genes in
the sample.

5.3 RNA-SEQ DATA ANALYSIS WORKFLOW

The first steps of the RNA-Seq are the same as in other sequencing applications. The
sequencing raw data (usually in FASTQ files) must pass through the quality control steps
that were discussed in detail in Chapter 1. In general, the steps of the workflow include
quality control, read alignment, read quantification, differential expression, annotation,
and interpretation (Figure 5.1).

5.3.1 Acquiring RNA-Seq Data

The RNA-Seq raw data are sequence reads produced by a sequencing instrument. RNA-
Seq sequence raw data for some projects are available in public databases and can be
 RNA-Seq Data Analysis   ◾    167

FIGURE 5.1 RNA-seq data analysis workflow.

downloaded and used for research purposes or for learning. Most RNA-Seq raw sequence
data are in FASTQ file format. When analyzing the RNA-Seq data, we must pay attention
to the design of the study. For instance, if the purpose is the differential gene expression,
there must be control raw data that we can use for comparison. The control raw data is
determined by the research goal. In conditions like cancers, researchers may use sequenc-
ing raw data of healthy tissue as control against the raw data of the affected tissue and
both from the same individual. However, researchers may also intend to compare gene
expression across individuals or samples. Most researchers include replicate samples in the
design of their study, and thus, there will be multiple raw data for a single sample. Replicate
samples will reduce errors generated by the laboratory technique used and also the possible
errors generated during the sequencing steps.
For practicing, we will use RNA-Seq raw data of a breast cancer study for differential
gene expression in tumor cells. The data is in six FASTQ files (three replicates for tumor
and three replicates for normal) containing paired-end reads of the size 151 bases. For the
sake of simplicity, the files include only the RNA-seq reads of chromosome 22. The data
was adapted to be as simple as possible, so its processing does not take too much time. To
keep the files organized, create a main directory “rnaseq” to be as the project directory and
create inside it the subdirectory “fastq”, and then, inside this subdirectory, download the
raw data from “https://github.com/hamiddi/ngs”. To avoid repetition, assume that the raw
data files have been cleaned from adaptors, duplicates, and the low-quality reads.

5.3.2 Read Mapping

The read mapping follows reprocessing and cleaning of the raw data. The accuracy of anal-
yses depends heavily on the read mapping. The mapping, as discussed in Chapter 2, is the
168   ◾    Bioinformatics

process of mapping reads to a reference genome producing a SAM/BAM file that contains
the mapping information. Refer to Chapter 2 for the read mapping and the content of
SAM/BAM files. When dealing with RNA-Seq data, we can either align the reads to a
reference genome or a reference transcriptome. When we align RNA-Seq to a eukaryotic
reference genome, we must use an aligning program like STAR that is able to detect the
splice junctions. The reads in this case will map to the exons leaving introns and other
non-coding regions of the genome uncovered. On the other hand, when aligning RNA-Seq
reads to a reference transcriptome, the aligned reads may cover the entire sequence. This
strategy is preferable when reads are very short (less than 50 bases). The downside of align-
ing reads to a transcriptome is that we may miss some novel genes since the transcriptome
is made up of only known transcripts. As discussed in Chapter 2, there are several align-
ers; however, for RNA-Seq, we prefer to use a splice-aware aligner that is able to introduce
long gaps to span introns when aligning reads to a reference genome. The commonly used
aligners for RNA-Seq data include STAR [5], segemehl [6], GEM [7], BWA [8], BWA-MEM
[8], and BBMap [9].
Before deciding on which of the aligners to use with RNA-Seq reads, make sure that the
aligner is splicing-aware and able to distinguish between reads aligned across exon–intron
boundaries and reads with short insertions [10]. The splicing-aware aligners include STAR
[5], GSNAP [11], MapSplice [12], RUM [13], and HISAT2 [14]. Each of these aligners has
different advantages and disadvantages in terms of memory efficiency, performance, and
speed. Refer to the user guide of any of these aligners to learn more about them. We will
use STAR (Spliced Transcripts Alignment to a Reference) as an example aligner for align-
ing RNA-Seq data. Several studies found that STAR is one of the most accurate aligners of
RNA-Seq reads [15]. However, STAR requires a large memory for indexing and mapping.
The reference sequence must be indexed by STAR before alignment. STAR begins mapping
process by aligning the longest reads that exactly match a single or multiple location on the
reference sequence. For partially aligned reads, STAR will attempt to align the unmapped
region to a different region. Those parts of the reads which align to different locations of
the reference sequence are called seeds. If STAR does not find an exact match to a read on
the reference sequence, the read will be extended by inserting gaps. If the extension does
not give a good alignment, it will be removed. In the second step of the STAR alignment
process, multiple seeds will be clustered based on proximity to a set of anchor seeds. The
clustered seeds are stitched together based on the best alignment score [5].
When reads are mapped to a reference sequence, the percentage of mapped reads reflects
the quality of the alignment. Low percentage indicates contamination of the DNA. Read
coverage and depth on exons are other factors that determine alignment quality.
Above, we have downloaded the FASTQ files in the directory “fastq”. We can map reads
in the FASTQ files to a reference genome using STAR program. STAR is a short read aligner
designed to align RNA-Seq reads to a reference sequence (genome or transcriptome). For
aligning reads in the FASTQ files using STAR, we need to download a reference sequence
together with its annotation file in GTF format. Since our example FASTQ files are from
human samples, we need to download the latest human genome and its annotation file.
 RNA-Seq Data Analysis   ◾    169

To keep the files organized, we will create two subdirectories in our project directory:
“­refgenome” where we will store the reference genome and “gtf” where we will save the
GTF annotation file.
The sequences of reference genomes and annotation are available in many sequence data-
bases such as Ensembl, UCSC, and NCBI genome database. iGenomes built by Illumina
has facilitated the process of downloading the reference data for the frequently analyzed
organisms. Genome builds in FASTA and their annotation in GTF/GFF files from the
above major databases are available for download. The iGenomes website that includes
the download links is available at “https://support.illumina.com/sequencing/sequencing_
software/igenome.html”. Reference data can also be downloaded from “https://hgdown-
load.soe.ucsc.edu/goldenPath/hg38/bigZips/”. For aligning with STAR, we will download
the UCSC human reference genome sequence in FASTA and gene annotation in GTF file
because the chromosomes are indicated by names rather than accession numbers. While
you are in the main directory “rnaseq”, run the following bash script to create the subdi-
rectories and to download the human reference genome and its gene annotation:

mkdir refgenome
wget \
-O “refgenome/hg38.fa.gz” \
“https://hgdownload.soe.ucsc.edu/goldenPath/hg38/bigZips/hg38.
fa.gz”
gzip -d refgenome/hg38.fa.gz
mkdir gtf
wget \
-O “gtf/hg38.ncbiRefSeq.gtf.gz” \
“https://hgdownload.soe.ucsc.edu/goldenPath/hg38/bigZips/genes/
hg38.ncbiRefSeq.gtf.gz”
gzip -d gtf/hg38.ncbiRefSeq.gtf.gz

Mapping reads to a reference genome using STAR is a two-step process: creating a refer-
ence sequence index and then mapping reads to the reference sequence.
The following command creates the STAR index for the reference genome sequence.
The “--runThreadN” specifies the number of processors to use, “--genomeDir” specifies
the directory where the index files will be saved, “--genomeFastaFiles” and “--sjdbGTFfile”
specify the directories of the reference genome file and annotation file, respectively, and
“--sjdbOverhang” specifies the read length -1 (read length minus one).

mkdir indexes
STAR --runThreadN 4 \
--runMode genomeGenerate \
--genomeDir indexes \
--genomeFastaFiles refgenome/hg38.fa \
--sjdbGTFfile gtf/hg38.ncbiRefSeq.gtf \
--sjdbOverhang 150
170   ◾    Bioinformatics

For indexing the human genome, STAR may require more than 32GB of RAM and it may
take a long time depending on the computer RAM and the number of processors used.
Once indexing has been created, you can align the reads in FASTQ files to the reference
genome sequence. Since the practice sequence data is paired end, there are two FASTQ files
(r1 and r2) for each sample. STAR requires these two files as inputs in addition to the refer-
ence genome index we created above. Instead of running STAR for each sample, we can use
bash script as follows to align the reads of the sample files in “fastq” directory:

mkdir starout
mkdir bam
cd fastq
for i in $(ls *.gz | rev | cut -c 13- | rev | uniq);
do
STAR --runThreadN 4 \
--runMode alignReads \
--outSAMtype BAM SortedByCoordinate \
--readFilesCommand zcat \
--genomeDir ../indexes \
--outFileNamePrefix ../starout/${i} \
--readFilesIn ${i}_r1.fastq.gz ${i}_r2.fastq.gz
done
for f in $(ls *.bam | rev | cut -c 30-|rev);
do
mv ${f}Aligned.sortedByCoord.out.bam ../bam/${f}.bam
done
cd ..

In the above, first we created a directory “starout” for STAR output and “bam” for the BAM
files. The Linux bash command “$(ls *.gz | rev | cut -c 13- | rev | uniq)” lists the names of
the FASTQ files found in the directory and cut the common name (ID) for the files of each
sample. The “for loop” will provide that common name as a variable “${i}” to assemble the
FASTQ file names and the output file name prefix to the STAR command each time until
all FASTQ files are processed. There will be output files including alignment log files and a
BAM file for each sample (6 BAM files for all example data). The second “for loop” renamed
the long BAM file names and moved the files to the “bam” subdirectory.
The bash scripting comes in handy whenever there is a task that needs to be repeated
multiple times.
The STAR output log files provide important statistics about the read alignment. You
can refer to STAR manual to read about the output log files.
Before moving to the next step, you need to index the BAM files using “samtools index”
command as follows:

#index bam
cd bam
for i in $(ls *.bam);
 RNA-Seq Data Analysis   ◾    171

do
samtools index ${i}
done
cd ..

5.3.3 Alignment Quality Assessment

After aligning the reads to a reference genome and obtaining the alignment information
in a BAM file, we may need to check the alignment quality before the next step in the data
analysis. The aligned reads may contain duplicate or overrepresented reads produced by
the sequencing process. STAR generates log files that contain some important statistics.
In our example, the log files of each replicate have been saved in “starout” directory. Open
these files and study the statistics included in each file. Figure 5.2 shows alignment statis-
tics in a STAR log file “*final.out”. The log file gives an idea about the mapped reads and
splice sites as well as other important statistics.

FIGURE 5.2 STAR alignment statistics.

172   ◾    Bioinformatics

As shown in Figure 5.2, the alignment statistics of one of the BAM files show the total
number of reads, the average reads length, the number and percentage of uniquely mapped
reads, splice statistics, statistics of the reads mapped to multiple genes, statistics of the
unmapped reads, and chimeric reads. Pay attention to the reads mapped to multiple loci
and chimeric; when their number is large, that indicates low-quality alignment. Remember
that this BAM file includes the alignments of chromosome 22 only. The number of reads
will be huge if the BAM file contains the alignments of all chromosomes.
In addition to the statistics in the STAR log files, there are a variety of programs for
assessing alignments in BAM files. Examples of those programs include Qualimap [16],
RNA-seQC [17], and RSeQC [18]. Those programs compute metrics for RNA-Seq data,
including per-transcript coverage, junction sequence distribution, genomic localization of
reads, 5′–3′ bias, and consistency of the library protocol. As an example, you can download
and use Qualimap to obtain an overall view about the alignment quality on an HTML for-
mat. You can download Qualimap from “http://qualimap.conesalab.org/” and unzip it in
your project directory. Run Qualimap for each sample and study the reports carefully. The
following script is an example of how to use it:

mkdir qc
qualimap_v2.2.1/qualimap rnaseq \
-outdir qc \
-a proportional \
-bam bam/norm_rep1.bam \
-p strand-specific-reverse \
-gtf gtf/hg38.ncbiRefSeq.gtf \
--java-mem-size=8G

The above script creates the directory “qc” where the Qualimap output files will be saved.
The program takes a BAM file and the reference annotation file as inputs and generates
an HTML report that includes summary statistics about read alignments, reads genomic
origin, transcript coverage profile, splice junction analysis, and figures about read genomic
origins, coverage profile along genes, coverage histogram, and junction analysis. As a biol-
ogist, you may need to study these metrics to have a general idea about the sample align-
ment before proceeding.

5.3.4 Quantification
Gene profiling or studying gene expression is centered in the quantification of aligned
reads per gene or locus. Quantification of reads begins by counting the number of reads
aligned to each gene annotated on the sequence of the reference sequence. Given a BAM
file with aligned RNA-Seq reads and a list of genomic features in an annotation file (GFT
format), the task of the read counting program is to count the number of reads mapping
to each feature. In general, a feature, in this case, is a gene which represents a transcript
or unions of exons of a gene for eukaryotic organisms. Some programs can also consider
exons as features. This is especially useful for checking alternative splicing in the eukary-
otic genes. A read in the BAM file may map to a single feature (unique) or may map or
 RNA-Seq Data Analysis   ◾    173

overlap with multiple features (non-unique). A read that maps or overlaps with multiple
features is considered as ambiguous and will not be counted. Only reads mapping unam-
biguously to a single feature are counted.
For RNA-Seq read counts, we can use a variety of programs but the most used free open-
source programs include HTSeq-count [19] (Python-based program) and FeatureCounts
[20], which is used as a Linux command-line program or as a function in the R package
Rsubread. Both these programs require BAM files and GFT file (reference annotation file)
as inputs.
In the following, we will use HTSeq-count to count RNA-seq reads aligned to the genes
in chromosome 22. Install HTSeq-count by following the installation instructions avail-
able at “https://htseq.readthedocs.io/en/master/install.html”. The following HTSeq-count
command counts aligned reads in all BAM files stored in the “bam” subdirectory and saves
the output in “features/htcount.txt”. The “-m union” option specifies the union mode to
handle reads overlapping more than one feature. We used “--additional-attr=transcript_
id” to add transcript accession numbers to the output.

mkdir features
htseq-count \
-m union \
-f bam \
--additional-attr=transcript_id \
-s yes bam/*.bam \
gtf/hg38.ncbiRefSeq.gtf \
> features/htcount.txt
sed ‘/^__/ d’ < htcount.txt > htcount2.txt

The “sed” command has been used to remove the last rows that begin with “__” and to save
that change in a new file “htcount2.txt”, which we will use in the next step of the analysis.

FIGURE 5.3 HTSeq-count feature count.

174   ◾    Bioinformatics

The HTSeq-count output file contains the feature count for each sample as shown in
Figure 5.3. The feature count file includes tab-delimited columns for gene symbols, tran-
script IDs, and a count column for each sample. We can notice that some genes have zero
reads aligned to them. Later, we will filter out the genes that have no aligned reads and the
one with low coverage.

5.3.5 Normalization
In general, when we analyze gene expression data, we may need to normalize it to avoid
some biases that may arise due to the gene lengths, GC contents, and library sizes (the total
number of reads aligned to all genes in a sample) [21]. The normalization of count data is
important for comparing between expression of genes within the samples and between
different samples. The normalized gene length fixes the bias that may affect within-sample
gene expression comparison. It is known that a longer gene would have a higher chance to
be sequenced than a shorter gene. Consequently, a longer gene would have a higher number
of aligned reads than a shorter one at the same gene expression level in the same sample.
The GC content also affects within-sample comparison of gene expressions. The GC-rich
and GC-poor fragments tend to be under-represented in RNA-Seq sequencing, and hence,
the gene with the GC content closest to 40% would have higher chance to be sequenced
[22]. The library size affects the comparison between the expressions of the same gene in
different samples (between-sample effect).
There are several normalization methods for adjusting the biases resulted from the
above-mentioned possible causes. Choosing the right normalization method depends on
whether the comparison is within-sample or between-samples. In the following, we will
discuss some of these normalization methods used by gene expression analysis program
like EdgeR and DESeq2 [23].

5.3.5.1 RPKM and FPKM

RPKM [24] (the reads per kilobase of transcript per million reads mapped) is a normalized
unit for the counts of reads aligned to genes (normalized gene expression unit). It scales
the count data by gene length to adjust for the sequencing bias arising from the differences
in gene lengths. The RPKM is used for within-sample gene expression comparison (i.e.,
comparison between genes in the same sample).
Assume that N reads are aligned to the reference sequence and only k reads are aligned
to the gene i of length li bp, the PRKM of the gene i is calculated as

ki
RPKMi = (5.1)
 i  N 
l
 3   6 
10 10

In the denominator of Formula 5.1, the length of gene (l) (in base) is divided by 1000 to
be in kilobase and the total number of reads aligned to the reference sequence is divided
by 1000,000 (million). When the number of reads aligned to a gene is divided by the
 RNA-Seq Data Analysis   ◾    175

denominator, we will obtain the reads per kilobase per million or RPKM. The RPKM is
for single-end reads. However, for paired-end reads, both forward and reverse reads are
aligned, and thus, “fragment” is used instead of “read” and the normalized unit of gene
expression in this case is FPKM (fragment per kilobase per million).

5.3.5.2 Transcripts per Million

The transcripts per million (TPM) [25] is proposed as an alternative to RPKM and FPKM
to adjust for the bias of gene length and to be used for within-sample differential gene
expression. The TPM represents the abundance of reads aligned to gene i in relation to
the abundance of the reads aligned to other genes in the same sample. To normalize the
RNA reads counts, first, for any gene, divide the number of reads aligned to it by its length,
forming the count per gene length in base. Then, divide the count per length in base by the
sum of all counts (per length in base of every gene) and multiply by 1000,000 forming the
transcript per million or TPM.

ki 1
TPMi = × × 106 (5.2)
li  kj 


∑ 
j lj 

5.3.5.3 Counts per Million Mapped Reads

The counts per million (CPM) mapped reads normalize the number of reads that map to a
particular gene after correcting for sequencing depth and transcriptome composition bias
[26]. CPM is used for between-sample differential analysis to compare between the gene
expressions of the same gene in different samples. It is not suitable for within-sample gene
expression comparison because it does not adjust for the gene length. The CPM of a gene is
defined as the number of reads mapped to a gene divided by the total number of mapped
read (or library size) multiplied by 1000,000.

ki
RPMi = × 106 (5.3)
N

5.3.5.4 Trimmed Mean of M-values

The Trimmed Mean of M-values (TMM) [27] is used by edgeR for between-sample dif-
ferential gene expression. It uses the relative gene expression of two samples: one is the
sample of interest (treated), and the other is the reference that we wish to use as a baseline
for comparison. The gene-wise log-fold change of gene g M g is given as:

 Ygj 
 Kj 
M g = log 2  (5.4)
Ygr 
 K r 
176   ◾    Bioinformatics

where Ygj is the observed read count of the gene g of interest in the sample j, K j is the size
of library of sample j (total number of aligned reads), and Ygr is the observed reads count of
the same gene g in the reference sample r of library size K r .
Then, the value of the gene expression fold change, M g , is trimmed by 30% followed
by taking the weighted average for the trimmed M g using inverse of the variances of read
counts of genes (Vgi ) as weight since the log-fold changes from gene with larger read counts
will have lower variance on the logarithm scale. The TMM adjustment, f j , is given as

∑V M
n
gi gi
f = i
(5.5)
∑V
j n
gi
i

The adjustment f j is an estimate for relative RNA production of two samples. The TMM
normalization factor for the sample j with m genes is given by

Nj = fj ∑Y g =1
gj (5.6)

TMM does not correct the observed read counts for the gene length, and hence, it is not
suitable for comparison between the gene expressions in the same sample.

5.3.5.5 Relative Expression

For a given sample j, the relative expression (RE) scaling factors are calculated as the
median of the ratios of observed counts to the geometric mean across all samples (pseudo-
reference sample, r) [28]. The scaling factors are calculated as follows:

median y gj
Nj = (5.7)
g 1/n

∏ Ygr 
n

 r =1 

5.3.5.6 Upper Quartile

The upper quartile (UQ) normalization factor is computed as the sample upper quartile
(75th percentile) of gene counts for the genes with no zero counts for all samples [29].

5.3.6 Differential Expression Analysis

Most differential expression programs have functions that normalize gene expression
count data as part of the analysis. The design of the study is crucial in the differential
expression analysis as we discussed in the introduction of this chapter. The conditions
that group the samples and sample replicates must be determined as metadata before the
analysis. A simple gene expression study is made up of two groups (e.g., treated and control
 RNA-Seq Data Analysis   ◾    177

or healthy and diseased) but it can also be a complex study that includes more than a
single factor (factorial design). Once the study design has been determined as metadata,
inferential statistics is used to identify which gene has statistically significant change in
expression compared to the same gene in a reference sample. The fundamental step in
the differential expression analysis is to model the association between gene counts (Y)
and the covariates (conditions) of interest (X). The number of replicates is crucial for the
statistical differential analysis. Most of the time, the number of replicates in an RNA-Seq
study is small. Instead of non-parametric statistical analysis, most programs for RNA-Seq
data analysis use generalized linear models (GLMs) by assuming that the count data fol-
lows a certain statistical distribution. That approach also assumes that each RNA-Seq read
is sampled independently from a population of reads and the read is either aligned to the
gene g or not. When the read is aligned to the gene g, we call this a success and otherwise
is a failure. The process of random trials with two possible outcomes (success or failure) is
called Bernoulli’s process. Thus, according to the probability theory, the number of reads
(successes), Yg , for a given gene g from sample j follows a binomial distribution.

Ygj ~ Binomial (n j , π gj ) (5.8)

Assume Ygj is the number of reads sequenced from sample j, n j represents the number of
independent trials in Bernoulli’s process, π gj is the probability of success (a read is aligned
to the gene g in sample j), and 1 − π gj is the probability of failure
Assume also that for the gene g on the sample j and that gene has the length l g and read
count Ygj . All possible positions in g that can produce a read can be described as Ygj l g [30].
Thus, probability of success, π gj , is given as

Ygj l g
π gj = (5.9)
∑
G
Ygj l g
g =1

where G is the number of genes in the sample.

According to the binomial distribution, the mean of the read counts is given as

µ gj = n j × π gj (5.10)

The probability that the number of reads (X gj = x gj ) for a given gene is given by

 nj 
P (Ygj = y gj ) =   π gj gj (1 − π gj )
y n j − y gj
(5.11)
 y g 

However, since in RNA-Seq count data, a very large number of reads are represented and
the probability of aligning a read to a gene is very small, the Poisson distribution is more
appropriate than the binomial distribution if the mean of read counts of a gene is equal to
the variance as the Poisson distribution assumes.
178   ◾    Bioinformatics

Yg ~ Poi ( λ g ) (5.12)

where λ g is the Poisson parameter which represents the rate of change in the count of the
gene g in a sample.
The Poisson distribution assumes that the rate of change is equal to the mean, λ g , which
is also equal to the variance.

µ g = λ g = var (Yg ) (5.13)

The probability that Yg = y g

λ g yg × e − λg
p (Yg = y g ; λ g ) = (5.14)
yg !

To model the RNA-Seq count data with the Poisson distribution it requires that the mean
is equal to the variance. A key challenge is the small number of replicates in typical RNA-
Seq experiments (two or three replicates per condition). Therefore, inferential methods
that deal with each gene separately may suffer, in this case, from lack of power, due to
the high uncertainty of within-group variance estimates. This challenge can be overcome
either by grouping the count data into groups and then calculating the variance and the
mean in each group or by pooling information across genes by assuming the similarity of
the variances of different genes measured in the same experiment. In general, RNA-Seq
count data suffers from over-dispersion, where variance is greater than the mean. There
are a variety of software that use different technique for modeling the RNA-Seq count
data, but most of them use quasi-Poisson, negative binomial, or quasi-negative binomial
distribution, which deal with over-dispersed data.
The quasi-Poisson is similar to the Poisson distribution, but the variance is linearly cor-
related to the mean of the counts [31].

Yg ~ qPoi ( µ g ,θ g ) (5.15)

var (Yg ) = θ g µ g (5.16)

where θ g is the dispersion parameter and µ g is the mean count.

In the negative binomial distribution, the variance is the function of the mean as

Yg ~ NB ( µ g ,α g ) (5.17)

var (Yg ) = µ g + αµ gP (5.18)

where α is the dispersion parameter and P is an integer but commonly we use P = 2 (NB2
or quadratic model).
 RNA-Seq Data Analysis   ◾    179

var (Yg ) = ( µ g + αµ g2 ) (5.19)

In the quasi-negative binomial distribution, the variance is modeled as follows:

(
var (Yg ) = σ g2 µ g + θµ g2 ) (5.20)

The RNA-Seq study design may include a single or several conditions called factors. A
researcher usually may be interested in testing the effect of a condition. For instance,
assume that a researcher wants to study breast cancer in women. She conducted an RNA-
Seq study on samples from healthy and cancer tissues of five affected women. The analysis
programs require a matrix that describes the design called a design matrix. The design
matrix defines the model (structure of the relationship between genes and explanatory
variables), and it is also used to store values of the explanatory variable [32]. The design
matrix will be created from the study metadata as shown in Table 5.1.
The design matrix will include dummy variables setting the level of each factor to either
zero or one as we will see soon.
The generalized linear model will fit the data of this study design so that the expression
of each gene will be described as a linear combination of the dummy explanatory variables.

y = β0 + β1 *Patient + β 2 *Condition + ε (5.21)

where y is the response variable that represents the gene expression in a specific unit, β0
is the intercept or the average gene expression when the other parameters are zero, and β1
and β 2 are the generalized linear regression parameters that represent the effect of each
explanatory variable. A log-linear model is used as

log µ gi = XiT β g + log N i (5.22)

where XiT is a vector of covariates (explanatory variables) that specifies the conditions/­
factors applied to sample i and β g is a vector of regression coefficients for the gene g.

TABLE 5.1 Sample Information or

Metadata for the Design Matrix
SampleID Condition Patient
norm_rep1 Norm Rep1
norm_rep2 Norm Rep2
norm_rep3 Norm Rep3
tumo_rep1 Tumo Rep1
tumo_rep2 Tumo Rep2
tumo_rep3 Tumo Rep3
180   ◾    Bioinformatics

5.3.7 Using EdgeR for Differential Analysis

EdgeR (Empirical Analysis of Digital Gene Expression Data in R) is an R Bioconductor
package for differential expression analysis of RNA-Seq data. It performs differential
expression of replicated count data using generalized linear model for the over-dispersed
count data and the models account for both biological and technical variability. EdgeR
uses negative binomial distribution to model the RNA-Seq count data.
Assume that for each sample i, the total number of reads (library size) is N i , φ g is the
dispersion coefficient, and pgi is the relative abundance of gene g in the experimental group
i. The mean and variance are estimated as follows:

µ gi = N i pgi (5.23)

variance = µ gi (1 + µ giφ g ) (5.24)

For differential expression analysis using the negative binomial regression, the parameters
of interest are the relative abundance of each gene (pgi).
As we have discussed above, the negative binomial distribution changes to the Poisson
distribution when the count data is not dispersed (φ g = 0) or to the quasi-Poisson distribu-
tion if the variance is linearly correlated to the mean. EdgeR estimates the dispersion (φ g )
as the coefficient of variation (CV) of biological variation between the samples. Dispersion
means biological coefficient of variation (BCV) squared that is estimated by dividing
Formula (5.24) by µ gi2 .

CV 2 = 1/ µ gi + φ g (5.25)

EdgeR calculates the common dispersion for all genes and it can also calculate gene-wise
dispersions and then it shrinks them toward a consensus value. Differential expression is
then assessed for each gene using an exact test for over-dispersed data [33].
In the following, we will analyze the non-normalized count data obtained by HTSeq-
count program in the previous step and saved as “htcount.txt” file in the “features” direc-
tory. The analysis will be carried out in R. Therefore, R must be installed on your computer.
The instructions of R installation are available at “https://cran.r-project.org/”. You can also
use R on Anaconda as well. We assume that you have R installed on your computer and it
is running. On R, you will also need to install Limma and EdgeR Bioconductor packages
by following the installation instructions available at “https://bioconductor.org/packages/
release/bioc/html/edgeR.html” to install EdgeR and “https://bioconductor.org/packages/
release/bioc/html/limma.html” to install limma. For the current versions, open R, and on
the R shell, run the following:

if (!require(“BiocManager”, quietly = TRUE))

install.packages(“BiocManager”)
BiocManager::install(“edgeR”)
BiocManager::install(“limma”)
 RNA-Seq Data Analysis   ◾    181

Once you have R, EdgeR, and limma package installed, you will be ready for the next steps
of the differential analysis which can be broken down into the following steps.

5.3.7.1 Data Preparation

For differential analysis, EdgeR requires the count data file and a sample info file. We have
already created the count data file in the previous step, but we need to create the sample
info file that describes the design of the study. We can create the sample info file manually
as shown in Table 5.1. The sample info file is tab-delimited, and the first column contains
the unique sample IDs or the BAM file names. Additional columns can contain the condi-
tions or factors depending on the study design. For our example data, we can create the
sample info file by executing the following bash script while you are in the main project
directory:

cd bam
ls *.bam \
| rev \
| cut -c 5-\
|rev > tmp.txt
echo -e “sampleid\tcondition\tpatient” \
> ../features/sampleinfo.txt
awk -F ‘_’ ‘{print $1 “_” $2 “\t” $1 “\t” $2}’ \
tmp.txt@ ../features/sampleinfo.txt
rm tmp.txt
cd ../features

This script creates the sample info file from the BAM file names and saves it in the “fea-
tures” subdirectory together with the read count data. For your own data, you may need
to modify this script or you can create yours using Linux bash commands or manually.
Then, you need to open R, make the “features” directory as the working directory, and
load both limma and edgeR packages.

library(limma)
library(edgeR)

Load both the count data file and sample info file to the R session as data frame.

seqdata <- read.delim(“htcount2.txt”, stringsAsFactors=FALSE)

sampleinfo <- read.delim(“sampleinfo.txt”, stringsAsFactors=FALSE)

Run the following command to display the first rows of the count data frame:

head(seqdata)

You will notice that the first two columns are the gene symbol and the transcript IDs. The
other six columns contain the read counts. In the next step, we need to separate the count
182   ◾    Bioinformatics

columns in a different data frame called “countdata” and then we need to add column
names and row names to that count data frame. The row names can be the transcript IDs
and the column names can be the sample IDs as listed in the sample info file. The sample
IDs in the sample info file must be in the same order as the columns of the read counts in
the read count file.

countdata0 <- seqdata[,-(1:2)]

head(countdata0)

The “head” function will display the first rows of the “countdata0” data frame. You can
notice that, as shown in Figure 5.4, the data frame is without row names and that the col-
umn names do not indicate the sample names. You can also notice that there are numer-
ous rows with all columns being zeros. This is mainly because we have aligned reads for
chromosome 22 only.
The second step is to make the gene symbol as the row names and sample IDs as col-
umn names of the “countdata0” data frame and to remove rows with zero for all samples
(Figure 5.5).

rownames(countdata0) <- seqdata[,1]

colnames(countdata0) <- sampleinfo$sampleid
countdata <- countdata0[rowSums(countdata0[])>0,]
head(countdata)

After creating a count data frame as in Figure 5.5, the next step is to create a DGEList
object to hold the read counts that will be analyzed by EdgeR. The DGEList object is a
container for the count data and the associated metadata, including sample names, sample,
group, library size, and normalization factors. The DGEList for our example data is created
by the following:

group = factor(sampleinfo$condition)
y <- DGEList(countdata, group=group)

Figure 5.6 shows the created DGEList object. At this time, it holds two slots: counts and
samples. The counts slot contains the count data and the samples slot contains the sample

FIGURE 5.4 The count data frame without row name and column names.
 RNA-Seq Data Analysis   ◾    183

FIGURE 5.5 The data frame after adding row and column names and removing rows with all zeros.

FIGURE 5.6 The DGEList object of the count data.

info (sample id, group, library size (lib.size), and normalization factors (norm.factor)) with
1 values. This data will be updated soon, and more slots will be added as well.

5.3.7.2 Annotation
The row names of the count data frame are the gene symbols as shown in Figure 5.5. For
some of the downstream analysis, we may need the count data to be annotated with the
NCBI Entrez IDs and full gene names which are not included in the count dataset at this
point. To add these annotations to the DGEList object (y), we need to make the Entrez IDs
as the row names instead of the gene symbols. To obtain the Entrez IDs and gene names,
we need to install and load the “org.Hs.eg.db” Bioconductor package, which is a genome-
wide annotation for human based on mapping using Entrez Gene identifiers [34]. You can
install and upload this package by running the following script on R prompt:

if (!require(“BiocManager”, quietly = TRUE))

install.packages(“BiocManager”)
BiocManager::install(“org.Hs.eg.db”)
library(org.Hs.eg.db)
184   ◾    Bioinformatics

FIGURE 5.7 Annotation columns available on “org.Hs.eg.db”.

FIGURE 5.8 Adding annotation to the count data.

The available annotation column names are displayed with the following (Figure 5.7):

columns(org.Hs.eg.db)

Each of these annotation columns has a row value corresponding to the gene annotated
in the reference sequence. We can select the annotation columns that we need and add
an annotation slot with the selected columns to the DGEList object. The following script
creates a vector of the Entrez IDs mapped to the gene symbol on the counts data, makes
the Entrez IDs as the row names, selects annotation columns mapped to the count data,
adds the annotation as a slot to the DGEList object, and finally removes any row without
an Entrez ID:

ENTREZID <- mapIds(org.Hs.eg.db,rownames(y),

keytype=”SYMBOL”,column=”ENTREZID”)
rownames(y$counts) <- ENTREZID
ann<-select(org.Hs.eg.db,keys=rownames(y$counts),
columns=c(“ENTREZID”,”SYMBOL”,”GENENAME”))
head(ann)
y$genes <- ann
i <- is.na(y$genes$ENTREZID)
y <- y[!i, ]

Figure 5.8 shows the annotation slot “genes” that includes Entrez IDs, gene symbols, and
gene names mapping to the count data “counts”.

5.3.7.3 Design Matrix

The design matrix includes dummy variables that define the covariates of the model,
depending on the study design to answer specific research questions. We will define the
 RNA-Seq Data Analysis   ◾    185

design matrix using the sample information in the “sampleinfo.txt” that we have created
above. In EdgeR, the design matrix can be defined with or without an intercept. The inter-
cept is used when there is a reference for the differential expression analysis. When the
design matrix is defined without an intercept, the differential analysis can be performed
by using a contrast as we will do. In the following, we define a design matrix without an
intercept (Figure 5.9):

condition <- factor(sampleinfo$condition)

design <- model.matrix(~ 0 + condition)
design

This design matrix defines two dummy variables representing the levels of the condition
studied (1 if the condition is correct and zero otherwise). When we fit a negative binomial
generalized log-linear model described in Formula 22, two coefficient estimates will be
calculated; one for each dummy variable.

5.3.7.4 Filtering Low-Expressed Genes

Some genes may not be expressed or may not have enough reads to contribute to the dif-
ferential analysis. Therefore, it is good practice to retain only the genes that have sufficient
read counts by filtering out the genes with zero or low counts keeping only the ones with at
least one count per million (1 cpm) reads in at least two samples. The following script filters
out the genes with low abundance and adjust the library size to reflect the new change:

keep <- filterByExpr(y, design)

y <- y[keep, , keep.lib.sizes=FALSE]

As shown in Figure 5.10, after filtering, the counts slot contains only genes with sufficient
abundance and the library size in the samples slot has been adjusted. Notice the difference
in the number of genes and library size between Figures 5.10 and 5.6. The new counts slot
contains only 133 genes compared to 632 genes before filtering and the library sizes have
been adjusted to reflect the new ones.

FIGURE 5.9 Design matrix without intercept.

186   ◾    Bioinformatics

FIGURE 5.10 DGEList object after filtering out genes with low gene expression.

5.3.7.5 Normalization
After filtering the low-expressed genes, we can normalize the count data. EdgeR uses TMM
to compute normalization factor that corrects sample-specific biases. Without normaliza-
tion, if only few genes have high expression, those genes will account for a substantial
proportion of the library size for a specific sample, causing other genes to be under-rep-
resented. The normalization factor is multiplied by the library size to yield the effective
library size, which is used for normalization. The following function calculates the TMM
normalization factor:

yNorm <- calcNormFactors(y)

Notice that as shown in Figure 5.11, the normalization factor was changed for each sample.

5.3.7.6 Estimating Dispersions

The next step is to use the above normalized count data to estimate the dispersions which
will be used to estimate the parameters of the negative binomial model as discussed above.
As there are only few replicates or samples, estimation of the gene-wise dispersions based
on the count vector of the gene across replicates will not be accurate. EdgeR uses informa-
tion sharing between genes to estimate dispersion; genes of closely similar abundance will
 RNA-Seq Data Analysis   ◾    187

FIGURE 5.11 DGEList object after computing the normalization factor.

FIGURE 5.12 DGEList objects slot including the estimated dispersions.

share a dispersion. Three types of dispersions are calculated: a common estimate across
all genes, mean-variance trend dispersion using genes’ similar abundance, and gene-wise
dispersion (tagwise dispersion).
The “estimateDisp(DGEList, design)” function estimates the common, trended, and
tagwise negative binomial dispersions by using weighted likelihood empirical Bayes algo-
rithm [33]. This function requires a DGEList object with normalized counts and a design
matrix.

yNorm <- estimateDisp(yNorm, design)

names(yNorm)

The three dispersions will be estimated and stored in the DGEList object (y) as shown in
Figure 5.12. You can display the common, trended, and tagwise dispersions by using the
following (Figure 5.13):

yNorm$common.dispersion
head(yNorm$trended.dispersion)
head(yNorm$tagwise.dispersion)
188   ◾    Bioinformatics

FIGURE 5.13 Printing dispersions.

FIGURE 5.14 Biological CV plot.

We can plot the BCV against gene abundance (in log2 counts per million) using
“plotBCV(y)” function. As described above, the BCV is the square root of the dispersion
parameter in the negative binomial model.

jpeg(‘myBCVPlot.jpg’)
plotBCV(yNorm, pch=16, cex=1.2)
dev.off()

The above script plots the biological coefficient of variations (based on the three types of
dispersions) against the average abundance of each gene. As shown in Figure 5.14, the plot
shows the square root estimates of the common, trended, and tagwise negative binomial
dispersions.
For RNA-Seq count data, the negative binomial dispersions tend to be higher for genes
with very low counts and the dispersion trend tends to decrease smoothly with the abun-
dance (CPM) increase and becomes asymptotic to a constant value for genes with a larger
abundance.
 RNA-Seq Data Analysis   ◾    189

5.3.7.7 Exploring the Data

Up to this point, we have prepared the count data for fitting the negative binomial general-
ized log-linear model. However, before that step, we can explore the data by visualizing the
library size and the distribution of cpm (log2 counts per million) of each sample.

png(file=”libsizeplot.png”)
x<-barplot(yNorm$samples$lib.size/1e06,
names=colnames(yNorm),
las=2, ann=FALSE,
cex.names=0.75,
col=”lightskyblue”,
space = .5)
mtext(side = 1, text = “Samples”, line = 4)
mtext(side = 2, text = “Library size (millions)”, line = 3)
title(“Barplot of library sizes”)
dev.off()

FIGURE 5.15 Library sizes.

190   ◾    Bioinformatics

As shown in Figure 5.15, the library sizes or the sequencing depths of the six samples are
similar. This bar chart gives an idea about the distribution of the library sizes and any
potential source of bias from the library sizes.
In the normalization step, we normalized the count data to eliminate composition biases
between libraries. We can assess the TMM normalization by the MD plot (mean-difference
plot), which displays the library size-adjusted log-fold change (difference) between two
libraries against the average log-expression across these libraries (the mean). The points on
the MD plot should be centered at a line of zero log-fold change if the biases between librar-
ies were removed successfully by the normalization. The “plotMD(y, column=i)” function
creates MD plot by converting the count (y) to log2-CPM values and then creating an
artificial array by averaging all samples other than the sample specified (column=i) in the

FIGURE 5.16 Mean-difference plots.

 RNA-Seq Data Analysis   ◾    191

function. The function then creates the MD plot from the specified sample and the arti-
ficial sample. The following script creates the MD plots for the six samples (Figure 5.16):

jpeg(‘mdplots.jpg’)
par(mfrow=c(2,3))
for (i in 1:6) {
plotMD(yNorm, column=i,
xlab=”Average log CPM (all samples)”,
ylab=”log-ratio (this vs others)”)
abline(h=0, col=”red”, lty=2, lwd=2)
}
dev.off()

We can also create boxplots for the unnormalized and normalized log-CPM values to show
the expression distributions in each sample using the following script:

png(file=”logcpmboxplot.png”)
par(mfrow=c(1,2))
logcounts <- cpm(y,log=TRUE)
boxplot(logcounts, xlab=””, ylab=”Log2 counts per million”,las=2)
abline(h=median(logcounts),col=”blue”)
title(“Unnormalized logCPMs”)
logcountsNorm <- cpm(yNorm,log=TRUE)
boxplot(logcountsNorm, xlab=””, ylab=”Log2 counts per
million”,las=2)
abline(h=median(logcountsNorm),col=”blue”)
title(“Normalized logCPMs”)
dev.off()

FIGURE 5.17 The boxplots of normalized log CPM.

192   ◾    Bioinformatics

Figure 5.17 shows the boxplot of the TMM-normalized data for each sample. The black line
dividing each box represents the median of the count data, the top of the box shows the
upper quartile, and the bottom of the box shows the lower quartile. The top and bottom
whiskers show the highest and lowest count values, respectively, and the circles indicate
the outliers.
We can use multidimensional scaling (MDS) [35] for representing relationships graphi-
cally between samples in multidimensional space and showing the overall differences
between the gene expression profiles for the different samples. The MDS uses the pairwise
dissimilarity Euclidean distances between samples in terms of the leading log-fold change
(logFC) for the genes that best characterize the pair of samples. The leading logFC is cal-
culated as the root-mean-square of the top log2-fold changes between the pair of samples
(the default is 500 (top=500) logFCs). Edger uses “plotMDS” function to plot the MDS plot.
The general syntax is as follows:

plotMDS(x, top=500, gene.selection = “pairwise”, method = “logFC”,

...)

The samples are then represented graphically in two dimensions such that the distance
between points on the plot approximates their multivariate dissimilarity. The objects that
are closer together on the MDS plot are more similar than the distant ones. In our example
data, if there is a difference between the normal and tumor samples, then we can see clear
patterns in multivariate dataset.

install.packages(“RColorBrewer”)
library(RColorBrewer)
png(file=”MDSPlot.png”)
pseudoCounts <- log2(yNorm$counts + 1)
colConditions <- brewer.pal(3, “Set2”)
colConditions <- colConditions[match(sampleinfo$condition,
levels(factor(sampleinfo$condition)))]
patients <- c(8, 15, 16)[match(sampleinfo$patient,
levels(factor(sampleinfo$patient)))]
plotMDS(pseudoCounts, pch = patients, col = colConditions, xlim =
c(-2,2))
legend(“topright”, lwd = 2, col = brewer.pal(3, “Set2”)[1:2],
legend = levels(factor(sampleinfo$condition)))
legend(“bottomright”, pch = c(8, 15, 16),
legend = levels(factor(sampleinfo$patient)))
dev.off()

As shown in Figure 5.18, the patterns are very clear that the samples are clustered by the
condition; normal samples are grouped together, and tumor samples are grouped together,
which indicates that the difference between the two groups is much larger than the dif-
ferences within groups and it is likely that the between-group difference is statistically
 RNA-Seq Data Analysis   ◾    193

FIGURE 5.18 Multidimensional scaling (MDS) plot.

significant. The distance between the normal samples and the tumor sample is about 2
log2-fold change on the x-axis or 4 folds.
We can also use heatmap to cluster the most variable genes in the samples. We expect
that some samples may have similar pattern depending on the given condition (normal
or tumor). The following heatmap script will describe the relationships between samples
using hierarchical clustering:

install.packages(“gplots”)
library(“gplots”)
png(file=”heatmap1.png”)
logcountsNorm <- cpm(yNorm,log=TRUE)
var_genes <- apply(logcountsNorm, 1, var)
select_var <- names(sort(var_genes, decreasing=TRUE))[1:10]
highly_variable_lcpm <- logcountsNorm[select_var,]
mypalette <- brewer.pal(11,”RdYlBu”)
morecols <- colorRampPalette(mypalette)
col.con <- c(rep(“purple”,3),
rep(“orange”,3))[factor(sampleinfo$condition)]
heatmap.2(highly_variable_lcpm,
col=rev(morecols(50)),trace=”none”,
main=”Top 10 most variable genes”,
ColSideColors=col.con,scale=”row”,
margins=c(12,8),srtCol=45)
dev.off()
194   ◾    Bioinformatics

FIGURE 5.19 Heatmap clustering samples and top 10 variable genes.

Figure 5.19 shows that samples are clustered into tumor and normal samples based on the
profiles of the genes in these samples.

5.3.7.8 Model Fitting

Once we have computed dispersion estimates, we can use them to fit the negative binomial
generalized linear model, and then we can carry out the testing procedures for determin-
ing the differential expression. EdgeR has two functions to fit the RNA-Seq count data
to the GLMs: the “glmQLFit” function which fits the data to a quasi-likelihood negative
binomial generalized log-linear model and the “glmFit” function which fits the data to
a negative binomial generalized log-linear model. The difference between the two GLM
functions is that “glmQLFit” uses the trended negative binomial dispersion for fitting and
then estimates the quasi-likelihood dispersion from the deviance, while “glmFit” uses the
tagwise negative binomial dispersion for model fitting. You can use any one of them to
fit the count data. Run the following to fit the count data to the quasi-likelihood negative
binomial model:

fitq <- glmQLFit(yNorm, design)

names(fitq)
 RNA-Seq Data Analysis   ◾    195

FIGURE 5.20 Quasi-likelihood negative binomial model slots.

FIGURE 5.21 Quasi-likelihood dispersions plot.

Fitting the count data to a negative binomial model generates several data as shown in
Figure 5.20. For instance, the GLM coefficients are in “coefficients” slot, fitted values are in
“fitted.values” slot, and the estimated quasi-likelihood dispersions are in “dispersion” slot.
The quasi-likelihood extends the negative binomial to account for gene-specific variability
from both biological and technical aspects. We can visualize the quasi-likelihood disper-
sion with “plotQLDisp” function. The quasi-likelihood gene-wise dispersion estimates are
squeezed toward a consensus trend, which will reduce the uncertainty of the estimates
and improves testing power. The following script creates a quasi-likelihood dispersion plot
showing the raw, squeezed, and trend dispersions (Figure 5.21):

jpeg(‘qlDispplots.jpg’)
fitq <- glmQLFit(yNorm, design)
plotQLDisp(fitq, pch=16, cex=1.2)
dev.off()

Once we have fitted the count data to a GLM log-linear model, we can then be able to con-
duct the gene-wise statistical tests for a given coefficient (coef) or we can use “contrast” to
196   ◾    Bioinformatics

perform differential analysis. Since our primary goal is to study the difference between the
normal and tumor samples, we can construct a contrast using “makeContrasts” function
and then we can conduct the statistical test using the “glmQLFTest” function as follows:

my.contrasts
<- makeContrasts(conditiontumo-conditionnorm,levels=design)
fitq <- glmQLFit(yNorm, design)
qlfq<- glmQLFTest(fitq,contrast=my.contrasts)
topTags(qlfq, n=10, adjust.method=”BH”, sort.by=”PValue”,
p.value=0.05)

The “qlfq” is a DGELRT object that stores the results of a GLM-based differential expres-
sion analysis for DGE data. The “topTags” function prints the top ten (n = 10) set of the
most significantly differential genes as shown in Figure 5.22. The p-value threshold is set
to “p.value=0.05” so only genes with p-value less than 0.05 will be listed. The negative log-
fold changes (logFC) represent genes that are downregulated (down-expressed) in tumor
samples over normal sample; logCPM is the log count-per-million; F is the test statistic
for the null hypothesis that no difference in the gene expression between the normal and
tumor samples; pvalue is the significance measure (p-value < 0.05 is significant); and FDR
is the false discovery rate.
To use GLM negative binomial model instead of the quasi-negative binomial model for
the differential expression, you can use the following script:

FIGURE 5.22 The top ten significantly expressed genes.

 RNA-Seq Data Analysis   ◾    197

my.contrasts
<- makeContrasts(conditiontumo-conditionnorm,levels=design)
fitg<- glmFit(yNorm, design)
qlfg<- glmLRT(fitg,contrast=my.contrasts)
topTags(qlfg, n=10, adjust.method=”BH”, sort.by=”PValue”,
p.value=0.05)

Notice that we used contrast to conduct differential analysis to study the effect of tumor
on the gene expression. However, the design may be more complex and EdgeR provides
several ways to construct contrast based on the study design. For instance, you can add
the intercept to the design matrix so the first column on the design will be the reference.
We can also use “coef=” with “glmQLFTest” or “glmLRT” functions to indicate that the
specified coefficients are zero. In more complex design like factorial design, an ANOVA-
like analysis can be conducted. Refer to Edger users’ manual for more information about
the designs.
EdgeR also provides “glmTreat” function to conduct gene-wise statistical tests for a
given coefficient or a contrast relative to a specified fold-change threshold. For instance,
assume that we need to test the hypothesis that the gene expression in normal cells and
tumor cells is equal at a threshold of 2 log-fold change (lfc=2), assuming that a gene with a
log-fold change less than 2 is equally expressed in both normal and tumor cells.

my.contrasts
<- makeContrasts(conditiontumo-conditionnorm,levels=design)
fitq <- glmQLFit(yNorm, design)
qlfq<- glmTreat(fitq, contrast=my.contrasts, lfc=2)
topTags(qlfq, n=10, adjust.method=”BH”, sort.by=”PValue”,
p.value=0.05)

Figure 5.23 shows the top ten significantly expressed genes using a log-fold change thresh-
old (lfc=2).
Since the FDR (the rate of incorrectly rejecting null hypothesis) may be inflated in
the case of multiple testing, it can be corrected using Benjamini–Hochberg method [36],
which orders hypotheses and then rejects or accepts a hypothesis based on the Benjamini–
Hochberg critical value.
The “decideTestsDGE” function can be used to display the total number of differentially
expressed genes identified at an FDR of 5% (upregulated, downregulated, and no-signifi-
cant regulation).

my.contrasts
<- makeContrasts(conditiontumo-conditionnorm,levels=design)
fitq<-glmQLFit(yNorm, design)
qlfq<-glmQLFTest(fitq, contrast=my.contrasts)
DEGenes<-decideTestsDGE(qlfq,
adjust.method=”BH”, p.value=0.05, lfc=2)
summary(DEGenes)
198   ◾    Bioinformatics

FIGURE 5.23 Differential analysis with log-fold change threshold.

FIGURE 5.24 The number of downregulated, upregulated, and nonsignificant genes.

Figure 5.24 shows the number of downregulated, nonsignificant, and upregulated genes at
the threshold of 2 log-fold change and 0.05 significance level.
The level of the differential gene expression change can be visualized with the mean-
difference plot (MD plot):

jpeg(‘mdPlotfitted.jpg’)
my.contrasts
<- makeContrasts(conditiontumo-conditionnorm,levels=design)
fitq<-glmQLFit(yNorm, design)
qlfq<-glmQLFTest(fitq, contrast=my.contrasts)
DEGenes<-decideTestsDGE(qlfq,
adjust.method=”BH”, p.value=0.05, lfc=2)
plotMD(qlfq, status=DEGenes,
 RNA-Seq Data Analysis   ◾    199

values=c(1,0,-1),
col=c(“red”,”black”,”blue”),
legend=”topright”)
dev.off()

The MD plot in Figure 5.25 shows upregulated (red color), downregulated (blue color), and
nonsignificant gene (black color) in log-fold change versus average log cpm (abundance)
plain. The points are usually dense when all chromosomes are investigated. Remember
that for the sake of simplicity, we demonstrate the differential analysis of gene expression
in a single chromosome (chromosome 22). Both upregulated and downregulated genes are
affected by the condition studied. That suggests that they may have a role on the condition
or the effect may be due to the condition. In this example case, the upregulated and down-
regulated genes may have implications for the breast cancer. These genes can be singled out
and studied to know their roles in the condition.
We can also use the volcano plots to display the differential gene expression results.
A volcano plot is a scatterplot that shows the p-values versus the fold changes in gene
expression. It enables a quick visual identification of genes with large fold changes that
are also statistically significant. In a volcano plot, the upregulated genes are shown toward
the right, the downregulated genes are shown toward the left, and the most statistically

FIGURE 5.25 MD plot showing upregulated (red), downregulated (blue), and nonsignificant
(black).
200   ◾    Bioinformatics

significant genes are toward the top of each side. We can create a volcano plot of the RNA-
Seq results using the data of the top differentially expressed genes as shown in Figure 5.26.

#############
#volcano plot
jpeg(‘volcano.jpg’)
fitq <- glmQLFit(yNorm, design)
qlfq<- glmTreat(fitq, contrast=my.contrasts, lfc=2)
resFilt<- topTags(qlfq, n=100, adjust.method=”BH”, sort.
by=”PValue”, p.value=1)
volcanoData <- cbind(resFilt$table$logFC,
-log2(resFilt$table$PValue))
colnames(volcanoData) <- c(“logFC”, “negLogPval”)
plot(volcanoData, pch=19)
dev.off()

Once again, when all chromosomes are studied, the point distribution on the volcano plot
will be very dense. Moreover, some programs are able to color the upregulated and down-
regulated genes for better visualization.
The heatmap has been known as the most popular graphical representation for visual-
izing complex and multidimensional data. We can use the heatmap to visualize differential
expression of the genes. First, we need to use “cpm” function to convert read abundance
into log2 CPM values. The heatmap can be created for the top differentially expressed
genes. The following script creates a heatmap for the top 20 differentially expressed genes:

FIGURE 5.26 Volcano plot showing regulated genes.

 RNA-Seq Data Analysis   ◾    201

jpeg(‘heatmap2.jpg’)
my.contrasts
<- makeContrasts(conditiontumo-conditionnorm,levels=design)
fitq<-glmQLFit(yNorm, design)
qlfq<-glmQLFTest(fitq, contrast=my.contrasts)
DEGenes<-decideTestsDGE(qlfq,
adjust.method=”BH”, p.value=0.05, lfc=2)
logCPM <- cpm(yNorm, prior.count=2, log=TRUE)
rownames(logCPM) <- yNorm$genes$SYMBOL
colnames(logCPM) <- paste(yNorm$samples$group, 1:3, sep=”-”)
o <- order(qlfq$table$PValue)
logCPM <- logCPM[o[1:20],]
logCPM <- t(scale(t(logCPM)))
col.pan <- colorpanel(100, “blue”, “white”, “red”)
heatmap.2(logCPM, col=col.pan, Rowv=TRUE, scale=”none”,
trace=”none”, dendrogram=”both”, cexRow=1, cexCol=1.4,
margin=c(10,9), lhei=c(2,10), lwid=c(2,6))
dev.off()

FIGURE 5.27 Heatmap for the top 20 most differentially expressed genes across the samples.
202   ◾    Bioinformatics

As shown in Figure 5.27, the heatmap clusters genes and samples based on Euclidean dis-
tance between the expression values. As expected, samples from the same group are clus-
tered together.

5.3.7.9 Ontology and Pathways

After identifying the differentially expressed genes, the next step is to study the func-
tions of these genes, their pathways, and the conditions associated with them based on
the accumulated knowledge that already we have from previous studies and discoveries.
This knowledge is available in databases like GO [37] and KEGG [38] databases as well as
other pathway databases. Many of the genes associated with given biological processes are
differentially expressed in a given condition like diseases. By performing GO analysis, we
will be able to identify those biological processes, cellular locations, and molecular func-
tions that are impacted by the condition studied. GO attempts to capture three aspects of
the gene: (i) Biological processes (BP) that the gene may involve in, (ii) Molecular functions
(MF), and (iii) Cellular components (CC), where the biological processes and molecular
activities take place in the cell. It is important to know that GO terms aim to describe the
normal functions, processes, or locations that gene products are involved in. It does not
capture pathological processes, experimental conditions, or temporal information. Given a
set of differentially expressed genes (upregulated or downregulated genes), GO and KEGG
analyses will identify the GO terms and pathways, respectively, for each gene. EdgeR uses
“goana” function for GO analysis and “kegga” function for KEGG analysis. Both functions
require a DGELRT object and Entrez Gene identifiers (IDs) to annotate the genes. The
NCBI Entrez Gene IDs must be present as row name as we did above. Also, it is important
to specify the species studied. The following script performs GO analysis and annotates the
significantly expressed genes (downregulated and upregulated genes) with the GO terms:

FIGURE 5.28 Ontology annotation of the significantly expressed genes.

 RNA-Seq Data Analysis   ◾    203

my.contrasts
<- makeContrasts(conditiontumo-conditionnorm,levels=design)
fitq<-glmQLFit(yNorm, design)
qlfq<-glmTreat(fitq,contrast=my.contrasts, lfc=2)
go <- goana(qlfq, species=”Hs”)
topGO20<-topGO(go, sort=”up”, n=20)
write.csv(topGO20,file=”topGO20.csv”)

Figure 5.28 shows GO IDs, terms, ontology (ONT), the total number of genes annotated
with each ontology term (N), the number of genes that are significantly upregulated (up)
and downregulated (down), the gene counts of the top significant GO ordered by the sig-
nificance of the p-value, and the p-values for the upregulated (P.Up) and p-values for the
downregulated (P.Down). Since the p-values are not adjusted for multiple testing, it is rec-
ommended to ignore GO terms with p-values greater than about 10−5.
Since this exercise is based on a single chromosome (chromosome 22), we do not expect
much information as when we analyze the entire genome. In general, GO analysis tells
us about the different biological processes, their localizations in the cells, and molecular
functions based on the upregulated and downregulated genes.
In the same way, we can perform KEGG pathway analysis to identify the molecular
pathways and disease signatures (Figure 5.29).

my.contrasts
<- makeContrasts(conditiontumo-conditionnorm,levels=design)
fitq<-glmQLFit(yNorm, design)

FIGURE 5.29 KEGG annotation of the significantly expressed genes.

204   ◾    Bioinformatics

qlfq<-glmTreat(fitq,contrast=my.contrasts, lfc=2)
keg <- kegga(qlfq, species=”Hs”)
keg20<- topKEGG(keg, sort=”up”, n=20)
write.csv(keg20,file=”keg20.csv”)

5.3.8 Visualizing RNA-Seq Data

We have already discussed some methods for visualizing the RNA-Seq data. For publica-
tion purpose, we can create high-resolution colored graphics using “vidger” Bioconductor
package, which is meant to generate information-rich visualizations for the interpretation
of differential gene expression results from edgeR, DESeq2, and cuffdiff [39]. The “vidger”
package can be installed in R using the following:

if (!require(“BiocManager”, quietly = TRUE))

install.packages(“BiocManager”)
BiocManager::install(“vidger”)
Once the package has been installed, we can load it using:
library(“vidger”)

In the following, we will use “vidger” package to create plots to visualize the example RNA-
Seq data. Open R and make the “features” directory where you saved the RNA-Seq count
data file as your working directory. The vidger functions require DGEList object with
group and normalized count data. The following script will create a DGEList, “yNorm”
that can be used as input for the functions:

#Loading packages
library(edgeR)
library(“vidger”)
library(org.Hs.eg.db)
#Loading data
seqdata <- read.delim(“htcount2.txt”, stringsAsFactors=FALSE)
sampleinfo <- read.delim(“sampleinfo.txt”, stringsAsFactors=FALSE)
#Preparing data
countdata0 <- seqdata[,-(1:2)]
rownames(countdata0) <- seqdata[,1]
colnames(countdata0) <- sampleinfo$sampleid
countdata <- countdata0[rowSums(countdata0[])>0,]
group = factor(sampleinfo$condition)
#Creating DGEList object
y <- DGEList(countdata, group=group)
#Adding annotation
ENTREZID <- mapIds(org.Hs.eg.db,rownames(y),
keytype=”SYMBOL”,column=”ENTREZID”)
rownames(y$counts) <- ENTREZID
ann<-select(org.Hs.eg.db,keys=rownames(y$counts),
columns=c(“ENTREZID”,”SYMBOL”,”GENENAME”))
y$genes <- ann
 RNA-Seq Data Analysis   ◾    205

#Removing rows without Entrez ids

i <- is.na(y$genes$ENTREZID)
y <- y[!i, ]
#Creating the design matrix
condition <- factor(sampleinfo$condition)
design <- model.matrix(~ 0 + condition)
#Filtering genes with low abundance
keep <- filterByExpr(y, design)
y <- y[keep, , keep.lib.sizes=FALSE]
# Normalizing count data
yNorm <- calcNormFactors(y)
#Estimating dispersions:
yNorm <- estimateDisp(yNorm, design)

Once you have run the above script successfully without any error, then you can use the
“vidger” functions to create plots as follows.

FIGURE 5.30 Box plots showing the distribution of normal and tumor counts in CPM.
206   ◾    Bioinformatics

5.3.8.1 Visualizing Distribution with Boxplots

We can use “vsBoxPlot()” function to create box plots to study the distribution of the count
data based on condition studied. The count data can be FPKM or CPM normalized values.
The function uses the DGEList object and creates boxplots for the log10 of either FPKM or
CPM (Figure 5.30).

#Creating box plot

jpeg(‘boxplot.jpg’)
vsBoxPlot(
data = yNorm,
d.factor = NULL,
type = ‘edger’,
title = “Box plots of normal and tumor count data”,
legend = TRUE, grid = TRUE)

FIGURE 5.31 A scatter plot showing comparisons of log10 values of CPM for the two conditions.
 RNA-Seq Data Analysis   ◾    207

dev.off()

5.3.8.2 Scatter Plot

The “vsScatterPlot()” function can be used to create a scatter plot that visualizes the com-
parisons of log10 values of either FPKM or CPM measurements under the two conditions
(Figure 5.31).

#Scatter plot
jpeg(‘CPMscatterplot.jpg’)
vsScatterPlot(
x = ‘norm’,
y = ‘tumo’,

FIGURE 5.32 MA plot showing the log-fold change against mean expression.
208   ◾    Bioinformatics

data = yNorm,
type = ‘edger’,
d.factor = NULL,
title = TRUE, grid = TRUE)
dev.off()

5.3.8.3 Mean-Average Plot (MA Plot)

The “vsMAPlot()” function graphs the log-fold change versus the mean count to visualize
the variance between the two conditions in terms of gene expression values.

jpeg(‘MAPlot.jpg’)
vsMAPlot(
x = ‘norm’, y = ‘tumo’,
data = yNorm,
d.factor = NULL,

FIGURE 5.33 Volcano plot showing differential gene expression.

 RNA-Seq Data Analysis   ◾    209

type = ‘edger’,
padj = 0.05,
y.lim = NULL,
lfc = 2,
title = TRUE, legend = TRUE, grid = TRUE)
dev.off()

As shown in Figure 5.32, the blue points on the graph represent the genes with statistically
significant log-fold changes above the specified fold change indicated in the graph by the
dashed lines. The green points represent the genes with statistically significant log-fold
changes less than the specified fold change. The gray points represent the genes without sta-
tistically significant log-fold changes. Moreover, the values in parentheses for each ­legend
color show the number of genes that meet the conditions. The triangular shapes represent
values that are not displayed. The point size indicates the magnitude of the log-fold change.

5.3.8.4 Volcano Plots

As discussed above, the volcano plot is used to visualize the differential gene expression
between two conditions. The “vsVolcano()” function graphs -log10 of p-values in the y-axis
against the log2-fold changes in the x-axis (Figure 5.33).

jpeg(‘volcanoPlot2.jpg’)
vsVolcano(
x = ‘norm’, y = ‘tumo’,
data = yNorm,
d.factor = NULL,
type = ‘edger’,
padj = 0.05,
x.lim = NULL,
lfc = 2,
title = TRUE,
legend = TRUE, grid = TRUE,
data.return = FALSE)
dev.off()

5.4 SUMMARY
The PCR allows studying gene expression in very narrow scale since only expression of
few known genes can be investigated. Microarrays were emerged as the first technique for
studying the gene expression of massive number of genes. However, recently, the RNA-
Seq seems to replace microarrays because it is better at detecting gene transcripts and
gene isoforms and it also provides more accurate and sensitive results for differential gene
expression.
Studying gene expression allows researchers to identify the roles of genes in the biologi-
cal activities in the cells and their implications for diseases. This specifically provides broad
insight when massively parallel sequencing is used in the investigation of gene expression
210   ◾    Bioinformatics

across the whole genome. Genes code for proteins that determine traits and biological
functions. RNA-Seq allows investigation of the entire transcriptome through the profiling
of genes so researchers will know which sets of genes are coregulated during the conditions
studied. RNA-Seq is used to study diseases like cancers and patient’s response to therapeu-
tics and other diseases.
The first steps of the workflow of RNA-Seq are similar to the steps of the other NGS
applications as discussed in the previous chapters. The raw data is obtained in FASTQ
files, which can be single end or paired end. The raw data is subjected to quality con-
trol for the trimming of adaptors and removal of duplicate sequences and low-quality
reads. The cleaned raw data is either aligned to a reference genome or a reference tran-
scriptome. The alignment program should be chosen depending on the read length and
splicing awareness. STAR is a good choice for aligning short RNA-Seq reads. However, it
requires sufficient memory and storage space for the indexing of the reference sequence
and read alignment. The alignment information is stored in SAM/BAM files. Programs
like Samtools and PICARD can be used to manipulate and generate alignment statistics
from the BAM files. Quantification of reads aligned to each gene is essential in the RNA-
Seq data analysis because it is the only way to see the gene transcript abundance, which is
an indication for its biological activities. The reads aligned to each gene in BAM files are
counted using reads counting programs like htseq-count and featureCount. In addition to
the BAM files as inputs, these programs require also a reference annotation file in GTF for-
mat. The reads counts produced by the counting programs are stored in a tab-delimited file
containing gene symbols or gene IDs and the corresponding counts of the aligned reads
for each gene. The count file may contain a count column for each sample. Sample counts
in different files can be merged for easy processing. The count data is usually normalized
to be suitable for within-sample gene comparisons and between-sample gene comparisons.
The normalization is made to adjust for the biases arising from the differences in the gene
lengths, GC contents, and library sizes across samples. The normalization methods like
RPKM/FPKM and CPM adjust for the biases arising from the difference in gene lengths;
therefore, they can be used when comparison between the expressions of different genes
within the sample is intended. Other normalization methods like TMM and RLE adjust
for the bias arising from the difference in the library sizes across samples; therefore, they
can be used for between-sample differential expression analysis. The count data generated
by reads counting programs are further analyzed by the differential programs like edgeR,
DESeq2, and cuffdiff. As an example, we used edgeR to demonstrate the steps of the dif-
ferential analysis of the RNA-Seq data.
The differential analysis requires a metadata file that holds the information of the
samples and the study design from which the design matrix is generated. Genes with low
abundance can be filtered out since they do not contribute to the differential analysis. The
normalized count data as response variable and the design matrix as explanatory vari-
ables are used for a GLM fitting. Since the count data is over-dispersed (variance is greater
than the mean), the negative binomial distribution is assumed to fit the RNA-Seq count
data to a log-linear model to estimate the dispersions and other statistics for the differen-
tial gene expression including log-fold change, p-values, and false discovery rates (FDR).
 RNA-Seq Data Analysis   ◾    211

The expression level of a gene is described by the log-fold change of the gene transcript
abundance with respect to a reference. This reference will be a different gene in the case of
intra-sample comparison or the same gene in a different sample in the case of inter-sample
differential analysis. The significance of differential gene expression is measured by the
fold change, test statistic, p-value, adjusted p-value, and FDR. In the case of multiple com-
parison, the adjusted p-value is used. The expressed genes are usually sorted by p-values
(from the lowest to the largest) so that it will be easy to identify the top upregulated and
downregulated genes. These genes can be further analyzed by annotating them with their
GO and KEGG pathways terms to gain insight into their biological interpretation under
the conditions studied.

REFERENCES
1. Reyes JL, Kois P, Konforti BB, Konarska MM: The canonical GU dinucleotide at the 5’
splice site is recognized by p220 of the U5 snRNP within the spliceosome. Rna 1996,
2(3):213–225.
2. Wilkinson ME, Charenton C, Nagai K: RNA Splicing by the spliceosome. Annu Rev Biochem
2020, 89: 359–388.
3. Haque A, Engel J, Teichmann SA, Lönnberg T: A practical guide to single-cell RNA-sequencing
for biomedical research and clinical applications. Genome Med 2017, 9(1):75.
4. Vu TN, Deng W, Trac QT, Calza S, Hwang W, Pawitan Y: A fast detection of fusion genes from
paired-end RNA-seq data. BMC Genomics 2018, 19(1):786.
5. Dobin A, Davis CA, Schlesinger F, Drenkow J, Zaleski C, Jha S, Batut P, Chaisson M, Gingeras
TR: STAR: ultrafast universal RNA-seq aligner. Bioinformatics 2013, 29(1):15–21.
6. Hoffmann S, Otto C, Kurtz S, Sharma CM, Khaitovich P, Vogel J, Stadler PF, Hackermüller
J: Fast mapping of short sequences with mismatches, insertions and deletions using index
structures. PLoS Comput Biol 2009, 5(9):e1000502.
7. Marco-Sola S, Sammeth M, Guigó R, Ribeca P: The GEM mapper: fast, accurate and versatile
alignment by filtration. Nat Methods 2012, 9(12):1185–1188.
8. Li H, Durbin R: Fast and accurate long-read alignment with Burrows-Wheeler transform.
Bioinformatics 2010, 26(5):589–595.
9. Bushnell B: BBMap: A Fast, Accurate, Splice-Aware Aligner. In.: Lawrence Berkeley National
Lab.(LBNL), Berkeley, CA (United States); 2014.
10. Kukurba KR, Montgomery SB: RNA Sequencing and Analysis. Cold Spring Harb Protoc 2015,
2015(11):951–969.
11. Wu TD, Nacu S: Fast and SNP-tolerant detection of complex variants and splicing in short
reads. Bioinformatics 2010, 26(7):873–881.
12. Wei X, Wang X: A computational workflow to identify allele-specific expression and epigen-
etic modification in maize. Genomics Proteomics Bioinfor 2013, 11(4):247–252.
13. Grant GR, Farkas MH, Pizarro AD, Lahens NF, Schug J, Brunk BP, Stoeckert CJ, Hogenesch
JB, Pierce EA: Comparative analysis of RNA-Seq alignment algorithms and the RNA-Seq
unified mapper (RUM). Bioinformatics 2011, 27(18):2518–2528.
14. Kim D, Paggi JM, Park C, Bennett C, Salzberg SL: Graph-based genome alignment and
­genotyping with HISAT2 and HISAT-genotype. Nat Biotechnol 2019, 37(8):907–915.
15. Corchete LA, Rojas EA, Alonso-López D, De Las Rivas J, Gutiérrez NC, Burguillo FJ:
Systematic comparison and assessment of RNA-seq procedures for gene expression quantita-
tive analysis. Sci Rep 2020, 10(1): 19737.
16. Okonechnikov K, Conesa A, García-Alcalde F: Qualimap 2: advanced multi-sample quality
control for high-throughput sequencing data. Bioinformatics 2016, 32(2):292–294.
212   ◾    Bioinformatics

17. DeLuca DS, Levin JZ, Sivachenko A, Fennell T, Nazaire MD, Williams C, Reich M,
Winckler W, Getz G: RNA-SeQC: RNA-seq metrics for quality control and process optimiza-
tion. Bioinformatics 2012, 28(11):1530–1532.
18. Wang L, Wang S, Li W: RSeQC: quality control of RNA-seq experiments. Bioinformatics
2012, 28(16):2184–2185.
19. Anders S, Pyl PT, Huber W: HTSeq--a Python framework to work with high-throughput
sequencing data. Bioinformatics 2015, 31(2):166–169.
20. Liao Y, Smyth GK, Shi W: featureCounts: an efficient general purpose program for assigning
sequence reads to genomic features. Bioinformatics 2013, 30(7):923–930.
21. Robinson MD, Oshlack A: A scaling normalization method for differential expression analy-
sis of RNA-seq data. Genome Biol 2010, 11(3):R25.
22. Benjamini Y, Speed TP: Summarizing and correcting the GC content bias in high-­t hroughput
sequencing. Nucleic Acids Res 2012, 40(10):e72–e72.
23. Maza E: In Papyro Comparison of TMM (edgeR), RLE (DESeq2), and MRN Normalization
Methods for a Simple Two-Conditions-Without-Replicates RNA-Seq Experimental Design.
Front Genet 2016, 7:164.
24. Mortazavi A, Williams BA, McCue K, Schaeffer L, Wold B: Mapping and quantifying mam-
malian transcriptomes by RNA-Seq. Nat Methods 2008, 5(7):621–628.
25. Wagner GP, Kin K, Lynch VJ: Measurement of mRNA abundance using RNA-seq data:
RPKM measure is inconsistent among samples. Theory Biosci 2012, 131(4):281–285.
26. Bushel PR, Ferguson SS, Ramaiahgari SC, Paules RS, Auerbach SS: Comparison of
Normalization Methods for Analysis of TempO-Seq Targeted RNA Sequencing Data. Front
Genet 2020, 11.
27. Robinson MD, Oshlack A: A scaling normalization method for differential expression analy-
sis of RNA-seq data. Genome Biol 2010, 11(3):R25.
28. Anders S, Huber W: Differential expression analysis for sequence count data. Genome Biol
2010, 11(10):R106.
29. Bullard JH, Purdom E, Hansen KD, Dudoit S: Evaluation of statistical methods for normal-
ization and differential expression in mRNA-Seq experiments. BMC Bioinformatics 2010,
11(1):94.
30. Finotello F, Di Camillo B: Measuring differential gene expression with RNA-seq: challenges
and strategies for data analysis. Brief Funct Genomics 2015, 14(2):130–142.
31. Ver Hoef JM, Boveng PL: Quasi-Poisson vs. negative binomial regression: how should we
model overdispersed count data? Ecology 2007, 88(11):2766–2772.
32. Law CW, Zeglinski K, Dong X, Alhamdoosh M, Smyth GK, Ritchie ME: A guide to creating
design matrices for gene expression experiments. F1000Res 2020, 9: 1444.
33. Robinson MD, McCarthy DJ, Smyth GK: edgeR: a Bioconductor package for differential
expression analysis of digital gene expression data. Bioinformatics 2010, 26(1):139–140.
34. Carlson M: org.Hs.eg.db: Genome wide annotation for human. R package version 3100 2019.
35. Mead A: Review of the development of multidimensional scaling methods. J R Stat Soc D
1992, 41(1):27–39.
36. Benjamini Y, Hochberg Y: Controlling the false discovery rate: a practical and powerful
approach to multiple testing. J R Stat Soc B 1995, 57(1):289–300.
37. Consortium GO: The Gene Ontology (GO) database and informatics resource. Nucleic Acids
Res 2004, 32(suppl_1):D258–D261.
38. Kanehisa M, Araki M, Goto S, Hattori M, Hirakawa M, Itoh M, Katayama T, Kawashima S,
Okuda S, Tokimatsu T et al: KEGG for linking genomes to life and the environment. Nucleic
Acids Res 2007, 36(suppl_1):D480–D484.
39. vidger: Create rapid visualizations of RNAseq data in R [https://bioconductor.org/packages/
release/bioc/html/vidger.html]
 Chapter 6

Chromatin
Immunoprecipitation
Sequencing

6.1 INTRODUCTION TO CHROMATIN IMMUNOPRECIPITATION

In complex organisms like multicellular eukaryotes, each somatic cell in the body carries
exactly the same genome (set of DNA molecules) but gene expression varies from a type
of cells to another. A genome is formed of a set of chromosomes found in pairs in most
eukaryotic cells. These cells are called diploids. A chromosome is a DNA molecule formed
by the four basic nucleotides of the nucleobases: adenine (A), cytosine (C), guanine (G),
and thymine (T). The lengths of chromosomes of an organism vary and each chromosome
includes a different set of genes that code for different proteins. For the set of chromosomes
to fit the nucleus, each chromosome includes multiple repeated units around 147 bp in
length. These repeated units of DNA sequences wrapped around eight units (octamer) of
histone proteins forming structures called nucleosomes. A histone octamer is formed of
four different core histone units (H2A, H2B, H3, and H4). Nucleosomes are connected by
other family of histones called linker histone H1, which is highly susceptible to modifica-
tions. A chromosome and packing proteins (histones) are together called chromatin, and
when we mention a chromatin segment, we refer to a piece of raw DNA with histones.
Some environmental factors may lead to changes in chromatin causing the so-called
epigenetic changes that involve covalent modifications to chromatin structure. These epi-
genetic chromatin modifications include DNA methylation, histone modification (acetyla-
tion, phosphorylation, methylation, and ubiquitylation), and chromatin remodeling. Such
modifications affect gene expression without altering the DNA sequence and they deter-
mine the expression status of the protein-coding RNA (mRNAs) and other non-coding
RNA molecules. In general, epigenetic changes include any process that alters gene expres-
sion without changing the DNA sequence and leads to modifications that can be transmit-
ted to daughter cells. A wide variety of illnesses including cancers and many other health
DOI: 10.1201/9781003355205-6 213
214   ◾    Bioinformatics

conditions were found to be linked to epigenetic mechanisms. An epigenome consists of

all epigenetic modifications, with the genome of an organism, that regulate the activity
(expression) of the genes; these modifications can be passed down to an organism’s off-
spring [1]. Epigenomics is the study of the complete set of epigenetic modifications on the
genome of an organism. Researchers use chromatin immunoprecipitation (ChIP) to exam-
ine the interactions between epigenetic components (proteins and DNA) and profiling of
DNA methylations in their original context. ChIP are used to identify specific genes and
sequences where a protein of interest binds, across the entire genome, providing critical
information about their regulatory functions and mechanisms. The laboratory protocol
of the ChIP includes fixation of in-vivo chromatin-bound proteins with formaldehyde to
stabilize the protein on the chromatin, and then sonication or restriction enzymes are
used to cut chromatins into short random fragments usually around 200 bp. ChIP can be
performed without formaldehyde crosslinking by digesting chromatins with micrococcal
nuclease, which is an enzyme that can break chromatins into the desired fragment size
(Native ChIP). Antibodies specific for the proteins of interest are added to the short chro-
matin fragment. These enzymes form immunoprecipitated DNA–protein–antibody com-
plexes that can be separated from the non-immunoprecipitated chromatin (DNA without
protein of interest) using beads. The cross-linked formaldehyde is then removed either by
heating or by digesting the protein component of the chromatins. In the final step, only
the DNA fragments, to which proteins of interest were bound, are isolated and purified.
Up to this step, we would have the DNA fragments that were the target for the epigenetic
modification. The next step is to characterize these fragments by identifying the sequences
of the binding sites and the affected genes and that provides important information about
the binding sites of the transcription factors (TFs), function and regulation of the genes,
and the impact of the activities of the genes on the condition studied.

6.2 CHIP SEQUENCING

Researchers use different techniques to characterize the ChIP purified DNA fragments.
Northern blot, polymerase chain reaction (PCR), and microarray are some of the methods.
However, only recently, sequencing these fragments with high-throughput methods has
become the most commonly used and effective method for studying epigenetic modifica-
tions. The technique of sequencing the DNA fragments isolated from immunoprecipita-
tion is called ChIP-Seq.
Highly specific antibodies to the targeted proteins are used for the ChIP-Seq so that only
the DNA fragments affected by the epigenetic modifications are isolated. The ChIP-Seq
also requires control DNA fragments extracted from the same samples but from the DNA
regions that were not affected by the epigenetic modification (non-immunoprecipitated
chromatin fragments) or the input DNA purified from the fragmented chromatin before
the antibody incubation step. The control DNA serves as a baseline to normalize the ChIP
data. Normalization with control DNA data can reduce false positives originated from
biases that cause overrepresentation of reads. Possible source of bias includes the nonuni-
form fragmentation during sonication (sonication bias), PCR amplification which tends to
over-amplify GC-rich regions (PCR bias), sequencing bias, and mapping bias.
 Chromatin Immunoprecipitation Sequencing   ◾    215

The library preparation of the ChIP-Seq DNA fragments follows the same steps as
that of the whole genome sequencing (WGS), which includes fragmentation, end repair,
­adaptor ligation, and enrichment. The sequencing steps follow the same steps used for
DNA sequencing by the sequencing technology. The sequencing raw data includes m ­ illions
of ChIP-Seq reads.
The sequencing strategies used for ChIP-Seq are the same as the ones followed for the
WGS and RNA-Seq. The design of the ChIP-Seq experiment is usually tailored to the condi-
tion studies and that design will guide the subsequent data analysis. The raw data produced
by the sequencer are raw reads in FASTQ files. Sequencing can be single end or paired end,
short reads (e.g., Illumina) or long reads (e.g., PacBio). However, most ChIP-Seq datasets
have been generated using single-end libraries and we should be aware that some programs
do not use paired-end libraries. The run can be for a single sample or multiplexed for sev-
eral samples; the fragments of each sample in the run are with a unique barcode.

6.3 CHIP-SEQ ANALYSIS WORKFLOW

In general, the ChIP-Seq analysis workflow includes raw data acquisition, quality control,
read alignment, alignment quality control, peak calling, combining peak calls, and final
analysis (visualization, motif discovery, and annotation and functional enrichment). You
are already familiar with the first four steps, which were discussed in detail in Chapters 1
and 2. ChIP-Seq raw data can be either provided by the sequencing facility for an experi-
ment or can be downloaded from a database. In either case, you may need to reprocess the
FASTQ files (refer to Chapter 1). There are several databases from which we can down-
load ChIP-Seq raw data submitted by other researchers either as supplementary material
for their publications or may be submitted as part of a project dedicated for investigating
some conditions and the data is made public for researchers. The NCBI SRA is the most
commonly used database for these purposes and it integrates most of the other public
databases. FASTQ files are downloaded from the NCBI SRA database using SRA-toolkit,
which we used in the previous chapters. ChIP reads can be subjected to quality control
by following the steps discussed in Chapter 1. FastQC program is used to assess the qual-
ity of the reads in the files. The reads then can be preprocessed, if needed, to remove the
reads with low quality, trim the low-quality ends or adaptor sequence, and remove dupli-
cate reads and other technical reads. The step of quality control is always crucial as in
other sequencing applications to avoid misleading interpretation of results. Read mapping
is performed by aligning ChIP reads and control reads to a reference genome. The same
aligners (e.g., BWA and Bowtie) used for aligning reads from WGS can also be used for
ChIP-Seq data (refer to Chapter 2). The alignment information produced by the aligner
is stored in SAM/BAM files. Before proceeding to the next step, we can remove duplicate
reads. Duplicate reads are generated from a single read; they are identical and aligned to
the same region forming low library complexity in that region. For ChIP-Seq data, reads
are aligned upstream and downstream around the binding sites, leaving the regions of
the binding site with low sequence coverage. The read pileup density (also called signal)
around the binding sites should form bimodal enrichment patterns, with Watson strand
tags enriched upstream of binding and Crick strand tags enriched downstream. The shape
216   ◾    Bioinformatics

of the ChIP-Seq signal may vary depending on the binding protein studied. The ChIP-Seq
signal can be sharp, broad, or a mix of sharp and broad signal. The sharp signal character-
izes the binding site of the TF which binds to a specific site in the DNA sequence called
motif. Histones form broad ChIP-Seq signals because they span several nucleosomes and
may cover several nucleotides on the DNA. The RNA polymerase II (Pol II) initiates the
process of transcription by localizing on the promoter region of the gene and then it moves
during the messenger RNA transcription. Therefore, the ChIP-Seq signal of Poly II may
include both sharp and broad signals (Figure 6.1).
Peak-calling programs use sliding windows to scan the genome for these patterns to
locate the binding regions by counting both Watson and Crick tags. However, for these
kinds of tags to fit in a single window, they must be shifted to the center so that Watson tags
are shifted toward the 3′ end and Crick tags are shifted toward the 5′ end to form a peak in
the putative binding site. Peak-calling programs like MACS take advantage of the bimodal
pattern to empirically model the shifting size to precisely locate the binding sites [2] on the
genomic DNA sequences.
Peak calling is a step unique to ChIP-Seq data analysis and it aims to identify the
genomic regions occupied by the protein of interest and enriched due to the ChIP. The
abundance of the aligned reads normalized by input reads in a sliding window is the basis
of the peak calling, which is performed using statistics that determine peak significance.
The ChIP-Seq tags are usually normalized by input read (control), but some peak callers
can also call peaks without using input reads. Instead, they assume even background signal

FIGURE 6.1 Sharp signal (TF), broad signal (histones), and mixed signal (Poly II).

FIGURE 6.2 ChIP-Seq read alignment. (The peaks represent reads aligned to the reference
genome.)
 Chromatin Immunoprecipitation Sequencing   ◾    217

or other strategies. There are several peak-calling programs that use their own algorithms
to define protein-binding sites in the genome by identifying regions where sequence reads
are enriched after mapping to a reference genome. The peak caller assumes that ChIP-Seq
reads should align in a larger number to the sites of protein binding than to other regions
on the genome (Figure 6.2).
Peak-calling programs use different strategies to compute the statistical significance of
peaks in the binding sites. Some peak callers assume Poisson or negative binomial distri-
bution to model the counts of the reads and to compute the p-value for the statistical sig-
nificance of the peak with respect to the background. Since multiple windows (thousands)
are tested generating multiple p-values, the chance of making Type I error (false positive)
will increase. Some peak callers adjust the p-value based on the number of windows by
computing the false discovery rate (FDR). Other callers use the height of peaks over back-
ground without providing statistical significance metric and others use machine learning
to generate statistical metrics that allow peak calling. Sequencing depth and library com-
plexity are crucial for statistical significance of the fold enrichment.
In general, most peak aligners perform well. However, attention should be paid to the
type of binding sites that a peak caller is good for. For instance, some callers are good for
TFs (e.g., SISSRs [3]) and some are good for histone modifications, and some can handle
both (e.g., MACS [2]). Some callers do not use paired-end library although paired-end
reads can be treated as single-end reads by using either forward or reverse reads but not
both. HOMER [4] caller was developed as a tool for de novo motif identification from
peak regions. JAMM [5] requires replicated samples to improve confidence in peak call-
ing. We should also pay attention to how a caller handles broad and sharp peaks. Callers
may merge the close peaks and that may lead to the loss of some resolution. ChIP-Seq
reads originated for histone modification generate a broad peak signal that requires a large
region. However, determining a region boundary for the histone enrichment is still a chal-
lenge for the peak callers. In contrast to the histone modifications, ChIP-Seq signals of TFs
and Poly II exhibit sharp peak signal, and therefore, the peak callers suitable for TFs and
Poly II should be able to identify those narrow regions. In the following, we will discuss the
steps of ChIP-Seq workflow with a worked example.

6.3.1 Downloading the Raw Data

For practicing with ChIP data, we will download data from ENCODE project at “https://
www.encodeproject.org”. The raw data is from a ChIP-Seq experiment with an acces-
sion number ENCSR000EZL that includes three samples from HeLa-S3 cell line estab-
lished from cervical adenocarcinoma of Henrietta Lacks, who was an African American
woman died on October 4, 1951, at the age of 31, but her cells continue to have impact on
the world by making significant contributions to the scientific progress and advances in
human health. This ChIP experiment targeted DNA-directed RNA polymerase II subunit
RPB1 (encoded by POLR2A gene), which is the largest subunit of RNA polymerase II that
synthesizes all mRNA in eukaryotes. It initiates the transcription by allowing a single-
stranded DNA template strand of the promoter of a targeted gene to position itself within
its central active site. The mRNA is formed as the complementary transcript to the template
218   ◾    Bioinformatics

strand forming an initial DNA–RNA hybrid from which the new mRNA transcript is
separated. The purpose of this exercise is to investigate the promoter regions in the gene
targeted by the DNA-directed RNA polymerase II subunit RPB1 during gene transcrip-
tion. The raw data consists of four single-end FASTQ files generated by Illumina Genome
Analyzer and available at ENCODE database with the accession numbers: ENCFF000XJP,
ENCFF000XJS, and ENCFF000XKD, and the accession number of the input data (control)
is ENCSR000EZM. For the sake of keeping the files organized, we can create a project
directory called “chipseq”, and inside that directory, we can create a subdirectory called
“data” where we can download the FASTQ files as follows:

mkdir chipseq; cd chipseq; mkdir data

wget \
-O “data/ENCFF000XJP_chp1.fastq.gz” \
“https://www.encodeproject.org/files/ENCFF000XJP/@@download/
ENCFF000XJP.fastq.gz”
wget \
-O “data/ENCFF000XJS_chp2.fastq.gz” \
“https://www.encodeproject.org/files/ENCFF000XJS/@@download/
ENCFF000XJS.fastq.gz”
wget \
-O “data/ENCFF000XKD_chp3.fastq.gz” \
“https://www.encodeproject.org/files/ENCFF000XKD/@@download/
ENCFF000XKD.fastq.gz”
wget \
-O “data/ENCFF000XGP_inp0.fastq.gz” \
“https://www.encodeproject.org/files/ENCFF000XGP/@@download/
ENCFF000XGP.fastq.gz”

The four files will be downloaded into the “data” directory. The four files are
ENCFF000XJP_chp1.fastq.gz, ENCFF000XJS_chp2.fastq.gz, ENCFF000XKD_chp3.fastq.
gz, and ENCFF000XGP_inp0.fastq.gz. The latter is the FASTQ file that contains the input
or control data.

6.3.2 Quality Control

The quality control is an important step in all sequencing data analysis workflows. The
quality of the reads in the FASTQ file can be assessed by an appropriate program like
FastQC to check the read quality, technical sequences such as adaptor dimer and PCR
duplicate reads, GC-content bias, and other sequencing biases. We should try to fix any
potential problem as possible before proceeding to the mapping step. Refer to Chapter 1 for
the quality assessment metrics and the approaches to fix the potential faults.

cd data
fastqc \
ENCFF000XJP_chp1.fastq.gz \
ENCFF000XJS_chp2.fastq.gz \
 Chromatin Immunoprecipitation Sequencing   ◾    219

ENCFF000XKD_chp3.fastq.gz \
ENCFF000XGP_inp0.fastq.gz

Then, we can display the reports in an Internet browser using Firefox command as follows:

firefox \
ENCFF000XJP_chp1_fastqc.html \
ENCFF000XJS_chp2_fastqc.html \
ENCFF000XKD_chp3_fastqc.html \
ENCFF000XGP_inp0_fastqc.html
cd ..

To avoid repeating what had been discussed in Chapter 1, we will assume that the four
FASTQ files are cleaned and ready for the next step.

6.3.3 ChIP-Seq and Input Read Mapping

The second step in the ChIP-Seq data analysis, after data acquisition and quality control,
is aligning both ChIP-Seq reads and input reads to a reference genome, following the same
steps discussed in Chapter 2. You can use an aligner of your choice; however, in this exam-
ple, we will use Bowtie2. First, we need to download the FASTA file of the current human
reference genome from a reliable database such as NCBI and UCSC. We prefer the USCS
reference genome version because the chromosomes are given names instead of acces-
sion numbers. After downloading the compressed FASTQ file, it must be decompressed
using “gunzip” and indexed with “samtools faidx” command. After the file is indexed with
Samtools, we must use Bowtie2 to build an index for the reference FASTA file. The follow-
ing commands, create the directory “ref” where the human reference genome is down-
loaded, decompressed, and indexed with both Samtools and Bowtie2. Refer to Chapter 2
for Samtools and Bowtie2 installation and uses.

mkdir ref; cd ref

wget https://hgdownload.soe.ucsc.edu/goldenPath/hg19/bigZips/hg19.
fa.gz
gunzip -d hg19.fa.gz
samtools faidx hg19.fa
bowtie2-build hg19.fa hg19
cd ..

Once the above operations have been performed successfully, we can use Bowtie2 to align
both ChIP-Seq reads and control reads to the reference genome; since each file is aligned
separately, four SAM files will be produced.

mkdir bam
bowtie2 \
-p 4 \
-x ref/hg19 \
220   ◾    Bioinformatics

-U data/ENCFF000XGP_inp0.fastq.gz \
-S bam/ENCFF000XGP_inp0.sam \
2> bam/inp0.log
bowtie2 \
-p 4 \
-x ref/hg19 \
-U data/ENCFF000XJP_chp1.fastq.gz \
-S bam/ENCFF000XJP_chp1.sam \
2> bam/chp1.log
bowtie2 \
-p 4 \
-x ref/hg19 \
-U data/ENCFF000XJS_chp2.fastq.gz \
-S bam/ENCFF000XJS_chp2.sam \
2> bam/chp2.log
bowtie2 \
-p 4 \
-x ref/hg19 \
-U data/ENCFF000XKD_chp3.fastq.gz \
-S bam/ENCFF000XKD_chp3.sam \
2> bam/chp3.log

The four SAM files produced by the above commands contain the alignment information
of the reads. However, they may also include alignment information that we do not need
and removing that will make us focus only on the regions of interest and also reduce the
computational complexity. We can remove the mitochondrion read alignments, which are
defined as “chrM” in the chromosome field of the SAM file and the unidentified, ran-
dom, and haploid reads, which are defined as “chrUn”, “random”, and “*hap*”, respectively,
keeping only the reads aligned to the human chromosomes. We can use “sed” Linux com-
mand to do that and the filtered alignments are saved in new files.

cd bam
sed ‘/chrM/d;/random/d;/chrUn/d;/hap/d’ ENCFF000XGP_inp0.sam >
ENCFF000XGP_inp0_filt.sam
sed ‘/chrM/d;/random/d;/chrUn/d;/hap/d’ ENCFF000XJP_chp1.sam >
ENCFF000XJP_chp1_filt.sam
sed ‘/chrM/d;/random/d;/chrUn/d;/hap/d’ ENCFF000XJS_chp2.sam >
ENCFF000XJS_chp2_filt.sam
sed ‘/chrM/d;/random/d;/chrUn/d;/hap/d’ ENCFF000XKD_chp3.sam >
ENCFF000XKD_chp3_filt.sam

We can then convert the SAM files into BAM files using “samtools view” command.

samtools view -S -b ENCFF000XGP_inp0_filt.sam > ENCFF000XGP_inp0_

filt.bam
 Chromatin Immunoprecipitation Sequencing   ◾    221

samtools view -S -b ENCFF000XJP_chp1_filt.sam > ENCFF000XJP_chp1_

filt.bam
samtools view -S -b ENCFF000XJS_chp2_filt.sam > ENCFF000XJS_chp2_
filt.bam
samtools view -S -b ENCFF000XKD_chp3_filt.sam > ENCFF000XKD_chp3_
filt.bam

The BAM file takes less storage space. Then, we can delete the SAM file to save some storage
space if we need to. Just be careful not to delete the BAM files.
Now, we have three BAM files for the three ChIP-Seq data and one file for the control
data. Before proceeding, we need to know the number of alignments in each file and then
draw a sample of control reads approximately equal to the reads of any of the ChIP-Seq
files to be the input reads for that ChIP-Seq file. We do that to avoid library coverage bias.
The following “samtools view” commands count the alignments in each BAM file:

samtools view -c ENCFF000XGP_inp0_filt.bam

samtools view -c ENCFF000XJP_chp1_filt.bam
samtools view -c ENCFF000XJS_chp2_filt.bam
samtools view -c ENCFF000XKD_chp3_filt.bam

Table 6.1 shows the number of aligned reads in each BAM file and the factor, which is the
read count of a ChIP-Seq file divided by the read count of the control file. This fraction is
used to sample input reads from the control file for that ChIP-Seq file.
The following commands store the counts in bash variables and then use “samtools
view” command to draw a subsample of reads from the control file and store them in a
separate control file for that ChIP-Seq file. The “-b” option is to output a BAM file and “-s”
option is to draw a subsample from the file.

inpc=$(samtools view -c ENCFF000XGP_inp0_filt.bam)

chp1=$(samtools view -c ENCFF000XJP_chp1_filt.bam)
fact1=$(echo “scale=6; $chp1/$inpc” | bc)
samtools view -b -s $fact1 ENCFF000XGP_inp0_filt.bam >
ENCFF000XGP_inp0_filt_inp1.bam

TABLE 6.1 Read Count in Each BAM File, the Fraction for Sampling Reads from the Control BAM file, and
Number of Reads in the Control File for Each ChIP-Seq File
Sample Read Count Sampling Factor Control Read Count
ENCFF000XGP_inp0_filt.bam 30,923,163 N/A N/A
ENCFF000XJP_chp1_filt.bam 8,942,010 0.289168673 8,941,151
ENCFF000XJS_chp2_filt.bam 12,748,871 0.412275775 12,744,729
ENCFF000XKD_chp3_filt.bam 13,217,349 0.427425519 13,212,672
222   ◾    Bioinformatics

chp2=$(samtools view -c ENCFF000XJS_chp2_filt.bam)

fact2=$(echo “scale=6; $chp2/$inpc” | bc)
samtools view -b -s $fact2 ENCFF000XGP_inp0_filt.bam >
ENCFF000XGP_inp0_filt_inp2.bam

chp3=$(samtools view -c ENCFF000XKD_chp3_filt.bam)

fact3=$(echo “scale=6; $chp3/$inpc” | bc)
samtools view -b -s $fact3 ENCFF000XGP_inp0_filt.bam >
ENCFF000XGP_inp0_filt_inp3.bam

You can then double check the read count in the new control files for the ChIP-Seq files.
The read counts for control files are shown in Table 6.1.

samtools view -c ENCFF000XGP_inp0_filt_inp1.bam

samtools view -c ENCFF000XGP_inp0_filt_inp2.bam
samtools view -c ENCFF000XGP_inp0_filt_inp3.bam

Up to this point, we have three ChIP-Seq BAM files and three control BAM files, one for
each ChIP-Seq file. Before proceeding to the next step, you may decide to view the align-
ment in the BAM file with the “samtools tview” command or any other BAM viewing pro-
gram. When we use “samtools tview”, the “-p” option is used to specify a specific position.
To view a BAM file, you need to sort the alignments by coordinate and then index it. As an
example, we will view only one file (Figure 6.3).

samtools sort ENCFF000XJP_chp1_filt.bam -o

ENCFF000XJP_chp1_filt_so.bam
samtools index ENCFF000XJP_chp1_filt_so.bam
samtools tview \

FIGURE 6.3 Visualizing reads aligned to the reference genome using “samtools tview” command.
 Chromatin Immunoprecipitation Sequencing   ◾    223

ENCFF000XJP_chp1_filt_so.bam \
-p chr17:56084561-56084661 \
../ref/hg19.fa

The “samtools tview” provides a quick way to visualize the alignments in a BAM file and
to inspect the binding site in a specific location. However, when you need to view a specific
gene, you may need to inspect the BAM file to see how the chromosomes are named. You
can also find gene coordinates in the reference genome from the NCBI Gene database.
Viewing a BAM file is not necessary unless you need to inspect the alignments.
The next step is to use one of the peak-calling programs or shortly a peak caller that uses
a ChIP-Seq BAM file and a control BAM file as inputs to perform the process of peak call-
ing as discussed above. There are several peak callers available as open source, but we will
use the oldest and the most commonly used one, MACS3.

6.3.4 ChIP-Seq Peak Calling with MACS3

MACS (Model-based Analysis of ChIP-Seq) [6, 7] is one of the popular peak callers,
­suitable for identifying TF binding sites. MACS utilizes library complexity to evaluate
the significance of enriched ChIP peaks by modeling the distribution of reads along the
genome using Poisson distribution and then it estimates the FDR for each detected peak
empirically. MACS can be used for peak calling using only ChIP-Seq data without pro-
viding control or with a control sample (input reads). Sliding window of different sizes is
used for finding peaks by calculating the expected number of reads aligned to that region.
MACS can perform removal of redundant reads and read shifting. It generates a number
of statistics for peak calling.
Basically, MACS separates upstream and downstream aligned reads and aligns them
by the midpoint between their peaks. Assume that the distance between the two peaks is
d bp, the all reads will be shifted by d/2 bp toward 3′ ends to cover the most likely DNA
protein-binding sites. MACS models read distribution along the genome by Poisson distri-
bution to perform peak calling and to provide p-value. After read shifting, the algorithm
uses a sliding window of 2d bp to identify candidate peaks with a significant enrichment.
It also merges the overlapping peaks and extends reads d bases from their center. The posi-
tion with the highest fragment pileup is called the summit, which is identified as the exact
location of the protein-binding site in the genome. A Poisson distribution parameter λ is
estimated for each peak and peaks with p-values less than a threshold will be called. The
ratio between the ChIP-Seq read count and λ is estimated as the fold enrichment.
For installation instruction, visit “https://github.com/macs3-project/MACS”. On Linux,
you can install MACS3 using “pip install macs3” and then you can test it by running
“macs3 callpeak -h” to display the help. In the following, we will use MACS3 to analyze the
three ChIP-Seq BAM files while using the corresponding control BAM file. Each “macs3”
command requires a ChIP-Seq BAM file “-t” and a control BAM file “-c”. The “-f” option
specifies the file format of the input file, “-g” option specifies the effective genome size,
“-n” option specifies the string that is prefixed to the output files, “-q” option specifies the
FDR threshold, “--bdg” option is to generate a bedGraph file, which is a format that allows
224   ◾    Bioinformatics

the display of continuous-valued data in a track in genome browsers like UCSC Genome
Browser or Integrated Genome Browser (IGB), and “--outdir” specifies the directory where
the output files are saved.

cd ..
mkdir macs3output
macs3 callpeak \
-t bam/ENCFF000XJP_chp1_filt.bam \
-c bam/ENCFF000XGP_inp0_filt_inp1.bam \
-f BAM \
-g hs \
-n chip1 \
-q 0.05 \
--bdg \
--outdir macs3output
macs3 callpeak \
-t bam/ENCFF000XJS_chp2_filt.bam \
-c bam/ENCFF000XGP_inp0_filt_inp2.bam \
-f BAM \
-g hs \
-n chip2 \
-q 0.05 \
--bdg \
--outdir macs3output
macs3 callpeak \
-t bam/ENCFF000XKD_chp3_filt.bam \
-c bam/ENCFF000XGP_inp0_filt_inp3.bam \
-f BAM \
-g hs \
-n chip3 \
-q 0.05 \
--bdg \
--outdir macs3output

Several output files for each ChIP-Seq data have been saved in the specified output ­directory
as shown in Figure 6.4. The description of these MACS3 output files is as follows:
Excel file: The “*_peaks.xls” is an Excel file, which is the most important file that con-
tains header information in the beginning of the file and information about the peaks
called by MACS3 in nine columns. These columns are Chromosome name (chr), start posi-
tion of peak (start), end position of peak (end), length of peak region (length), absolute peak
summit position (abs_summit), pileup height at peak summit (pileup), -log10(pvalue) for
the peak summit (-log10(pvalue)), fold enrichment for this peak summit against random
Poisson distribution with local lambda (fold_enrichment), -log10(qvalue) at peak summit
(-log10(qvalue)), and name.
BedGraph format: The “*_treat_pileup.bdg” contains the peak enrichment signal and
“*_control_lambda.bdg” contains local biases for each location in the reference genome.
 Chromatin Immunoprecipitation Sequencing   ◾    225

FIGURE 6.4 MACS3 output files for the three ChIP-Seq data.

These two files are the largest and they are in bedGraph format that can be visualized in the
UCSC browser. The bedGraph format is a format developed to display genomic informa-
tion in a track on the genomic browser. It consists of four tab-separated columns: chromo-
some, start, end, and value. The chromosome coordinates are 0-based. The positions (start
and end) are listed in ascending order. The data displayed in the track are the values in the
value column and they can be integer or real, positive or negative.
BED format: The “*_summits.bed” file is in BED format. The BED (Browser Extensible
Data) format defines the data lines that are viewed in an annotation track of a genomic
browser. The BED file contains three required fields (chromosome, start, and end) and nine
additional optional fields (name, score, strand, thick start, thick end, RGB (an RGB color
value), block count, block size, and block start).
The “summits.bed” file contains the summit locations for every peak. The fifth column
in this file is the same as what is in the narrowPeak file. This file can be used for finding
motifs at the binding sites.
R script file: The “*_model.r” file is an R script to produce a PDF file containing peak
model plot and cross-correlation plot. The pdf files for our ChIP-Seq data are generated
using the “Rscript” command as follows:

Rscript chip1_model.r
Rscript chip2_model.r
Rscript chip3_model.r

BED6+4 format: The “*_peaks.narrowPeak” file is in BED6+4 format, which stores infor-
mation about signal enrichment of the called peaks based on pooled, normalized read
counts. This format consists of ten tab-separated columns: chromosome, start, end, name
(region name), score (peak density score from 0 to 1000) based on the signal value, strand
(+/-), signal value (average enrichment), p-value (int(-10*log10pvalue), FDR or q-value
226   ◾    Bioinformatics

(int(-10*log10qvalue), and peak (o-based offset from start). Either p-value or q-value is
present in “*_peaks.narrowPeak” file depending on which threshold option is used: “-p”
or “-q”.

6.3.5 Visualizing ChIP-Seq Enrichment in Genome Browser

In this step, we will visualize the enrichment signals contained in the bedGraph files
­(*control_lambda.bdg, and *treat_pileup.bdg) and peak densities in the BED file (*peaks.
narrowPeak) in a genome browser like a UCSC Genome Browser or IGB. There are three
files for each ChIP-Seq sample. First, we need to convert bedGraph format into BigWig for-
mat, which is an indexed binary file created from bedGraph file using “bedGraph­ToBigWig”
utility for fast display. We can download the conversion utilities from the UCSC using the
following commands (download these programs in the directory where the MACS3 output
files are found for easy use):

rsync -aP \
rsync://hgdownload.soe.ucsc.edu/genome/admin/exe/linux.x86_64/
bedGraphToBigWig ./
rsync -aP \
rsync://hgdownload.soe.ucsc.edu/genome/admin/exe/linux.x86_64/
bigWigToWig ./
rsync -aP \
rsync://hgdownload.soe.ucsc.edu/genome/admin/exe/linux.x86_64/
fetchChromSizes ./

The “bedGraphToBigWig” utility converts a file from the bedGraph format to the BigWig
format. The “bigWigToWig” utility converts a file from the BigWig format to the Wig for-
mat. The “fetchChromSizes” utility creates a text file (from a reference genome) containing
the chromosome names and sizes in bases. For more information about these commands
and file format, refer to the UCSC website.
Before proceeding, we need to create a text file containing the name of the human chro-
mosomes and their sizes using “fetchChromSizes”.

fetchChromSizes hg19 > hg19.chrom.sizes

The file “hg19.chrom.sizes” will be created in the working directory; you can view it using
“less hg19.chrom.sizes”. This file consists of two columns: chromosome name and chromo-
some size in bases.
Now, we can convert the bedGraph files to BigWig format and then from BigWig format
to Wig format using the following commands while the working directory is the one where
the MACS3 output files are found:

mkdir vis
#Convert *control_lambda.bdg from bedGraph to BigWig
bedGraphToBigWig \
 Chromatin Immunoprecipitation Sequencing   ◾    227

chip1_control_lambda.bdg hg19.chrom.sizes \
vis/chip1_control_lambda.bw
bedGraphToBigWig \
chip2_control_lambda.bdg hg19.chrom.sizes \
vis/chip2_control_lambda.bw
bedGraphToBigWig \
chip3_control_lambda.bdg hg19.chrom.sizes \
vis/chip3_control_lambda.bw

#Convert *treat_pileup.bdg from bedGraph to BigWig

bedGraphToBigWig \
chip1_treat_pileup.bdg \
hg19.chrom.sizes \
vis/chip1_treat_pileup.bw
bedGraphToBigWig \
chip2_treat_pileup.bdg \
hg19.chrom.sizes \
vis/chip2_treat_pileup.bw
bedGraphToBigWig \
chip3_treat_pileup.bdg \
hg19.chrom.sizes \
vis/chip3_treat_pileup.bw

#Convert control_lambda.bw from Bigwig to wig

bigWigToWig \
vis/chip1_control_lambda.bw \
vis/chip1_control_lambda.wig
bigWigToWig \
vis/chip2_control_lambda.bw \
vis/chip2_control_lambda.wig
bigWigToWig \
vis/chip3_control_lambda.bw \
vis/chip3_control_lambda.wig

#Convert chip1_treat_pileup.bw from Bigwig to wig

bigWigToWig \
vis/chip1_treat_pileup.bw \
vis/chip1_treat_pileup.wig
bigWigToWig \
vis/chip2_treat_pileup.bw \
vis/chip2_treat_pileup.wig
bigWigToWig \
vis/chip3_treat_pileup.bw \
vis/chip3_treat_pileup.wig

We need also to modify the BED file “peaks.narrowPeak” by keeping only columns 1–4.
228   ◾    Bioinformatics

cut -f1,2,3,4 chip1_peaks.narrowPeak > vis/chip1_peaks.bed

cut -f1,2,3,4 chip2_peaks.narrowPeak > vis/chip2_peaks.bed
cut -f1,2,3,4 chip2_peaks.narrowPeak > vis/chip3_peaks.bed

The converted files were saved in “macs3output/vis” directory as shown in Figure 6.5. Those
files can be visualized in a genome browser. If you are working with a Linux with a graphi-
cal desktop, that works for the next step; otherwise, you can copy “macs3output” directory
including “vis” to a Windows or Mac desktop using FileZilla, which is an open-source FTP
application for file transfer, available at “https://filezilla-project.org/”. The “macs3output”
directory contains all MACS3 output files that we will use in later analyses and “vis” sub-
directory contains files for visualization in a genome browser.
In the next step, we will use a genome browser. In this exercise, we will use the IGB [8],
which is a standalone software available for Linux, Windows, and Mac OS X. It can be
downloaded from “https://www.bioviz.org/” and installed on a local computer. Once it has
been installed, open it, click “H. sapiens”, open the directory where the Wig and BED files
are stored, and for each ChIP-Seq sample, use the mouse to drag “*control_lambda.wig”,
“*treat_pileup.wig”, and “*peaks.bed” to the viewer just above “RefSeq curated (+)” track
and every time click “Load Data” button on the top right. You can also change the color of
the tracks of each sample by clicking the right button on the track and select “Customize”
from the popup menu and then choose a color (Figure 6.6).
As shown in Figure 6.6, each ChIP-Seq sample has three tracks: control (input data),
treated (ChIP-Seq), and the peaks. It is clear that the signal of input data (control) is flat but
the ChIP-Seq signals are with peaks. The peak track shows the peak position. Compare the
control signal of each sample to its corresponding treated (ChIP-Seq) signal. You can move
along the genome by using grab tool (hand icon), or zoom in and out using the horizontal
zoom slider on the top. You can also use the mouse to move left or right. The IGB comes
with search tools that can be used to search for a gene or to navigate to a specific position
in a chromosome.
Figure 6.7 shows a typical mixed (sharp and broad) signal of Poly II for the three sam-
ples after zooming in.

FIGURE 6.5 Wig and BED files ready to be visualized in a genome browser.
 Chromatin Immunoprecipitation Sequencing   ◾    229

FIGURE 6.6 The IGB displaying the tracks of the three ChIP-Seq samples.

FIGURE 6.7 The IGB displaying POLR2A TF binding site.

6.3.6 Visualizing Peaks Distribution

In this time, we will use peaks files “*peaks.narrowPeak” to visualize the peak enrich-
ment distribution using R Bioconductor packages. You can use R in any platform (Linux,
Windows, or Mac OS). Follow the instructions at “https://cran.r-project.org/” to download
230   ◾    Bioinformatics

and install R. You can also download and install RStudio by following the instructions
at “https://www.rstudio.com/products/rstudio/download/”. Once you have R installed on
your computer, run R, and on R prompt, run the following to install the Bioconductor
packages required for the remaining ChIP-Seq data analysis:

if (!require(“BiocManager”, quietly = TRUE))

install.packages(“BiocManager”)
BiocManager::install(“clusterProfiler”)
BiocManager::install(“ChIPseeker”)
BiocManager::install(“TxDb.Hsapiens.UCSC.hg19.knownGene”)
BiocManager::install(“EnsDb.Hsapiens.v75”)
BiocManager::install(“clusterProfiler”)
BiocManager::install(“AnnotationDbi”)
BiocManager::install(“org.Hs.eg.db”)
nstall.packages(“dplyr”)

After installing the packages, load them as follows:

library(clusterProfiler)
library(ChIPseeker)
library(TxDb.Hsapiens.UCSC.hg19.knownGene)
library(EnsDb.Hsapiens.v75)
library(AnnotationDbi)
library(org.Hs.eg.db)
library(“dplyr”)

Now, we are ready to finish the remaining analysis in R. Copy “chip1_peaks.narrowPeak”,

“chip2_peaks.narrowPeak”, and “chip3_peaks.narrowPeak” files produced by MACS3
above to a directory that you can browse from inside R. Use R to choose your working
directory where those three files are copied. In the following, we can use ChIPseeker[9]
package functions to create different plots.

6.3.6.1 ChIP-Seq Peaks’ Coverage Plot

The “covplot()” function is used to create a plot that shows peak distribution over the
whole genome. This function calculates the coverage of peak regions over chromosomes
or regions of chromosomes and generates a peaks’ coverage plot. The function requires
the peak data as Granges (genomic ranges) object (see ChIPseeker documentations). The
following codes load each file as a data frame, add column names, create a Granges object,
and finally create the ChIP-Seq peaks’ coverage plots for each sample:

peaks1 <- read.table(“chip1_peaks.narrowPeak”,header=FALSE)

colnames <- c(“chrom”, “start”, “end”, “name”, “score”,”strand”,
“signal”, “pvalue”, “qvalue”, “peak”)
colnames(peaks1) <- colnames
 Chromatin Immunoprecipitation Sequencing   ◾    231

peaks2<- read.table(“chip2_peaks.narrowPeak”,header=FALSE)
colnames(peaks2) <- colnames
colnames(peaks2)
peaks3<- read.table(“chip3_peaks.narrowPeak”,header=FALSE)
colnames(peaks3) <- colnames

#head(peaks1)
peaks1Ranges<- GRanges(seqnames=peaks1$chrom,
ranges=IRanges(peaks1$start,peaks1$end),
peaks1$name,
peaks1$score,
strand=NULL,
peaks1$signal,
peaks1$pvalue,
peaks1$qvalue,
peaks1$peak)
covplot(peaks1Ranges, weightCol=”peaks1$peak”)

peaks2Ranges<- GRanges(seqnames=peaks2$chrom,
ranges=IRanges(peaks2$start,peaks2$end),
peaks2$name,
peaks2$score,
strand=NULL,
peaks2$signal,
peaks2$pvalue,
peaks2$qvalue,
peaks2$peak)
covplot(peaks2Ranges, weightCol=”peaks2$peak”)

peaks3Ranges<- GRanges(seqnames=peaks3$chrom,
ranges=IRanges(peaks3$start,peaks3$end),
peaks3$name,
peaks3$score,
strand=NULL,
peaks3$signal,
peaks3$pvalue,
peaks3$qvalue,
peaks3$peak)
covplot(peaks3Ranges, weightCol=”peaks3$peak”)

Three ChIP-Seq peak coverage plots will be created but we will display only a single plot
to save space. Figure 6.8 shows the coverage plot for the first sample (chip1); it shows the
distribution of the peaks of each human chromosome.
Rather than the entire genome, “covplot()” can also display the coverage in a single
chromosome, a group of chromosome, a specific region of a chromosome, or it can be
232   ◾    Bioinformatics

FIGURE 6.8 ChIP-Seq peak coverage plot.

FIGURE 6.9 ChIP-Seq peak coverage plot comparing between a region in Chromosomes 1 and 2.

used to compare between chromosomes. The following codes display peak coverage in the
region between 1.0e8 and 1.5e8 bp for chromosome 1 and 2 (Figure 6.9):

covplot(peaks1Ranges, weightCol=”peaks1$peak”,
chrs=c(“chr1”, “chr2”), xlim=c(1.0e8, 1.5e8))
 Chromatin Immunoprecipitation Sequencing   ◾    233

6.3.6.2 Distribution of Peaks in Transcription Start Site (TSS) Regions

The heatmap can be used to show the peak distribution on the transcription start site (TSS)
region as shown in Figure 6.10.

txdb <- TxDb.Hsapiens.UCSC.hg19.knownGene

promoter <- getPromoters(TxDb=txdb, upstream=2000,
downstream=2000)
tagMatrix <- getTagMatrix(peaks1Ranges, windows=promoter)
tagHeatmap(tagMatrix, xlim=c(-2000, 2000), color=”blue”)

The distribution of peaks in the TSS region can also be visualized with the line plot that
profiles the average peaks in the TSS region.

plotAvgProf(tagMatrix, xlim=c(-2000, 2000),

xlab=”Genomic Region (5’->3’)”,
ylab = “Read Count Frequency”)

We can notice that as shown in Figure 6.11, the peaks are normally distributed (bell-
shaped) with the mean in the TSS region.

FIGURE 6.10 ChIP-Seq peak profiling in the TSS regions of the genes.
234   ◾    Bioinformatics

FIGURE 6.11 Average profile of ChIP-Seq peaks.

A line plot provided with confidence interval can also be created. The confidence inter-
val is estimated by resampling the peaks several times without replacement and computing
the average and variance each time.

plotAvgProf(tagMatrix, xlim=c(-2000, 2000),

conf = 0.95, resample = 1000)

6.3.6.3 Profile of Peaks along Gene Regions

The average peak profile in the different gene regions across genes will give a clear idea
about the site of the protein–DNA interaction. The “plotPeakProf2()” creates a profile plot
for the peak signal from the TSSs to the transcription termination sites (TTSs).

#Profile of ChIP peaks binding to body regions

plotPeakProf2(peak = peaks1Ranges, upstream = rel(0.2), downstream
= rel(0.2),
conf = 0.95, by = “gene”, type = “body”, nbin = 800, TxDb = txdb,

weightCol = “peaks1$peak”,ignore_strand = F)
 Chromatin Immunoprecipitation Sequencing   ◾    235

FIGURE 6.12 Average profile of ChIP-Seq peaks across.

Figure 6.12 shows that Poly II localization is centered in the TSS where most peaks are observed.

6.3.7 Peak Annotation

We will continue using R to perform annotation of the peaks called with MACS3 program
that stored the peak information in “*peaks.narrowPeak” files. The peaks represent the
most likely locations of protein–DNA interaction in the genome (the content of “*peaks.­
narrowPeak” files is discussed above). The main goal of ChIP-Seq data analysis is to inves-
tigate the biological implications of the epigenomic changes like genomic binding sites of
proteins such as TFs, histones, and Poly II. Annotation of the protein–DNA interaction
sites and functions will provide important information about the biological implications.
Peak annotation is the process of associating the sites identified by the peaks to the genes
and region of the genes affected by the epigenetic change. Most interactions like TFs, ini-
tial localization of Poly II and histones occur in the cis-regulatory site of the gene which
is close to TSS and it includes a promoter, an enhancer, a silencer, insulators, etc., which
play crucial roles in controlling gene expressions in specific cell types, conditions, and
developmental stages. An annotation program annotates ChIP-Seq peaks by associating
these peaks to the closest TSS of a gene, either upstream or downstream. A cis-regulatory
region can also be in distance from the TSS or between the TSSs of two different genes.
We will continue using R Bioconductor packages and “*peaks.narrowPeak” files as
inputs for annotation. The following codes create a list of the sample file names and a label
for each sample as “chip1”, “chip2”, and “chip3” respectively:

bedfiles <- list.files(“vis”, pattern= “.bed”, full.names=T)

bedfiles <- as.list(bedfiles)
names(bedfiles) <- c(“chip1”, “chip2”, “chip3”)

Then, we can assign the database of the known human genes to a variable so that we can
use the annotation information and associate them to the peaks.
236   ◾    Bioinformatics

txdb <- TxDb.Hsapiens.UCSC.hg19.knownGene

The above annotation data package is for human (Homo sapiens) data from UCSC build
hg19 based on the knownGene Track. The available annotation packages for the genomes of
other organisms are available at “http://bioconductor.org/packages/3.5/data/annotation/”.
In the next step, we will use “annotatePeak()” function from ChIPseeker package[9] to
annotate the peaks by associating them to the nearest genes. This function also provides
the option “tssRegion=” that allows us to specify a max distance from the TSS in which the
peaks can be associated to the gene.

annotated_peaks <- lapply(bedfiles,

annotatePeak,
TxDb=txdb,
tssRegion=c(-1000, 1000), verbose=FALSE)
annotated_peaks

The above R codes apply the “annotatePeak()” function to annotate the peaks in the peak
signal files. The peak region was also set to any distance in the range (−1000, 1000) from
the TSS of the gene. Figure 6.13 shows the annotation summaries for the three samples.
The summary includes the number of peaks annotated on the top and then the peak anno-
tation frequencies based on the genomic features (gene regions). We can notice that the
maximum frequencies are in the promoter region, which in this case is an indication for
the transcriptional activity in the gene associated to the peaks.
The ChIPseeker package provides several functions to visualize the annotated peaks. The
“plotAnnoBar()” creates a bar chart for the peak representation in the different genomic
regions (features).

plotAnnoBar(annotated_peaks)

Figure 6.14 shows a bar plot that depicts peak enrichment representation in the different
genomic regions of the genes. We can notice that most peaks are centered in the promoter
regions. This may look different if the ChIP-Seq is for TFs or histone marks.
Distribution of peaks relative to TSS:
The sites of TF binding and Poly II localization are found in the promoter regions of the
genes. Thus, distribution of peaks around TSS will give an idea about the activity of the

FIGURE 6.13 Annotation summaries for three ChIP-Seq samples.

 Chromatin Immunoprecipitation Sequencing   ◾    237

FIGURE 6.14 Bar chart of the genomic feature representation.

protein studied. The “plotDistToTSS()” function creates a plot showing the distribution of
the peaks relative to the TSS.

plotDistToTSS(annotated_peaks,
title=”Distribution of Poly II relative to TSS”)

As shown in Figure 6.15, most interaction sites are in the region of 0–1 kb from the TSS of
the genes.

6.3.7.1 Writing Annotations to Files

The annotation information of the peaks can be written to a file so that it can be used for
further analysis. The annotation file will include chromosome name, start (peak start posi-
tion), end (peak end position), width (number of bases), strand, peak name, annotation
(gene region), gene chromosome, gene start, gene end, gene length (bases), gene strand,
gene Id, transcript Id, distance to TSS (in bases), and gene name. This information is usu-
ally used by researchers to investigate individual genes. The following R codes write the
annotations for the three ChIP-Seq samples to three text files: “Chip1_peak_annotation.
txt”, “Chip2_peak_annotation.txt”, and “Chip3_peak_annotation.txt”.

#Write Chip1 annotation to a file

#separet Chip1 annotation in a data frame
chip1_annot <- data.frame(annotated_peaks[[“chip1”]]@anno)
238   ◾    Bioinformatics

FIGURE 6.15 Distribution of Poly II relative to the TSS.

#get the NCBI Entrez IDs

entrez1 <- chip1_annot$geneId
# Obtain the gene symbols for the set of Entrez IDs from the
database
annotations_edb1 <- AnnotationDbi::select(EnsDb.Hsapiens.v75,
keys = entrez1,
columns = c(“GENENAME”),
keytype = “ENTREZID”)
#Convert IDs to character type to merge
annotations_edb1$ENTREZID
<- as.character(annotations_edb1$ENTREZID)
# Write Chip1 annotation to file
chip1_annot %>% left_join(annotations_edb1,
by=c(“geneId”=”ENTREZID”)) %>%
write.table(file=”Chip1_peak_annotation.txt”,
sep=”\t”, quote=F, row.names=F)

# Write Chip2 annotation to a file

chip2_annot <- data.frame(annotated_peaks[[“chip2”]]@anno)
entrez2 <- chip2_annot$geneId
annotations_edb2 <- AnnotationDbi::select(EnsDb.Hsapiens.v75,
keys = entrez2,
columns = c(“GENENAME”),
keytype = “ENTREZID”)
 Chromatin Immunoprecipitation Sequencing   ◾    239

annotations_edb2$ENTREZID
<- as.character(annotations_edb2$ENTREZID)
chip2_annot %>% left_join(annotations_edb2,
by=c(“geneId”=”ENTREZID”)) %>%
write.table(file=”Chip2_peak_annotation.txt”,
sep=”\t”, quote=F, row.names=F)

# Write Chip3 annotation to a file

chip3_annot <- data.frame(annotated_peaks[[“chip3”]]@anno)
entrez3 <- chip3_annot$geneId
annotations_edb3 <- AnnotationDbi::select(EnsDb.Hsapiens.v75,
keys = entrez3,
columns = c(“GENENAME”),
keytype = “ENTREZID”)
annotations_edb3$ENTREZID
<- as.character(annotations_edb3$ENTREZID)
chip3_annot %>% left_join(annotations_edb3,
by=c(“geneId”=”ENTREZID”)) %>%
write.table(file=”Chip3_peak_annotation.txt”,
sep=”\t”, quote=F, row.names=F)

Those files will be saved in the working directory. Open these files in Excel to study their
contents.

6.3.8 ChIP-Seq Functional Analysis

After peak annotation and functional enrichment, the next step in ChIP-Seq data analysis
is the identification of the biological implications of the expression of the genes associated
with the sites. For this purpose, we will use knowledge from Gene Ontology (GO) and
KEGG pathway and other pathway databases as we did in the RNA-Seq data analysis. The
GO enrichment analysis is performed using “enrichGO()”, which is one of the “clusterPro-
filer” package [10]. This function will return the enrichment GO categories based on the
FDR threshold. The following codes perform GO analysis for each sample and then write
the results into a file and create a dot plot showing a specified number of the top GO terms.
The dot size on the plot represents the number of the genes related to GO term divided by
the total number of significant genes and the size of the circle describes the significance in
adjusted p-value (p-adjusted).

ego1 <- enrichGO(gene = entrez1,

keyType = “ENTREZID”,
OrgDb = org.Hs.eg.db,
ont = “BP”,
pAdjustMethod = “BH”,
qvalueCutoff = 0.05,
readable = TRUE)
ego2 <- enrichGO(gene = entrez2,
240   ◾    Bioinformatics

keyType = “ENTREZID”,
OrgDb = org.Hs.eg.db,
ont = “BP”,
pAdjustMethod = “BH”,
qvalueCutoff = 0.05,
readable = TRUE)
ego3 <- enrichGO(gene = entrez3,
keyType = “ENTREZID”,
OrgDb = org.Hs.eg.db,
ont = “BP”,
pAdjustMethod = “BH”,
qvalueCutoff = 0.05,
readable = TRUE)
#GO output
# Chip1
cluster_summary1 <- data.frame(ego1)
write.csv(cluster_summary1, “chip1_GO.csv”)
# Dotplot visualization
dotplot(ego1, showCategory=10)
# Chip2
cluster_summary2 <- data.frame(ego2)
write.csv(cluster_summary2, “chip2_GO.csv”)
# Dotplot visualization
dotplot(ego2, showCategory=10)
# Chip2
cluster_summary3 <- data.frame(ego3)
write.csv(cluster_summary3, “chip3_GO.csv”)
# Dotplot visualization
dotplot(ego3, showCategory=10)

FIGURE 6.16 GO dot plot showing the top 10 genes.

 Chromatin Immunoprecipitation Sequencing   ◾    241

Figure 6.16 shows the dot plot of the first ChIP-Seq sample.
The IDs, descriptions, and statistics of the significant GO terms are stored in “chip1_
GO.csv”, “chip2_GO.csv”, and “chip3_GO.csv”. In Figure 6.16, we can notice that those top
ten GO terms are associated with gene transcription which reflects the Poly II biological
activity. The definitions of the GO terms can be searched at “http://www.informatics.jax.
org/vocab/gene_ontology/”. Thus, ChIP-Seq provides information about the functions of
the protein studied.
We can also use KEGG database for gene pathways to annotate the genes with signifi-
cant peaks. The “enrichKEGG()” function returns the enrichment KEGG categories with
FDR control. The following codes generate KEGG signaling pathway annotation and cre-
ate dot plot for each sample (Figure 6.17):

ekegg1 <- enrichKEGG(gene = entrez1, organism = ‘hsa’,

pvalueCutoff = 0.05)
cluster_kegg1 <- data.frame(ekegg1)
write.csv(cluster_kegg1, “kegg_chip1.csv”)
dotplot(ekegg1)
#Chip2
ekegg2 <- enrichKEGG(gene = entrez2, organism = ‘hsa’,
pvalueCutoff = 0.05)
cluster_kegg2 <- data.frame(ekegg2)
write.csv(cluster_kegg2, “kegg_chip2.csv”)
dotplot(ekegg2)
#Chip3
ekegg3 <- enrichKEGG(gene = entrez3, organism = ‘hsa’,
pvalueCutoff = 0.05)
cluster_kegg3 <- data.frame(ekegg3)
write.csv(cluster_kegg3, “kegg_chip3.csv”)
dotplot(ekegg3)

The significant KEGG signaling pathways show the most likely active pathways in the cells.
We can also compare enrichment across samples by using “compareCluste()” function,
which requires the list of genes from each sample (Figure 6.18).

# Create a list with genes from each sample

genes = lapply(annotated_peaks, function(i) as.data.
frame(i)$geneId)

# Run KEGG analysis

compKEGG <- compareCluster(geneCluster = genes,
fun = “enrichKEGG”,
organism = “human”,
pvalueCutoff = 0.05,
pAdjustMethod = “BH”)
dotplot(compKEGG, showCategory = 10, title = “KEGG Pathway
Enrichment Analysis”)
242   ◾    Bioinformatics

FIGURE 6.17 KEGG dot plot showing the top 10 genes.

FIGURE 6.18 KEGG signaling pathway comparison across the three samples.
 Chromatin Immunoprecipitation Sequencing   ◾    243

6.3.9 Motif Discovery

The major goal of ChIP-Seq is the determination of the binding sites, where TFs, Poly II, and
histone marks interact with the genomic DNA to control the transcription of genes. Those
sites have sequence patterns that are recognized by the targeted proteins. The genomic
sequence pattern that has such biological activity is called a motif. The motifs are usually
found in the genes’ regulatory regions. Therefore, they are most likely to be found in the
peak enrichment regions. The motif enrichment analysis is used to detect enrichment of
known binding motifs in the regulatory regions of genes. The researchers use motif analy-
sis to detect the binding site patterns of known library of TFs that are believed to regulate
a specific set of genes. Motifs are searched around the ChIP-Seq peaks of a specified win-
dow size. Remember that we have peak enrichment stored in “*peaks.narrowPeak” files.
However, the motif detection programs require FASTA sequence as input. Therefore, we
need to generate FASTA sequences from the BED file. We can create BED files by extract-
ing the first three columns from “*peaks.narrowPeak” files as follows:

mkdir motifs
cut -f 1,2,3 \
macs3output/chip1_peaks.narrowPeak \
> motifs/chip1_peaks.bed
cut -f 1,2,3 \
macs3output/chip2_peaks.narrowPeak \
> motifs/chip2_peaks.bed
cut -f 1,2,3 \
macs3output/chip3_peaks.narrowPeak \
> motifs/chip3_peaks.bed

The above commands create a new directory, “motifs”, and store the new created BED files
in it. We will extract FASTA sequences from each of these three files using bedtools, which
is a collection of programs for manipulation of BED files. On Ubuntu, you can install bed-
tools using “apt-get install bedtools”.
Visit the program website “https://bedtools.readthedocs.io/en/latest/content/installa-
tion.html” for more information.
The “bedtools getfasta” command is used to extract a FASTA file from each BED file.
This command requires the FASTA file of the reference sequence and a bed file as input.
We will use the same reference sequence that we used to generate BAM files.

bedtools getfasta \
-fi ref/hg19.fa \
-bed motifs/chip1_peaks.bed \
-fo motifs/chip1_peaks.fasta
bedtools getfasta \
-fi ref/hg19.fa \
-bed motifs/chip2_peaks.bed \
-fo motifs/chip2_peaks.fasta
244   ◾    Bioinformatics

bedtools getfasta \
-fi ref/hg19.fa \
-bed motifs/chip3_peaks.bed \
-fo motifs/chip3_peaks.fasta

Those three FASTA files contain the sequences of enriched peaks for each sample and we
will use them as inputs for the motif detection programs.
There are two approaches for motif detection: de novo method when no prior informa-
tion is assumed and a position weight matrix (PWM) method for known motif.
The de novo approach searches for motifs in an input FASTA sequences without prior
information about the motifs. The search is conducted in a window around the peak. The
motif discovery programs either create k-mers from the sequences and perform exhaustive
search to identify the most frequent consensus substring of the sequences as motifs or use
sequence alignments iteratively to create consensus motifs from the PWM that identifies
motifs as the consensus motifs with the most frequent nucleobases. An example of de novo
motif discovery program is MEME Suite [11], which has DREME, MEME, or STREME
programs for discovering ungapped motifs. DREME is k-mer based, but it is depreciated
and will not be supported in the future. MEME is an alignment-based motif discovery
tool but it is recommended for motifs discovery in less than 50 sequences. STREME is a
k-mer based and it is recommended for detecting motifs in a dataset with more than 50
sequences. MEME SUITE is available as web server and command-line programs. To use
the web server or to download and install MEME SUITE, visit “https://meme-suite.org/
meme/”. On Linux, you can download and install MEME SUITE by using the following
steps:

wget https://meme-suite.org/meme/meme-software/5.4.1/meme-
5.4.1.tar.gz
tar vxf meme-5.4.1.tar.gz
cd meme-5.4.1
./configure --prefix=$HOME/meme --enable-build-libxml2
--enable-build-libxslt
make
make test
make install

Once you have installed it, you can add the following to “.bashrc” file:

export PATH=$HOME/meme/bin:$HOME/meme/libexec/meme-5.4.1:$PATH

The version may change so the best way is to visit the MEME SUITE website for the lat-
est installation instruction.
After adding the above line to the “.bashrc” file, you may need to restart the terminal or
use “source ~/.bashrc” for the change that you have made to take effect.
The MEME Suite programs require the ChIP-Seq dataset in FASTA (primary data-
set) and control dataset (secondary dataset). If no control dataset is used, MEME Suite
 Chromatin Immunoprecipitation Sequencing   ◾    245

programs will create a control dataset by shuffling each of the sequences in the primary
input dataset.
The following DREME command will search for motif in the FASTA sequences of the
three ChIP-Seq samples. However, because the process may take a long time and this is just
a practice, you can run the command for a single sample only to save time. Run the com-
mands from inside “motifs” directories, where FASTA files are found.

dreme -verbosity 2 \
-oc dreme_motifs_chip1 \
-dna \
-p chip1_peaks.fasta \
-t 14400 \
-e 0.05
dreme -verbosity 2 \
-oc dreme_motifs_chip2 \
-dna \
-p chip2_peaks.fasta \
-t 14400 \
-e 0.05
dreme -verbosity 2 \
-oc dreme_motifs_chip3 \
-dna \
-p chip3_peaks.fasta \
-t 14400 \
-e 0.05

The “-oc” specifies the output directory, “-dna” specifies the type of sequence, “-p” specifies
the primary dataset, “-t” specifies an elapsed time as a stopping criterion, and “-e” specifies
the E-value threshold.
The output files will be saved in the directories “dreme_motifs_chip”. The motifs are
reported in an HTML file, an XML file, and a text file. You can open each of these files by
using the right program. You can change into each of the output directory and display the
HTML file using Firefox as follows:

firefox dreme.html

Figure 6.19 shows the motifs as displayed on the HTML file. The figure shows motif
sequence, logo, RC logo (reverse complement logo), and E-value. The motif sequence logo
is a graphical representation of the sequence conservation of DNA nucleotides. A DNA
sequence logo consists of the four nucleobase letters A, C, G, and T at each position. The
relative sizes of the letters reflect their frequency in the aligned sequences. The sequence
of the motif uses the IUPAC codes for nucleotides for representing each of the 15 possible
combinations as shown in Table 6.2.
246   ◾    Bioinformatics

FIGURE 6.19 The motifs found by DREME.

DREME is depreciated by the developer. So, it is recommended to use MEME or

STREME instead. Since we wish to find motifs in more than 50 FASTA sequences, we will
use STREME program. STREME will attempt to find motifs by iterating five steps until the
user-specified stopping criterion is met. The stopping criterion can be either the p-value of
the motif or the total number of motifs found. The five steps include building of a suffix
tree for both the ChIP-Seq FASTA sequences and control sequences, evaluation of the seed
words, motif refinement, motif significance computation, and motif erasing (converting
the site of the best motif to N in the primary sequences and control sequences).
The following STREME command searches for motifs in the FASTA sequences of the
three ChIP-Seq samples:

streme --p chip1_peaks.fasta \

--oc streme_motifs_chip1 \
--dna \
--thresh 0.05
streme --p chip2_peaks.fasta \
--oc streme_motifs_chip2 \
--dna \
--thresh 0.05
 Chromatin Immunoprecipitation Sequencing   ◾    247

TABLE 6.2 IUPAC Codes for DNA Alphabet

Nucleobase Name IUPAC Symbol Complement
Adenine A T
Cytosine C G
Guanine G C
Thymine TU A
Name Symbol(s) Matches
Any base N.X ACGT
Not T V ACG
Not G H ACT
Not C D AGT
Not A B CGT
Amino M AC
Purine R AG
Weak W AT
Strong S CG
Pyrimidine Y CT
Keto K GT

FIGURE 6.20 The motifs found by STREME.

streme --p chip3_peaks.fasta \

--oc streme_motifs_chip3 \
--dna \
--thresh 0.05

Each of the above commands will create a directory, “streme_motifs_*”. Four files will be
produced inside each directory: “sequences.tsv”, “streme.html”, “streme.txt”, and “streme.
xml”. Figure 6.20 shows how “streme.html” is displayed in the Internet browser.
248   ◾    Bioinformatics

HOMER is another motif discovery program that has an algorithm which performs de
novo motif discovery in the regulatory regions of genes using a primary dataset and a sec-
ondary dataset. HOMER has two programs, findMotifs.pl and findMotifsGenome.pl, that
perform the motif discovery in promoter and genomic regions, respectively. However, for
ChIP-Seq enrichment sequences, we will use findMotifs.pl or homer2. Like MEME Suite
programs, HOMER can also create a secondary random dataset if you do not have one.
For the instructions of HOMER installation, visit “http://homer.ucsd.edu/homer/intro-
duction/install.html”. Homer includes a collection of perl and c++ programs designed to
run in a UNIX/Linux environment. Refer to the website for the program requirements.
On Ubuntu, after making sure that all required programs have been installed, you can
create a directory “homer”, download the installation Perl script, and install the HOMER
as follows:

mkdir homer; cd homer

wget http://homer.ucsd.edu/homer/configureHomer.pl
perl configureHomer.pl -install

After the installation is complete, you will get a message as follows:

Add this line to your .bash_profile or .bashrc file (or other depending on your shell):

PATH=$PATH:/home/username/downloads/homer/.//bin/

The path will depend on where you have installed HOMER. Copy the path and add it to the
end of “.bashrc” file as follows:

export PATH=$PATH:/home/username/downloads/homer/.//bin/:$PATH

You may need to restart the terminal or use “source ~/.bashrc” for the change to take effect.
To test HOMER installation, run “findMotifs.pl”. That will display HOMER usage and
option. The “findMotifs.pl” command requires your target sequences’ file and a background
sequence file in FASTA format specified by “-fasta” option. However, if you do not provide
a background sequences’ file, the command will use your target sequences to create back-
ground sequences. The general syntax for the use of “findMotifs.pl” command is as follows:

findMotifs.pl <targetSequences.fa> fasta <output directory> -fasta

<background.fa> [options]

We can use this command without providing background sequences’ file as follows:

findMotifs.pl chip1_peaks.fasta fasta homer_motifs_chip1/

findMotifs.pl chip2_peaks.fasta fasta homer_motifs_chip2/
findMotifs.pl chip3_peaks.fasta fasta homer_motifs_chip3/

Three output directories will be created and several files are produced in each of the output
directories:
 Chromatin Immunoprecipitation Sequencing   ◾    249

The files “homerMotifs.motifs*” are the files that contain information of the motifs
found by the de novo motif discovery method, separated by motif length (“*” simply rep-
resents an integer for the length of the motif sequence). The motif information includes the
sequence, statistics, and PWM.
The file “homerMotifs.all.motifs” contains the information of all motifs found by the
de novo method. It is the concatenated file made up of all the “homerMotifs.motifs*” files.
The file “motifFindingParameters.txt” contains the command line used to execute the
program for the motif discovery.
The file “knownResults.txt” contains the statistics about known motif enrichment,
including motif name, consensus, p-value, Log p-value, q-value (Benjamini), number of
target sequences with motif, percentage of target sequences with motif, number of back-
ground sequences with motif, and the percentage of background sequences with motif.
The file “seq.autonorm.tsv” contains autonormalization statistics for the oligo.
The files “homerResults.html” and “knownResults.html” contain the output of de novo
motif finding and known motif finding, respectively. These HTML files can be displayed
on the Internet browser or using Firefox.

firefox homerResults.html knownResults.html

Figure 6.21 shows the motifs (logos and statistics) found by the de novo method using
“findMotifs.pl”. The motifs are ordered by the p-value. The significant motifs must have
very small p-value. Each motif has a link to the motif file.
The PWM motif discovery methods use prior information about the binding or inter-
action sites, obtained via laboratory means. Motif information are stored in a PWM file
format that describes the probability of finding the respective nucleotides A, C, G, and T
on each position of a motif. The word or PWM search method is used only to detect known
binding sites of transcription factors. PWM files of known motifs are used to scan windows

FIGURE 6.21 The motifs found by HOMER.

250   ◾    Bioinformatics

of a sequence for detecting the presence of the known binding sites of interest. PWM files
of known motifs can be download from motifs’ database such as JASPAR at “https://jaspar.
genereg.net/”. We will use MAST, which is one of the MEME Suite programs, to search for
known motifs in our sequences. Assume that we wish to search for TATA binding site in
our example sequences. First, we need to download the motif file from a database and then
run the program as follows:

wget https://jaspar.genereg.net/api/v1/matrix/MA0108.1.meme
mast -mt 5e-02 \
-oc mast_chip1 \
MA0108.1.meme \
chip1_peaks.fasta

Three output files (mast.html, mast.xml, and mast.txt) will be saved in the “mast_chip1”
directory. You can use “firefox mast.html” to view the results.

6.4 SUMMARY
Identification of binding sites of proteins on the genomic DNA is critical for understand-
ing gene regulation, pathways, and role of specific proteins in gene regulation and their
implications of some diseases. Therefore, ChIP-Seq is used to study epigenetic change that
affects gene expression and the impact of such changes on diseases. The ChIP-Seq is the
most effective way to identify protein-binding sites on the genomic DNA. The binding
sites of transcription factors and RNA polymerase II are found in the promoter regions of
genes. In a ChIP experiment, the genomic DNA is cut into fragments. The DNA regions,
where the protein of interest binds, are precipitated using a specific antibody. The protein
molecules are then removed from the DNA fragments. The isolated DNA fragments are
then sequenced using one of the sequencing techniques. The DNA library preparation and
sequencing are similar to that of other sequencing applications. The sequence reads (in
FASTQ files) produced by the sequencer are for the ChIP-Seq DNA reads that are likely to
contain the binding sites for the protein of interest. The quality control step is carried out to
reduce the error and to trim and remove adaptors and other technical sequences that may
affect the analysis results. The cleaned reads are then aligned to a reference genome to pro-
duce BAM files that contain the alignment information of the ChIP reads. The unaligned,
random, and mitochondrial reads are usually removed from the BAM files to reduce the
computational burden. The peak enrichment regions, where the binding sites are most
likely to be found, are called using one of the peak-calling programs. The peak information
for each sample is saved in a BED file. We have used R Bioconductor package to visualize
the distribution of the peaks and to perform annotation and functional analysis including
GO and KEGG pathways. GO and KEGG enrichment analyses provide knowledge-based
biological information. Finally, we used motif discovery programs to identify the motifs
on the promoter regions.
 Chromatin Immunoprecipitation Sequencing   ◾    251

REFERENCES
1. Bernstein BE, Meissner A, Lander ES: The mammalian epigenome. Cell 2007, 128(4):669–681.
2. Zhang Y, Liu T, Meyer CA, Eeckhoute J, Johnson DS, Bernstein BE, Nusbaum C, Myers
RM, Brown M, Li W et al: Model-based Analysis of ChIP-Seq (MACS). Genome Biol 2008,
9(9):R137.
3. Narlikar L, Jothi R: ChIP-Seq data analysis: identification of protein-DNA binding sites with
SISSRs peak-finder. Methods Mol Biol 2012, 802:305–322.
4. Heinz S, Benner C, Spann N, Bertolino E, Lin YC, Laslo P, Cheng JX, Murre C, Singh H, Glass
CK: Simple combinations of lineage-determining transcription factors prime cis-regulatory
elements required for macrophage and B cell identities. Mol Cell 2010, 38(4):576–589.
5. Ibrahim MM, Lacadie SA, Ohler U: JAMM: a peak finder for joint analysis of NGS replicates.
Bioinformatics 2015, 31(1):48–55.
6. Feng J, Liu T, Qin B, Zhang Y, Liu XS: Identifying ChIP-seq enrichment using MACS. Nat
Protoc 2012, 7(9):1728–1740.
7. Feng J, Liu T, Zhang Y: Using MACS to identify peaks from ChIP-Seq data. Curr Prot Bioinfor
2011, Chapter 2:Unit2 14–12.14.
8. Freese NH, Norris DC, Loraine AE: Integrated genome browser: visual analytics platform for
genomics. Bioinformatics 2016, 32(14):2089–2095.
9. Yu G, Wang LG, He QY: ChIPseeker: an R/Bioconductor package for ChIP peak annotation,
comparison and visualization. Bioinformatics 2015, 31(14):2382–2383.
10. Yu G, Wang LG, Han Y, He QY: clusterProfiler: an R package for comparing biological themes
among gene clusters. Omics 2012, 16(5):284–287.
11. Bailey TL, Gribskov M: Combining evidence using p-values: application to sequence homol-
ogy searches. Bioinformatics 1998, 14(1):48–54.
 Chapter 7

Targeted Gene Metagenomic

Data Analysis

7.1. INTRODUCTION TO METAGENOMICS

We can use any of the sequencing applications discussed in the previous chapter to study
an individual bacterium. Rather than a single species of bacteria, metagenomics involves
studying the genomes of a community of bacteria recovered from environmental or ­clinical
samples to obtain a variety of knowledge of the microbial species present in the sample and
their impacts on other living organisms. The most common samples targeted by metage-
nomics include human skin and cavities, digestive system, water, plants, soil, waste (liquid
and solid, feces), and food products. Metagenomics has a variety of uses including identi-
fication of an unknown pathogen in outbreaks of a disease, clinical diagnosis, monitoring
human and animal health, identification of bioactive compounds (terragines, violacein,
and indirubin) [1], drugs from marine microorganisms such as cytarabine (anti-cancer)
[2], cephalosporins (anti-microbial) [3], and vidarabine (anti-virus) [2], discovery of novel
antibiotics, production of some enzyme (lipases, proteases, lyases, amylases, etc.), explor-
ing new industrial and healthy products (indigo, probiotics) [4], and investigation and
monitoring of wildlife health.
In a typical metagenomics study, genomic DNA of the whole bacterial community in
a sample is extracted and then sequenced, and a pipeline of analyses is conducted on the
acquired sequence data. There are two sequencing approaches to characterize microbial
taxonomic groups in environmental samples. The first one targets a specific marker gene
or a region of a marker gene after being amplified with polymerase chain reaction (PCR).
The 16S rRNA gene is often used for this purpose. The second approach targets all bacte-
rial genomes in the samples and uses the shotgun whole genome sequencing to achieve
that. We will discuss the shotgun sequencing in the next chapter. In this chapter, we will
focus on the amplicon-based sequencing to identify bacteria in environmental or clinical
samples.

DOI: 10.1201/9781003355205-7 253

254   ◾    Bioinformatics

The analysis of the metagenomics data involves classification of an individual sequence

to a bacterial taxon or taxa. Bacteria are classified into hierarchical taxonomic groups in
descending order (kingdom, phylum, class, order, family, genus, species, and subspecies).
The analysis also includes construction of phylogenetic tree for the bacterial community,
quantification of bacterial presence in the sample (abundance), and the microbial diversity
in an individual sample and across samples. A number of bioinformatics tools are available
for analyzing metagenomic data.
In the amplicon-based metagenomics, the targeted region is a marker gene or any part
of a gene or any genomic region appropriate for microbial identification. The ideal tar-
get region is the one that includes a highly conserved sequence surrounded by less con-
served and it must be present ubiquitously in all target species and with available reference
sequences in the sequence databases. The 16S rRNA gene is one of the best candidate
marker genes. The 16S rRNA gene codes for a component of the 30S small subunit of the
bacterial ribosome. It is around 1500 bp consisting of nine conserved regions surrounding
hyper-variable regions. The sequences of the conserved regions are labeled as C1, C2, …,
C9 and the variable regions are labeled as V1, V2, …, V9. PCR primers are designed from
the conserved region close to the variable region so that the species-specific targeted region
is amplified and enriched by the PCR amplification. Then, the amplicons are sequenced
and analyzed.
Several samples are usually sequenced in a single run using multiplexing approach in
which unique barcode sequences are ligated to the DNA of each sample in the library
preparation step. After sequencing, the reads are demultiplexed by separating the reads of
the individual samples into separate FASTQ files before analysis. Since the amplicon-based
metagenomics depends on a targeted region of the genomes, it has less resolution than the
shotgun whole genome sequencing but is less expensive.

7.2 ANALYSIS WORKFLOW

In the following, we will discuss the steps of the workflow of the amplicon-based metage-
nomics data analysis, which include raw data preprocessing, read clustering, denoising
(error removal), taxonomic group assignment, construction of phylogenetic tree, and
diversity analysis.

7.2.1 Raw Data Preprocessing

After sequencing the targeted marker, raw sequence data is obtained in FASTA or FASTQ
format. In the case of FASTA format file produced by Sanger sequencing method, the per
base quality score may be provided in a separate file. The format of FASTQ files allows
base quality scores to be in the same file. The base quality scores reveal base call qual-
ity of each base and enable us to assess the sequence reads and to determine if the reads
require preprocessing before the analysis. The quality control step should not be taken
lightly. Errors in base calls may occur due to the library preparation or sequencing. The
reads of low-quality scores are usually filtered or truncated so that the errors do not affect
the final results. The preprocessing of the raw sequence data may also include demultiplex-
ing if multiple samples are sequenced in a single run. The demultiplexing step depends on
 Targeted Gene Metagenomic Data Analysis   ◾    255

how the reads and barcodes are stored. The reads also might be sequenced as single end or
paired end. The paired-end reads are usually stored in two FASTQ files.

7.2.2 Metagenomic Features

After the preprocessing, the raw reads of the targeted gene (16S rRNA) are dereplicated
by reducing the abundance of the similar reads based on a Hamming distance threshold.
The dereplicated reads are then clustered by collapsing the complete sets of reads into a
grouping units called operational taxonomic units (OTUs) [5]. Once the OTUs have been
formed, representative sequences (features) can either be selected from those OTUs or the
sequences of each OTUs are denoised to remove potential errors before selecting represen-
tative sequences. In the following, we will discuss both clustering and denoising.

7.2.2.1 Clustering
There are three approaches for clustering the preprocessed amplicon reads. The first is the
de novo clustering [6] in which the reads are clustered into OTUs according to their simi-
larity at a specified threshold. It compares each read against each other and then it groups
sequences into OTUs. The second is called open-reference clustering in which reads are
clustered around previously annotated reference sequences by sequence classification or
searching in a database (Greengenes database) and the reads that do not cluster around
reference sequences are clustered with de novo method. The third is the closed-reference
clustering in which reads are clustered around reference sequences and the reads that do
not find reference sequences are removed [7]. The closed-reference clustering may intro-
duce reference bias to the clustering process since the reads without references will be
removed, while they may represent novel species or taxa. The de novo clustering is pre-
ferred, although it may be computationally expensive other than the other two clustering
methods. The algorithms for de novo clustering include hierarchical clustering and heu-
ristic clustering. The hierarchical clustering method requires a matrix for the pairwise dis-
tances between all unique sequences. A hierarchical tree is constructed from the distance
matrix, and then the OTUs are constructed from the tree based on a distance threshold.
The heuristic clustering uses pairwise sequence alignment to construct a distance
matrix one by one instead of computing all distances in a single step. The heuristic cluster-
ing method generates OTUs in a greedy incremental strategy that utilizes one sequence as
a seed to represent a cluster. Then, each one of the unique reads is compared with the seeds
of existing clusters. A read is assigned to a cluster if the pairwise distance between a unique
read and a seed meets a predefined threshold. If a read does not join any cluster, that read
will become a seed for a new cluster. The final clusters represent the OTUs. The heuristic
clustering method is more computationally efficient than the hierarchical clustering. There
are a variety of heuristic clustering methods that vary in seed selection and distance cal-
culation. Examples of the tools that use heuristic clustering for constructing OTUs include
USEARCH and CD-HIT.
In general, an OTU constructed by any of the clustering methods is considered as a tax-
onomic group. From this point, the analysis can move on to the taxonomic group assign-
ment, construction of the phylogenetic tree, and diversity analysis but the latest analytic
256   ◾    Bioinformatics

tools proposed an extra step before taxonomic assignment. This step is meant to reduce the
potential errors that may produce noises; therefore, the step is called denoising.

7.2.2.2 Denoising
There are two possible types of errors that may occur on deciding whether the variation
within an OTU represents errors or real diversity. The first type is the base calling error
which may arise from the sequencing. This type of errors may occur due to the incorrect
base pairing during the PCR amplification, polymerase slippage, or PCR chimeras that
are formed when the DNA strand extension is aborted during the PCR process and the
aborted products act as primers in the next PCR cycle producing artifacts. The second
type of errors is the misclassification of a read to an incorrect taxonomic group. This error
can be corrected by constructing OTUs at a particular similarity threshold such as 97%.
However, that may come at the cost of taxonomic sensitivity. Denoising is attempting to
handle these errors by using the reads to infer the correct biological sequences. This way
the misclassification can be avoided.
Several computational approaches have been proposed for sequence denoising. The most
commonly used approaches are DADA2, Deblur, and UNOISE3 which are able to infer
error-free biological sequences at a single-nucleotide resolution. Those inferred sequences
that will be used for taxonomic assignment are called features, zero-radius OTUs (ZOTUs),
exact sequence variants (ESVs), or amplicon sequence variants (ASVs). In the following, we
will discuss those three popular denoising methods.

7.2.2.2.1 DADA2 Denoising

DADA2 (Divisive Amplicon Denoising Algorithm 2) [8] was adapted to use with Illumina
sequencing and available as an open-source R package and as plugin in QIIME2, which
is an open-source command-line Linux program. DADA2 implements a new model of
Illumina-sequenced amplicon errors that incorporates quality information of the within-
sequence errors and between-sequence errors. The model quantifies the error rate (λ) at
which an amplicon read is produced from a sample sequence as a function of a sequence
composition and quantity. The number of amplicons or abundance follows the Poisson
distribution with the parameter λ ji , which is the error rate at which an amplicon read
with sequence i is produced from sample sequence j. The abundance of the sequence i has
an expected value equal to an error rate λ ji multiplied by the expected reads of sample
sequence j.
The DADA2 model assumes that errors occur independently with a read and indepen-
dently between reads. The model then estimates the error rate as the product over the
transition probabilities between the L aligned nucleotides and associated quality score of
the original nucleotide as follows:
L

λ ji = ∏ p ( j ( l ) → i ( l ) , q ( l ))
l =0
i (7.1)
 Targeted Gene Metagenomic Data Analysis   ◾    257

where p ( j ( l ) → i ( l ) , qi ( l )) is the transition probability between aligned nucleotides j ( l ) and

i ( l ) and the associated quality score qi ( l ), e.g., p(T→G, 40).
The divisive partitioning algorithm begins with all amplicon reads in a single parti-
tion. The error rate is then used to model the number of observed reads of each unique
sequence to compute the p-value of the hypothesis that the number of amplicons of each
unique sequence is consistent with the error model. According to the DADA2 model, for
the unique sequence i with abundance ai be in partition j containing n j reads, the abun-
dance p-value is given as

pA ( j → i ) =
1
∑p
1 − ppoisson (n j λ ji ,0 ) a=a
i
poisson (n λ , a )
j ji (7.2)

These p-values of the unique sequences are used as the division criteria for an iterative
partitioning. A threshold is specified for partition; if the smallest abundance p-value falls
below the threshold, a new partition is formed with that unique sequence allowing other
similar unique sequences to join it. The division continues iteratively until all unique
sequences falling within a OTUs are consistent with abundance p-values greater than the
specified threshold.
The output of the divisive amplicon denoising algorithm is a collection of ASVs, which
are exact sequences with defined statistical confidence. Because ASVs are exact sequences,
generated without clustering or reference databases, they can be readily compared between
studies using the same target region. DADA2 pipeline generates an ASV table that can be
used for downstream analysis.

7.2.2.2.2 Deblur Denoising

Deblur [9] is a denoising method that uses error profiles of amplicons sequenced by
Illumina MiSeq and HiSeq sequencing platforms to infer error-free sequences. Unlike
DADA2, Deblur operates on each sample independently. The Deblur algorithm begins by
comparing the pairwise Hamming distances of all sequences within a sample to an upper-
bound error profile. The unique sequences are sorted by abundance by ascending from
the most to the least. Neighboring reads are formed for each read based on a Hamming
distance threshold. The number of incorrect reads is then subtracted from the abundance
of the neighboring reads using an upper bound on the error probability. After subtraction,
the sequences with zero abundance are considered as a noise and dropped from the list
of the valid sequences. Deblur can infer the correct sequences. However, it may decline to
remove PCR chimeras that are produced from the aborted PCR cycle.

7.2.2.2.3 UNOISE2 Denoising

Unlike DADA2, UNOISE [10] denoising does not use quality score and it utilizes one-
pass clustering strategy with only two parameters (α and β) with pre-set values. A unique
read sequence (M) in a cluster is evaluated based on its Levenshtein distance (d) from
258   ◾    Bioinformatics

the centroid sequence (C) and skew of its abundance (aM ) with respect to the centroid
sequence abundance (aC), which is given as

skew ( M , C ) = aM a (7.3)
C

When a member unique sequence has both an enough small distance and an enough small
skew with respect to the centroid sequence, then it is likely that sequence is incorrect read
of the centroid sequence with d points errors. The maximum skew (β) allowed for a cluster
member with d differences from the centroid sequence is given by

1
β (d ) = α d +1 (7.4)
2
where α is set to 2 by default.
We can notice that as the distance d between the member sequence and centroid
increases, the maximum skew β decreases exponentially.
The unique sequences with low abundance are removed by the UNOISE2 algorithm.

The final products of any of the clustering and denoising methods are feature table and the
list of representative sequences. The feature table provides the feature abundance or the
number of a times a feature has been observed in a sample. A feature is a unit of observa-
tion that can be an OTU or an ASV. The feature table is needed for the downstream ­analysis
such as taxonomy assignment, construction of phylogenetic tree, and diversity analysis.

7.2.3 Taxonomy Assignment

Given a set of representative sequences generated by the above-discussed clustering or
denoising methods, the taxonomy assignment step will attempt to assign taxa for each
sequence. There are several methods for assigning taxonomy but, in general, they can
be categorized into (i) alignment-based methods such as BLAST and VSEARCH and (ii)
machine learning methods such as Ribosomal Database Project (RDP) Classifier. The out-
put of the taxonomy assignment methods is mapping a representative sequence to taxa and
providing an assignment quality score.

7.2.3.1 Basic Local Alignment Search Tool

The Basic Local Alignment Search Tool or BLAST [11] is a widely used seed-based heuris-
tic sequence search tool whose algorithm is adopted from the Smith-Waterman algorithm
for local sequence alignment. Providing a representative sequence (generated from cluster-
ing or denoising) as a search query to BLAST, the search is conducted against a database of
sequences with known taxonomy. Rather than aligning to a single sequence, the taxonomy
assignment is based on the consensus of hits in the reference database that exceed the prede-
termined percent identity. If the blast hits agree on the same taxonomy, then the representa-
tive sequence will be given that taxonomy level with consensus greater than a threshold.
 Targeted Gene Metagenomic Data Analysis   ◾    259

7.2.3.2 VSEARCH
VSEARCH [12] is another alignment-based and open-source tool like BLAST. However,
unlike BLAST, which uses seed-based heuristic search algorithm, it uses a fast heuris-
tic search by identifying a small set of database sequences that have many k-mer words
in common with the query sequence (a representative sequence). VSEARCH algorithm
counts the number of shared words between a representative sequence and each data-
base sequence (a word will count once if it appears multiple times). Thus, the similarity
between a representative sequence and a target sequence is based on the statistics of shared
words rather than alignment. Then, the algorithm performs pairwise global alignment
(Needleman-Wunsch) for the query sequence with each sequence of the database begin-
ning with the sequences with the largest number of words in common with the representa-
tive sequence. The optimal global alignment score is computed for each alignment. Unlike
BLAST, exhaustive pairwise alignment is computationally expensive. However, VSEARCH
employs parallel strategies by using vectorization and multiple threading that reduce the
computational cost.

7.2.3.3 Ribosomal Database Project

Ribosomal Database Project (RDP) [13] Classifier is a machine learning taxonomy classifi-
cation method. It is a model trained by known reference sequences with known taxonomic
groups. After training, the classifier is used to predict the taxonomy of the representative
sequences. RDP uses the naïve Bayesian algorithm to accurately predict the taxonomy of a
sequence. The model is trained by N sequences with known taxonomy. First, the algorithm
forms a dictionary from all possible 8-base subsequences or words from all sequences. The
words of all N sequences are called corpus. Let W = {w1, w2, …, wd} be the set of all possible
eight-character words in the corpus and n(wi) be the number of sequences containing word
wi, where i=1, 2, …, d; thus, the expected-likelihood estimate of observing word wi in a
sequence is calculated as

Pi =
[n(wi ) + 0.5] (7.5)
( N + 1)
The numbers 0.5 and 1 are used to keep the value of Pi to be between 0 and 1 (e.g., 0 Pi <1).
Assume that m(wi) is the number of the word wi contained by M training sequences
whose genus is G. Thus, the conditional probability that a sequence of genus G contains the
word wi is estimated as:

P ( wi | G ) =
[m(wi ) + Pi ] (7.6)
M +1

Now assume that a sequence S of genus G has the observed set of words V={v1, v2, …, vf }
V⊑W; thus, the joint probability or likelihood of observing the sequence S is the product of
the probability of each word given genus G as
260   ◾    Bioinformatics

P ( S|G ) = ∏ P (v | G )
i (7.7)

Finally, the probability that an unknown query sequence S is a member of genus G is

given as

P (G )
P (G|S ) = P(S | G ) × (7.8)
P(S)

where P(G│S) is the posterior probability, P(G) is the class prior probability of a sequence
being a member of G, and P(S) is the predictor prior probability.
Since both class prior probability and predictor prior probability are constant for all
sequences, we can drop both P(G) and P(S). So, the above formula can be rewritten as

P (G|S ) = P ( S|G ) = ∏ P (v | G )
i (7.9)

Based on the naïve Bayesian, each representative sequence obtained from the denoising
step is assigned to the genus G or a taxonomic group that is giving the highest probability
score after calculating the score for each taxonomic group.

7.2.4 Construction of Phylogenetic Trees

A phylogenetic tree is a diagram that represents evolutionary relationships among organ-
isms. The pattern of branching in a phylogenetic tree describes how species or taxonomic
groups evolved from a series of common ancestors and how they are related. Two spe-
cies are more related if they have a more recent common ancestor and less related if they
have a less recent common ancestor. In amplicon-based metagenomic studies, phyloge-
netic trees based on targeted genes (e.g., 16S rRNA gene) are commonly used to describe
and compare the taxonomic groups of the bacterial community in the sample. There are
several tree-building algorithms implemented in software. The construction of the phylo-
genetic tree begins with computing genetic distances between the unique representative
sequences identified in the above steps. The genetic distance between two representative
sequences is the number of mutations or evolutionary events between two sequences since
their divergence. The traditional Hamming method calculates genetic distance between
any two sequences by counting the number of differences between them. However, this
method does not capture the evolutionary history of the sequences since not all muta-
tions are shown in the current sequences. The Hamming distance is corrected with Jukes–
Cantor logarithmic correction. Let p be the distance between two sequences, the distance
(d) after Jukes–Cantor correction is given as

3  4 
d = log  1 − p  (7.10)
4  3 
The tree is then built from the pairwise distances of the representative sequences. Several
tree-building methods are available. The most popular distance-based methods are the
 Targeted Gene Metagenomic Data Analysis   ◾    261

unweighted pair group method with arithmetic mean (UPGMA), weighted pair group
method with arithmetic mean (WPGMA), and neighbor joining (NJ) [14].
Both UPGMA and WPGMA assume a randomized molecular clock that measures the
evolutionary divergence of sequences. The molecular clock is defined as the average rate at
which a sequence accumulates mutations. Both UPGMA and WPGMA also have a similar
algorithm. They use a cluster procedure that assumes each representative sequence as a
cluster on its own and then they join the closest clusters and recalculate the distance of the
joint pair by the average. These steps are repeated until all sequences are connected in a
single cluster. However, the difference between the two methods is that in UPGMA, equal
weight is assigned on the distances, while in the WPGMA different weights are assigned
on the distances.
The algorithm of the NJ method does not make an assumption of the molecular clock
and it adjusts for the rate variation among branches. The algorithm begins with an initial
unsolved star-like tree made up of the representative sequences. The distance between each
pair is evaluated. The first joint is created by joining the closest two neighboring sequences
and a branch is inserted between them and the rest of the star-like tree. The value of the
branch is recalculated on the basis of their average distance. This process is repeated until
only one terminal is present from the initial tree.
The above briefly described tree construction methods are distance-based and less
computationally expensive. However, there are other methods including maximum par-
simony (MP) and maximum likelihood (ML) which make use of all known evolutionary
­information (individual substitutions) to determine the most likely ancestral relationships.
Refer to a book for phylogenetic tree for more details about the various tree construction
methods.
A phylogenetic tree is either rooted (with a common ancestor for all sequences) or
unrooted (without common ancestor). The unrooted trees are constructed when we do not
make the assumption that the molecular clock evolution is valid and they only reflect the
relationship among representative sequences but not the evolutionary path. However, if we
can make the assumption that sequences evolve at rates that remain constant through time
for different lineages, then the root of a tree is estimated as the midpoint of the longest span
across the tree.

7.2.5 Microbial Diversity Analysis

The microbial diversity or richness is calculated from the feature table, obtained in the
denoising step above, to describe the number of different species of microbes present
within individual samples and between samples. The diversity of the microbial community
within a sample is called alpha diversity, while the measure of similarity or dissimilarity
of microbial communities in two samples is called beta diversity. For alpha diversity, there
are several diversity metrics including Shannon’s diversity index, observed features, Faith’s
phylogenetic diversity, and evenness. The beta diversity metrics include Jaccard distance,
Bray–Curtis distance, and unweighted UniFrac distance.
262   ◾    Bioinformatics

7.2.5.1 Alpha Diversity Indices

7.2.5.1.1 Shannon’s Diversity Index
The Shannon’s Index also called Shannon–Wiener Index is the most commonly used mea-
sure for alpha diversity of an individual sample. It is defined as:

H= ∑p × ln( p )
i =1
i i (7.11)

where pi is the probability of finding species/taxon i in the sample and S is the number of
taxa in the sample or richness.
The values of Shannon diversity usually fall between 1.5 and 3.5. The higher the value
of H, the higher the diversity of taxa in a sample. The lower the value of H, the lower the
diversity. A sample with a single taxon will have H = 0 (not diverse sample).

7.2.5.1.2 Pielou’s Evenness

The evenness of taxa in a sample refers to how similar the abundances of the different taxa
in the sample are and it is obtained by dividing Shannon–Wiener Index by the natural
logarithm of the number of unique taxa in the sample as follows:

H
Evenness = (7.12)
ln( S )
The evenness value ranges from 0 to 1, where 1 indicates complete evenness or the taxo-
nomic groups in the sample have similar abundances.

7.2.5.1.3 Faith’s Phylogenetic Diversity Index

Faith’s phylogenetic diversity (PD) index is a qualitative measure of community richness.
It is based on phylogeny and is defined as the phylogenetic diversity of a set of sequences
being equal to the sum of the lengths of all the branches on the tree. A higher PD indicates
more richness or diversity in the sample.

7.2.5.2 Beta Diversity

Beta diversity quantifies the difference among biological communities in different samples.
It is defined as the ratio between regional (gamma) and local (alpha) diversities. Most of the
beta diversity metrics are based on dissimilarity measures such as Jaccard distance, Bray–
Curtis distance, and weighted and unweighted UniFrac distance. Let a be the number of
species/taxa found in both samples, b is the number of species/taxa in the first sample but
not in the second sample, and c is the number of species/taxa in the second sample but not
in the first sample. The beta diversity metrics are computed as follows.

7.2.5.2.1 Jaccard Diversity Distance

 b+c 
β jac =  ×100 (7.13)
 a + b + c 
 Targeted Gene Metagenomic Data Analysis   ◾    263

If the samples share all the taxonomic groups, the percent Jaccard index will be 100%
similar (β jac = 100% ); the closer it is to 100%, the more similar the samples are. If the two
samples share no species/taxa, they will be 0% similar (β jac = 0 ). If the percent Jaccard
index is 50%, the two sample will share half of the taxonomic groups.

7.2.5.2.2 Bray–Curtis Dissimilarity Index

 2c 
β br =  1 − × 100 (7.14)
 a + b 
The percent Bray–Curtis dissimilarity is always a number between 0 and 100. If it is 0, then
the two samples share all the same species; if it is 100, that means the two samples do not
share any species.

7.2.5.2.3 Unweighted and Weighted UniFrac Distance Index

The UniFrac is phylogenetic-based beta diversity index that takes into account the evolu-
tionary relatedness of the communities in the two samples. The UniFrac distance index
is defined as the fraction of the observed branch lengths of the phylogenetic tree that is
unique in either sample. If the communities of the two samples are identical, UniFrac
index will be zero. If the two communities are evolutionarily unrelated, the UniFrac index
would be 1.0. The UniFrac index is closer to zero if the communities of the two samples are
more evolutionarily related. The weighted UniFrac uses relative abundances of species in
the samples as a weight on the branch lengths (thus, it emphasizes the dominant species).
While unweighted UniFrac uses only presence or absence (thus, it emphasizes the rare
species).

7.3 DATA ANALYSIS WITH QIIME2

Now it is time to get your hand dirty with some worked examples that cover raw data pre-
processing, read clustering, denoising, taxonomic assignment, phylogenetic tree, and diver-
sity analysis. For this purpose, we will use QIIME2 (Quantitative Insights Into Microbial
Ecology 2) [15], which is the most commonly used free program for analysis of amplicon-
based microbial sequencing data. QIIME2 can be used for any analysis of any targeted gene
sequencing data but the program modules for the analysis of metagenomic data based on
16S rRNA gene are very well established. QIIME2 can be installed in different platforms.
For the detailed installation instructions, visit “https://docs.qiime2.org/2022.2/install/”.
If you have Anaconda installed on you Linux computer, you can install it with “conda
install -c qiime2 qiime2”; however, make sure that all requirements are met. We will use
QIIME2 under the Anaconda environment. After installing QIIME2 under Anaconda,
run “conda activate qiime” on the Linux terminal to activate QIIME2 environment. Once
it has been activated, the terminal prompt will change into something like “(qiime2)$”.
Then, you can run any QIIME command. To display the available QIIME2 commands,
run the following:

(qiime2)$ qiime
264   ◾    Bioinformatics

This will display the program usage, options, and available commands. A command can
be a QIIME2 utility tool (info, tools, dev) or QIIME2 plugins. We will use the “tools” com-
mand very often. To use a QIIME2 command or plugin, it must be in the “qiime <com-
mand>”. For instance, to use “tools” command, run the following:

(qiime2)$ qiime tools

The QIIME2 plugins are software packages developed by programmers to be used for anal-
ysis. Plugins work through two functional units: methods and visualizers. Before any data
processing, QIIME2 converts input raw data into a special format called QIIME2 artifacts.
An artifact is a compressed file with the “.qza” file extension. For the output, QIIME2

FIGURE 7.1 General QIIME2 workflow for amplicon-based sequence data analysis.
 Targeted Gene Metagenomic Data Analysis   ◾    265

generates another file format, with the “.qzv” file extension, called visualization file. This
visualization file is a standalone and sharable file that may contain any kinds of output such
as images, tables, and interactive representations. Plugin methods take QIIME2 ­artifacts
as input and produce an output. While a plugin visualizer produces a single visualiza-
tion file for the purpose of visualizing or sharing. In the following, we will show you how
to use QIIME2 to analyze targeted gene metagenomic data. The general workflow of the
amplicon-based metagenomic data analysis is shown in Figure 7.1.

7.3.1 QIIME2 Input Files

Metagenomic amplicon-based raw data may be acquired from a study conducted at the
laboratory or may be downloaded from a database. A study will have a design tailored to
address research objectives. The study design usually dictates the workflow of the analysis.
Whether raw data is from your own project or downloaded from a database, it usually
includes (i) raw sequence data and (ii) metadata (information of the samples and study
design). The analysis with QIIME2 requires importing these two inputs and converting
them into artifacts. Then, only the artifacts are the files that are used for the analysis.
Therefore, the first task is to import the raw sequence data and metadata into artifacts. The
QIIME2 artifacts have their semantic type that enables QIIME2 to identify the suitable
artifact for an analysis. In the following, we will discuss the raw sequence and metadata
with more details.

7.3.1.1 Importing Sequence Data

QIIME2 accepts inputs in a variety of file formats including FASTA, FASTQ files, and
feature files of OTUs or representative sequences. The FASTQ (single-end or paired-
end) files are the most commonly used. The reads in FASTQ files may be multiplexed or
demultiplexed. If they are multiplexed, they can either be multiplexed following the Earth
Microbiome Project (EMP) protocol (the barcode sequences are in a separate file) or non-
EMP (the reads are with in-sequence barcodes) [16]. On the other hand, the demultiplexed
reads can either be in Casava 1.8 format or not.
Some laboratories may have their own sequencer and others may depend on genomic
core facilities for sequencing. In either case, the raw data would be provided in one of the
above formats. Raw sequence data can also be downloaded from metagenomics databases.
Examples of metagenomic databases include NCBI SRA database available at “https://www.
ncbi.nlm.nih.gov/sra” and the EMBL-EBI-hosted MGnify database available at “https://
www.ebi.ac.uk/metagenomics/”. Both databases provide data generated from a variety of
microbiome studies on specific environment such as human body sites, soil, seawater, and
others. MGnify microbiome data can also be accessed from the NCBI SRA. Sequence data
like FASTQ files generated from those studies are stored in the NCBI SRA as sequence read
archives, which are compressed files and can be downloaded using SRA-toolkit.
Whether the input file for QIIME2 is FASTA file, FASTQ files, or feature file, it must be
imported by QIIME and converted into QIIME2 artifact. Importing an input file into an
artifact depends on the raw data file format; each file format is imported in a unique way.
But, in general, to import any input data, you must use “qiime tools import”. We already
266   ◾    Bioinformatics

know that to execute any QIIME2 command, you must use “qiime”. We also mentioned
above that “tools” is a command and command or plugin may have methods or visualizers
or both. The “import” above is a tools command method whose function is to import any
input data. However, for each input file format, different arguments and options will be
used. In general, importing an input file into an artifact will follow this format:

qiime tools import \

--type xxx #raw data type
--input-path xxx #the directory
--input-format xxx #the raw data format
--output-path xxx #where artifact file is to be saved

The “import” options are self-explanatory, where “--type” specifies the type of the data to
be imported, “--input-path” specifies the path to the raw data files, “--input-format” speci-
fies the input file format, and “--output-path” specifies the path where the artifact file will
be saved.
You must specify the right data type and format on “--type” and “--input-format” for
your input files to be imported successfully.
In the following, we will show you the formats for importing different output files, but
for practice and detailed information, visit “https://docs.qiime2.org/”.
We will discuss the import of the common types of raw data below.

7.3.1.1.1 Importing Raw Data in FASTA

If the raw data is in FASTA format, all sequences must be in a single FASTA file consisting
of exactly two lines for each sequence record; a line for the definition line and a single line
for sequence (not in multiple lines). The ID in each in the definition line must be as per the
FASTA specification (no white space between the “>” and the ID). Moreover, the ID must
follow the “SAMPLEID_SEQID” format, where SAMPLEID is a unique sample identifier
and SEQID is a sequence identifier. Assuming that your sequences are in the “sequences.
fasta” file, you can use the following “qiime tools import” to create an artifact for the input
file:

qiime tools import \

--type ‘FeatureData[Sequence]’ \
--input-path sequences.fna \
--output-path sequences.qza

The above command will import the FASTA sequences as a QIIME2 artifact “sequences.
qza” that is ready for the next step of the analysis.

7.3.1.1.2 Importing EMP-Multiplexed FASTQ Reads

The EMP-multiplexed FASTQ files contain multiplexed reads, but the barcode sequences
are in a separate file. Therefore, there will be two files (forward FASTQ file and barcode
FASTQ file) for single-end reads and three files for paired-end reads (forward, reverse,
 Targeted Gene Metagenomic Data Analysis   ◾    267

and barcode FASTQ files). We can import such multiplexed raw data to QIIME2 and later
demultiplex them using a QIIME2 command. To import a single-end EMP-multiplexed
reads, the forward FASTQ file name must be “sequences.fastq.gz” and the barcode file is to
be “barcodes.fastq.gz”. These two files can be in a directory say “data” for example. Then,
you can use “qiime tools import” to import the raw data using the following:

qiime tools import \

--type EMPSingleEndSequences \
--input-path data \
--output-path artifacts/multiplexed-emp-single-end.qza

Notice that we use “input-path” to specify the directory where the two files are found.
For the paired-end EMP-demultiplexed raw data, we can use the following:

qiime tools import \

--type EMPPairedEndSequences \
--input-path data \
--output-path artifacts/multiplexed-emp-paired-end.qza

Notice that both artifacts created by the above commands are for multiplexed reads. Those
artifacts required demultiplexed as we will discuss later.

7.3.1.1.3 Importing non-EMP-Multiplexed FASTQ Files

In the non-EMP-multiplexed reads, the barcode sequences are attached to the reads. This
read format will have a single FASTQ file (“sequences.fastq.gz”) for the single-end reads
and two files (“forward.fastq.gz” and “reverse.fastq.gz”) for the paired-end reads. The fol-
lowing are the commands to import single-end and paired-end raw data, respectively:

qiime tools import \

--type MultiplexedSingleEndBarcodeInSequence \
--input-path data/sequences.fastq.gz \
--output-path artifacts/multiplexed-single-end.qza

qiime tools import \

--type MultiplexedPairedEndBarcodeInSequence \
--input-path data \
--output-path artifacts/multiplexed-paired-end.gza

Again, the artifacts created by the above commands contain multiplexed reads. So, they
must be demultiplexed before proceeding. Demultiplexing will be discussed later.

7.3.1.1.4 Importing Casava Demultiplexed FASTQ Files

The Casava 1.8 demultiplexed formatted FASTQ file includes a barcode as the sample iden-
tifier in the FASTQ file name, which is made of four IDs separated by underscore: the
sample identifier, barcode identifier, slide lane number, direction of the read, and the set
268   ◾    Bioinformatics

number. QIIME2 can use that information in the file name to automate importing the files
and to create an artifact for that demultiplexed raw data. The following are the commands
for importing Casava 1.8 formatted demultiplexed, single-end and paired-end FASTQ
files, respectively:

qiime tools import \

--type ‘SampleData[SequencesWithQuality]’ \
--input-path data \
--input-format CasavaOneEightSingleLanePerSampleDirFmt \
--output-path artifacts/demultiplexed-single-end.qza

qiime tools import \

--type ‘SampleData[PairedEndSequencesWithQuality]’ \
--input-path data \
--input-format CasavaOneEightSingleLanePerSampleDirFmt \
--output-path artifacts/demultiplexed-paired-end.qza

These commands create artifacts for demultiplexed data that are ready for processing (the
samples are already separated).

7.3.1.1.5 Non-Casava Demultiplexed FASTQ

When the reads of each sample are in a separate FASTQ file for single-end reads or in two
FASTQ files for the paired-end reads, these demultiplexed files are non-Casava 1.8 demul-
tiplexed FASTQ files. Those are the most commonly used files. In our worked example,
we will practice in files with this format. Since the analysis of metagenomic data involves
multiple samples, QIIME2 will need to know the sample identification when multiple
FASTQ files for individual samples are used in the analysis. Thus, importing the non-
Casava 1.8 demultiplexed FASTQ files requires an additional text file that maps sample
identifiers to the FASTQ files. This file is called the manifest file and it will be created
by the user. A manifest file is a comma-separated value (CSV) file with three columns
with these column headers: “sample-id”, “absolute-filepath”, and “direction”. The headers
are self-explanatory; the sample-id column includes the sample IDs, the absolute-filepath
includes the absolute bath for a FASTQ file, and the direction column shows the direction
of the FASTQ file (forward or reverse). Figure 7.2 shows example manifest files for single-
end FASTQ files (on the left) and paired-end FASTQ files (on the right).
When you have multiple FASTQ files, each file (or two files) for an individual sample,
you will need to create a manifest file for importing those FASTQ files into QIIME2 arti-
fact. The single-end FASTQ files and paired-end FASTQ files are imported, respectively,
as follows:

qiime tools import \

--type ‘SampleData[SequencesWithQuality]’ \
--input-format SingleEndFastqManifestPhred33 \
--input-path data \
--output-path artifacts/xxxx.qza
 Targeted Gene Metagenomic Data Analysis   ◾    269

FIGURE 7.2 Examples of manifest files for single-end and paired-end FASTQ files.

qiime tools import \

--type ‘SampleData[PairedEndSequencesWithQuality]’ \
--input-format PairedEndFastqManifestPhred33 \
--input-path data \
--output-path artifacts/xxxx.qza

The artifacts created by the above commands are demultiplexed and they are ready for
processing by QIIME2.

7.3.1.2 Metadata
In the above, we discussed importing the raw sequence data into QIIME2 artifacts. But we
have also mentioned that a sample metadata describing the study is required for the analy-
sis. A metadata file can also be created by the user and will be used later in the analysis to
show the sample grouping. A metadata file is a tab-separated value (TSV) file containing
sample and study design information in columns and the first column must be the sample
identifiers for the IDs of the sample in the study. The top line is for the column names. The
ID column name must be one of those: id, sampleid, sample id, sample-id, featured, fea-
ture id, or feature-id. The IDs listed in the metadata files should be unique and not empty.
The ID column is not optional and the metadata file must contain at least one ID. The ID
column can be followed by additional columns defining metadata associated with each
sample or feature such as groups, treatment, factor, and barcode sequence. The column
name should not be one of the reserved words and the type of the data contained in the
metadata file is either categorical or numeric. The sample metadata is created by the user
manually. However, if the data is downloaded from the NCBI SRA database, the metadata
can also be downloaded and modified to meet our goal. This will be discussed below in the
worked example.

7.3.2 Demultiplexing
Above, we discussed the EMP and non-EMP-multiplexed FASTQ files and how they can
be imported into QIIME2 artifacts. Since the reads are multiplexed, demultiplexing must
be carried out before reads processing and analysis. QIIME2 uses “demux” and “cutadapt”
plugins for demultiplexing EMP-multiplexed reads and non-EMP-multiplexed reads,
respectively.
270   ◾    Bioinformatics

qiime demux emp-single  # EMP-multiplexed single-end reads

qiime demux emp-paired  # EMP-multiplexed paired-end reads
qiime cutadapt demux-single 
# non-EMP-multiplexed single-end reads
qiime cutadapt demux-paired # non-EMP-multiplexed paired-end reads

The following is a format for demultiplexing EMP-multiplexing single-end reads. Notice

that the input for the command is an artifact that is created by importing a FASTQ file as
discussed above.

FIGURE 7.3 Demultiplexing plugin for different multiplexed data types.

 Targeted Gene Metagenomic Data Analysis   ◾    271

qiime demux emp-single \

--i-seqs artifacts/multiplexed-emp-single-end.qza \
--m-barcodes-file data/sample-metadata.tsv \
--m-barcodes-column barcode-sequence \
--o-per-sample-sequences artifacts/demux-single-end.qza \
--o-error-correction-details artifacts/demux-details.gza

It is also required to provide the sample metadata file as an input “m-barcodes-file” and the
column which includes the barcode sequence in the metadata to be specified. The output is
two artifacts: an artifact for the demultiplexed reads and an artifact for the demultiplexing
details.
Figure 7.3 shows the commands and usages for demultiplexing both file formats.
Once the raw data is imported into QIIME2 artifact and the multiplexed reads were
demultiplexed, all types of raw data will be preprocessed and analyzed in the same way.
Therefore, we will discuss the remaining steps of the analysis with QIIME2 through a
worked example.

7.3.3 Downloading and Preparing the Example Data

As an example, we will download sequence raw data from the NCBI SRA database. The
data is for amplicon-based 16S rRNA gene sequences obtained from NGS for a study to
examine the effect of a yoga-based intervention against a low-FODMAP diet on patients
with irritable bowel syndrome. FODMAP is an acronym for Fermentable Oligosaccharides,
Disaccharides, Monosaccharides, and Polyols, which are short-chain carbohydrates and
poorly absorbed in the small intestine. The metagenomic 16S rRNA gene data sequenced
from fecal samples are available for 86 patients, with irritable bowel syndrome, grouped
into (i) patients who received yoga sessions and (ii) patients who received low-FODMAP
diet. The NCBI BioProject accession for this study is PRJEB24421. We will download the
FASTQ files from the NCBI SRA database and then we will follow through the QIIME2
pipeline to analyze these data step by step. The data are for demultiplexed paired-end
sequences: two FASTQ files (forward and reverse) for each sample.

7.3.3.1 Downloading the Raw Data

To download the files of all experiments, we need to obtain the run accessions of the
­experiments in the BioProject. To keep files organized, we will create a directory to store
the raw data of this project. Open the Linux terminal and create a directory with the
BioProject accession, and inside that directory, create a subdirectory with the name “data”
as follows:

mkdir PRJEB24421
cd PRJEB24421
mkdir data

In the next step, we will save the run accessions of the BioProject in a text file in the
“data” subdirectory. To do that, you can open the NCBI SRA database and search for the
272   ◾    Bioinformatics

BioProject by the above accession number or simply copy and paste the following URL on
the Internet browser:
https://www.ncbi.nlm.nih.gov/sra/?term=PRJEB24421
Then, use “Send to” dropdown menu to download the runinfo text file. After download-
ing the text file, open the file in Excel, delete all columns except the column with the run
accessions and remove the column name as well, and save the file as “runids.txt” in the
“data” subdirectory.
Instead of the above, you can also use the following EDirect script, which extracts the
run accessions and stores them in a file named “runids.txt” in “data” subdirectory (you
should have the NCBI Entrez Direct installed):

esearch -db sra -query ‘PRJEB24421[bioproject]’ \

| efetch -format runinfo \
| cut -f1 -d, > data/runids.txt
sed -i ‘/^$/d’ data/runids.txt
sed -i ‘/^Run/d’ data/runids.txt

Check to see if the file has been saved successfully by using “ls data/” command or you can
display the file content by using “vim data/runids.txt” command.
After saving the text file with the 86 run accessions in the “data/runids.txt” file, you
can then download the raw FASTQ files from the NCBI SRA database either by saving
the following script in a bash file “download.sh” and then run it as “bash download.sh” or
you can just enter the script on the terminal command-line prompt, while you are in the
project directory:

while read f;
do
fasterq-dump --progress --outdir data “$f”
done < data/runids.txt

You will see the downloading progress. The files require only 771.29MB of storage space.
The 172 FASTQ files will be downloaded in the “data” subdirectory, two files for each
sample. When the files have been downloaded successfully, you can check the content
of the “data” subdirectory and count the number of the FASTQ files using the following
command:

ls data/*.fastq | wc -l

The number of files should be 172. If it is not, you may need to run the download script
again.

7.3.3.2 Creating the Sample Metadata File

Open the NCBI SRA using the above URL. Then, open Run Selector from “Send to” drop-
down menu. All runs will be displayed on the Run Selector. Click “Metadata” button
 Targeted Gene Metagenomic Data Analysis   ◾    273

FIGURE 7.4 Downloading metadata from SRA database using Run Selector.

FIGURE 7.5 The sample metadata text file “sample-metadata.tsv”.

(Figure 7.4) to download the metadata text file. Open the file in a spreadsheet, remove all
columns except “run”, “group”, and “time”. Then, rename “run” to “SampleID”, insert a
data type row as shown in Figure 7.5, and save the file as a TSV file with the name: “data/
sample-metadata.tsv”.
If you have EDirect installed on your computer, you can search the NCBI database and
retrieve metadata from it. EDirect is a collection of Linux-based command-line E-Utilities
programs. The instructions for installing EDirect programs are available at the NCBI
274   ◾    Bioinformatics

website at “https://dataguide.nlm.nih.gov/edirect/install.html”. You can use the following

script to use EDirect to create a metadata file. The script searches the NCBI SRA database
for the BioProject “PRJEB24421” and then it retrieves the sample metadata and stores them
in a TSV file “sample-metadata.tsv”.

esearch -db sra -query ‘PRJEB24421[bioproject]’ \

| efetch -format runinfo \
| tr -s ‘,’ ‘\t’ > sample-metadata.tsv

Then, you can edit the file as above.

7.3.3.3 Importing Microbiome Yoga Data

Our example raw data is non-Casava 1.8 demultiplexed reads. To import the FASTQ files
into QIIME2 artifact, we need a manifest file listing the file names and their absolute path
as described above. We can create the manifest file manually or we can use the following
bash script. Before running the script, change to the “data” directory “cd data”, where the
FASTQ files are found.

#Creating a manifest file

###############################
#a- make file name and absolute path
find “$PWD”/*.fastq -type f -printf ‘%f %h/%f\n’ > tmp.txt
#b- remove _1/2.fastq
awk ‘{ gsub(/_[12].fastq/,”,”, $1); print } ‘ tmp.txt > tmp2.txt
#remove space
cat tmp2.txt | sed -r ‘s/\s+//g’ > tmp3.txt
n=$(ls -l *1.fastq|wc -l)
#create a direction column
seq $n | sed “c forward\nreverse” > tmp4.txt
#add direction column
paste tmp3.txt tmp4.txt | column -s $’’ -t > tmp5.txt
#replace space with comma
sed -e ‘s/\s\+/,/g’ tmp5.txt > manifest.txt
#add column names
sed -i ‘1s/^/sample-id,absolute-filepath,direction\n/’ manifest.
txt
rm tmp*.txt

The “manifest.txt” file will be created in “data” directory, and it looks as shown in
Figure 7.6.
After running the above script, you can display the file content using the text editor of
your choice. Then, move back to the project main directory using “cd ..”.
The next step is to import the FASTQ files into a QIIME2 artifact. To keep files orga-
nized, you can create a new subdirectory “input” for the artifact files.

mkdir inputs
 Targeted Gene Metagenomic Data Analysis   ◾    275

FIGURE 7.6 Manifest file for the example FASTQ files.

The following script uses the “qiime tools import” to import the 172 FASTQ files onto a
single QIIME2 artifact “demux-yoga.qza” in the “inputs” directory:

qiime tools import \

--type ‘SampleData[PairedEndSequencesWithQuality]’ \
--input-format PairedEndFastqManifestPhred33 \
--input-path data/manifest.txt \
--output-path inputs/demux-yoga.qza

We will use this artifact “demux-yoga.qza” in the next step.

7.3.4 Raw Data Preprocessing

All the raw FASTQ files are contained in a single QIIME2 artifact. The next step would be
data preprocessing, which includes quality control and denoising.

7.3.4.1 Quality Assessment and Quality Control

QIIME2 has plugins for quality assessment and quality control. We can use “q2-demux
summarize” to assess the quality of the raw data in the artifact. The output of that com-
mand is a visualization file containing the quality assessment report that can be viewed
on an Internet browser. To keep the files organized, we can create a subdirectory to the
visualization file called it “viz”.

mkdir viz

Then, we can run the following script to save the report visualization file in the “viz”
subdirectory:

qiime demux summarize \

--i-data inputs/demux-yoga.qza \
--o-visualization viz/demux-yoga-qc.qzv
276   ◾    Bioinformatics

We can view the visualization file on the Internet browser using “q2-tools view” as follows:

qiime tools view viz/demux-yoga-qc.qzv

The report displayed on the Internet browser will have two tabs: Overview and Interactive
Quality Plot. The Overview report tab shows Demultiplexed sequence count summary,
forward and reverse reads frequency histogram, and Per-sample sequence counts. The
summary count section shows minimum, median, mean, maximum, and total number
of reads across the samples for forward reads and reverse reads (this is only in paired
end). Figure 7.7 shows that the summary statistics for forward reads and reverse reads are
the same. The minimum number of reads in a sample is 7702 and the maximum num-
ber of reads is 95075. The median and mean of reads across all samples are 25588.0 and
30018.709302, respectively. The total number of forward reads in all samples is 2581609
and is the same for reverse reads. The histogram shows the distribution of the number
of reads versus number of samples. The histograms can be downloaded as pdf files. The
Per-sample sequence counts table, below the histogram, lists sample ID and number of
forward reads and reverse reads in each sample. The Interactive Quality Plot section shows
plots for the per base Phred quality score across reads for both forward and reverse reads
(Figure 7.8). These plots are similar to the QC plots discussed in Chapter 1. They will help
us to determine if the sequences need trimming or filtering. If you hover the mouse over a
specific position on the plot, the table below the plot will update itself to display the seven-
number summary statistics for that base position. This table corresponds to what is visu-
ally represented by the box plot at that position.

FIGURE 7.7 Sequence count summary and histograms of reads of yoga data.
 Targeted Gene Metagenomic Data Analysis   ◾    277

FIGURE 7.8 Per base quality plots of the yoga data.

You can notice that the overall quality scores of the reads are high but there are also
some reads with quality score less than 20 (99% accuracy) toward the end of the reads.
We can trim the low-quality bases from the end of the reads. The demultiplexed sequence
length summary at the bottom of the Interactive Quality Plot tab shows that the reads have
equal length (275 base). This table will help us to determine if we need to make the length
of reads equal or not.
If you decide to filter out the reads with poor quality scores, you can use the “qual-
ity-filter” plugin with “q-score” method. However, this can be done for single-end reads.
For paired-end reads, you can join forward and reverse reads and then run “quality-fil-
ter q-score” on the merged reads. This will be discussed later with clustering. However,
denoising methods also have their way to filter low-quality reads as we will see soon. But,
if your data is single-end reads, you can use “quality-filter q-score” to remove low-quality
reads using the following script:

qiime quality-filter q-score \

--i-demux demux.qza \
--p-min-quality 20 \
--p-quality-window 5 \
--p-min-length-fraction 0.8 \
--p-max-ambiguous 0 \
--o-filtered-sequences demux-filtered.qza \
--o-filter-stats demux-filter-stats.qza

The default settings are “--p-min-quality 4”, “--p-quality-window 3”, “--p-min-length-frac-

tion 0.75”, and “--p-max-ambiguous 0”, if those parameters are not included in the above.
For more information about these parameters, use “qiime quality-filter q-score --help”. We
will discuss this in more detail with clustering.
If there are PCR primer sequences or any other non-biological sequences, you can
remove them using “cutadapt” plugin. Once more, this is not applicable to our yoga data
but, just in case, if you had sequences with primers at this stage of the analysis, it would
278   ◾    Bioinformatics

be the right moment to remove them using “cutadapt” plugin with “trim-single” or “trim-
paired” for single-end or paired-end reads, respectively.

qiime cutadapt trim-single \

--i-demultiplexed-sequences demux.qza \
--p-front GTGCCAGCMGCCGCGGTAA \
--p-error-rate 0 \
--o-trimmed-sequences trimmed-demux.qza \
--verbose

About our yoga data, now we have a clear view about its quality after we have assessed it
using “demux summarize”. Since the data is paired-end reads, we will carry out the quality
control later with clustering and denoising.

7.3.4.2 Clustering and Denoising

After the above preprocessing, the next step in the data analysis is to create features by
either clustering or denoising as discussed above. QIIME2 supports de novo, closed-refer-
ence, and open-reference clustering using VSEARCH plugin and denoising with DADA2
and deblur. Either clustering or denoising is performed to create feature tables and rep-
resentative sequences. Denoising (with DADA2 or deblur) attempts to remove the noises
generated from errors. The features generated by DADA2 and deblur are also called ampli-
con sequence variant (ASVs). Whatever you choose to continue with clustering or with
denoising, it is your sole choice. These techniques were discussed above in detail. In the
following, we will show you how to perform clustering and denoising with QIIME2.

7.3.4.2.1 Clustering
If your plan is to cluster reads into OTUs without denoising, QIIME2 provides “q2-vsearch”
plugin to do just that. This plugin has methods for the three types of clustering: de novo,
closed-reference, and open-reference. The “q2-vsearch” plugin can also perform quality
control; therefore, before running clustering, you may need to do some preprocessing to
the data. The paired-end reads must be merged before processing. In the following, we will
walk you through the steps of clustering to the point of generating feature tables and OTU
representative sequences.

7.3.4.2.1.1 Merging Paired-End Reads

If the data is paired-end reads, the forward and reverse reads must be merged before clus-
tering. The merging is achieved with “join-pairs” method of “q2-vsearch” plugin.
The artifact “demux-yoga.qza” of our example data is in the “inputs” directory. Since the
reads are paired end, we can merge them before clustering. The following script takes the
“demux-yoga.qza” artifact as an input, joins the forward and reverse reads, and creates a
new artifact for the merged reads “demux-yoga-merged.qza”:

qiime vsearch join-pairs \

--i-demultiplexed-seqs inputs/demux-yoga.qza \
 Targeted Gene Metagenomic Data Analysis   ◾    279

--p-allowmergestagger \
--o-joined-sequences inputs/demux-yoga-merged.qza

For checking the merging, run “demux summarize” to create a visualization file for the
summary and quality of the merged reads as discussed above.

qiime demux summarize \

--i-data inputs/demux-yoga-merged.qza \
--o-visualization viz/demux-yoga-merged-qc.qzv
qiime tools view viz/demux-yoga-merged-qc.qzv

If you open Interactive quality plot on the summary report, you will notice that the for-
ward and reverse reads have been joined as shown in Figure 7.9.
If, at this point, you notice that the forward reads (for single-end data) or merged reads
(for paired-end data) require filtering, you can then perform that with “quality-filter
q-score” as follows:

qiime quality-filter q-score \

--i-demux inputs/demux-yoga-merged.qza \
--o-filtered-sequences inputs/demux-yoga-merged-filter.qza \
--o-filter-stats inputs/demux-yoga-merged-filter-stats.qza

7.3.4.2.1.2 Sequence Dereplication

Before clustering, the reads must be dereplicated to reduce the redundancy of the reads
based on similarity. The dereplication is carried out with “dereplicate-sequences” method

FIGURE 7.9 Per base quality plots of the merged yoga data.
280   ◾    Bioinformatics

of “q2-vsearch” plugin. The input is the last preprocessed data artifact “demux-yoga-
merged.qza”.

qiime vsearch dereplicate-sequences \

--i-sequences inputs/demux-yoga-merged.qza \
--o-dereplicated-table inputs/derep-yoga-table.qza \
--o-dereplicated-sequences inputs/derep-yoga-seqs.qza

The outputs from the “dereplicate-sequences” command are two artifacts: (i) feature table
containing the OTU features with their observed abundances (frequencies) for each of the
samples of the study and (ii) feature data in which each feature identifier is mapped to a
feature.

7.3.4.2.1.3 Clustering Methods

Clustering then follows the dereplication. Both feature table and feature data artifacts gen-
erated with the dereplication are required as inputs for clustering. We will use QIIME2
to perform the three types of clustering (de novo, closed-reference, and open-reference
clustering).

7.3.4.2.1.3.1 De Novo Clustering

The de novo clustering does not require a database but it uses sequence similarity to cluster
features into groups. The threshold for similarity can be set to 0.99% (only reads similar
to centroid sequence with identity of 99% are allowed to join the cluster). QIIME2 uses
“cluster-features-de-novo” method of “q2-vsearch” plugin to perform the de novo cluster-
ing. The input artifacts are the feature table and feature data artifacts generated by the
dereplication in the previous step. To keep the clustering in a separate directory, we will
create the “denovo” subdirectory.

mkdir denovo
qiime vsearch cluster-features-de-novo \
--i-table inputs/derep-yoga-table.qza \
--i-sequences inputs/derep-yoga-seqs.qza \
--p-perc-identity 0.99 \
--o-clustered-table denovo/table-yoga-denovo.qza \
--o-clustered-sequences denovo/rep-seqs-yoga-denovo.qza

The outputs are two artifacts: a feature table for the OTUs and feature data that contains
the centroid sequences defining each OTU cluster. De novo clustering usually consumes
more computational resources compared to the other two methods.

7.3.4.2.1.3.2 Closed-Reference Clustering

Closed-reference clustering requires a curated database for the16S rRNA gene sequences as
reference sequences. Only the representative sequences that have matches on the database
are clustered, while the ones that do not have matches will be discarded. Examples of widely
used databases include Greengenes (16S rRNA) at “https://greengenes.secondgenome.
 Targeted Gene Metagenomic Data Analysis   ◾    281

com”, Silva (16S/18S rRNA) at https://www.arb-silva.de, and UNITE (fungal ITS) at

“https://unite.ut.ee/”. The database must be downloaded and then imported in QIIME2 as
artifact before being used for clustering. For example, we will download the latest release
of curated OTUs from GreenGenes database:

wget ftp://greengenes.microbio.me/greengenes_release/gg_13_5/
gg_13_8_otus.tar.gz
tar vxf gg_13_8_otus.tar.gz
rm gg_13_8_otus.tar.gz

Make sure that the URL is a single line with no white space. You can visit the website to
download the latest release.
The files will be extracted into a directory (gg_13_8_otus). Display the contents of this
directory and its subdirectories using “ls” Linux command. You will find four subdirec-
tories: “otus” (for reference OTUs), “rep_set” (for the reference representative sequences),
“rep_set_aligned” (for aligned representative sequences), “taxonomy” (for taxonomy
files), and “trees” (for phylogenetic trees). The files in these directories contain data at dif-
ferent identities (e.g., 99%, 97%, and 94%). Keep this database as we will use it for other
applications.
To use the reference database for clustering, you need to import the file of the database
representative sequences (FASTA file). You need to choose at which identity you wish to
perform clustering. Assume that you want to cluster your sample sequences at 97% iden-
tity, then you can import “rep_set/97_otus.fasta” onto QIIME2 as artifact using “tools
import”. To keep the files organized, we will create the subdirectory “closed_ref_cl_97” for
closed-reference clustering files.

mkdir closed_ref_cl_97

Then, import the database representative sequences into QIIME2 artifact.

qiime tools import \

--type ‘FeatureData[Sequence]’ \
--input-path gg_13_8_otus/rep_set/97_otus.fasta \
--output-path inputs/97_otus-GG_db.qza

Then, you can use the “cluster-features-closed-reference” method of the “q2-vsearch” plu-
gin to perform the closed-reference clustering on the features generated in the derepli-
cation steps. The input artifacts are: dereplicated feature table “derep-yoga-table.qza”,
dereplicated representative sequences “derep-yoga-seqs.qza”, and the reference representa-
tive sequences from the database “97_otus-GG_db.qza”.

qiime vsearch cluster-features-closed-reference \

--i-table inputs/derep-yoga-table.qza \
--i-sequences inputs/derep-yoga-seqs.qza \
282   ◾    Bioinformatics

--i-reference-sequences inputs/97_otus-GG_db.qza \
--p-perc-identity 0.97 \
--o-clustered-table closed_ref_cl_97/table-yoga-closed_cl.qza \
--o-clustered-sequences closed_ref_cl_97/rep-seqs-yoga-close_
cl.qza \
--o-unmatched-sequences closed_ref_cl_97/
unmatched-yoga-close_cl.qza

The above script outputs three artifacts: A feature table, clustered sequences (the sequences
defining the features in the feature table), and unmatched sequences (the sequences that
didn’t match reference sequences at 97% identity). The unmatched sequences will be com-
pletely ignored.

7.3.4.2.1.3.3 Open-Reference Clustering

The open-reference clustering is hybrid of the above two clustering methods. First, it uses
reference sequences for clustering the matched sequences and then it performs de novo
clustering on the unmatched sequences. The open-reference clustering is performed with
“cluster-features-open-reference” method. The input and output artifacts are the same
as that of the closed-reference clustering except that there are no unmatched sequences;
instead, there is an artifact for the new reference sequences used as an input in addition to
the sequences clustered as part of the internal de novo clustering step. We will create the
new subdirectory “open_ref_cl_97” for files of the open-reference clustering.

mkdir open_ref_cl_97
qiime vsearch cluster-features-open-reference \
--i-table inputs/derep-yoga-table.qza \
--i-sequences inputs/derep-yoga-seqs.qza \
--i-reference-sequences inputs/97_otus-GG_db.qza \
--p-perc-identity 0.97 \
--o-clustered-table open_ref_cl_97/table-yoga-open_cl.qza.qza \
--o-clustered-sequences open_ref_cl_97/rep-seqs-yoga-open_cl.qza \
--o-new-reference-sequences open_ref_cl_97/
new-ref-seqs-open_cl.qza

The three clustering methods use dereplicated feature table and representative sequences
and produce a final feature table and OTU representative sequences to be used in the
downstream analysis for phylogeny, diversity analysis, assignment of taxonomic group,
and differential taxonomic analysis.

7.3.4.2.2 Denoising
Like clustering, denoising also produces a feature table and representative sequences.
However, denoising attempts to remove errors and to provide more accurate results.
There are two denoising methods available in QIIME2: DADA2 and deblur. Both meth-
ods output feature tables containing feature abundances and ASVs. Moreover, they also
 Targeted Gene Metagenomic Data Analysis   ◾    283

perform quality filtering and chimera removal. You may not need to perform any quality
control prior using any of these two methods. Only for deblur denoising, you may need to
merge paired-end reads as we did for the clustering.

7.3.4.2.2.1 Denoising with DADA2

The “q2-dada2” plugin is used to denoise single-end and paired-end reads (no need for
read merging). Use “denoise-single” method for the single-end reads and “denoise-paired”
method for the paired-end reads. You can always use “--help” with any of the methods to
display the usage and options. DADA2 methods denoise sequences, dereplicate them, and
filter chimeras.

qiime dada2 denoise-single --help

qiime dada2 denoise-paired --help

Now, we can use “q2-dada2” to denoise the yoga data that we imported and saved as
“demux-yoga.qza”. We will use “denoise-paired” method. To keep the files organized, we
will create the “dada2” subdirectory for DADA2 denoising files.

mkdir dada2
qiime dada2 denoise-paired \
--i-demultiplexed-seqs inputs/demux-yoga.qza \
--p-trim-left-f 0 \
--p-trim-left-r 0 \
--p-trunc-len-f 250 \
--p-trunc-len-r 250 \
--p-n-threads 4 \
--o-representative-sequences dada2/rep-seqs_yoga_dada2.qza \
--o-table dada2/table_yoga_dada2.qza \
--o-denoising-stats dada2/stats_yoga_dada2.qza

The parameters “--p-trim-left-f”, “--p-trim-left-r”, “--p-trunc-len-f”, and “--p-trunc-len-r”

are optional, and they are used to trim and truncate the forward and reverse sequences,
respectively, to improve the quality of the reads if required. If you use “--p-trunc-len-f 0”
and “--p-trunc-len-r 0”, the truncation will be disabled. The parameter “--p-n-threads”
specifies the number of threads used for denoising. If the “denoise-single” method is used
with paired-end reads instead of “denoise-paired”, only forward reads will be used as input
while the reverse reads will be ignored.
We set “--p-trunc-len-f 250” and “--p-trunc-len-r 250” to truncate the forward and
reverse reads to 250 bases. However, we did not trim the left ends of the reads because they
do not need trimming. Like clustering, DADA2 feature table and representative sequences
artifacts are used for the downstream analysis for phylogeny, diversity analysis, taxonomy
assignment, etc. Moreover, the DADA2 stats summary artifact contains useful information
regarding the filtering and denoising. Users can use “q2-metadata tabulate” to generate a
visualization file that can be displayed on the Internet browser with “tools view” as follows:
284   ◾    Bioinformatics

qiime metadata tabulate \

--m-input-file dada2/stats_yoga_dada2.qza \
--o-visualization dada2/stats_yoga_dada2.qzv
qiime tools view dada2/stats_yoga_dada2.qzv

The DADA2 stats summary table is a TSV. It shows the ID, number of input reads, number
of reads filtered, percentage of reads passed the filter, number of denoised reads, number of
merged reads, percentage of reads merged, number of non-chimeric reads, and percentage
of non-chimeric reads for each sample in the study.

7.3.4.2.2.2 Denoising with Deblur

Deblur plugin denoises single-end reads and merged paired-end read. Paired-end sequences
must be merged using “vsearch join-pairs”, as shown above, before being denoised by
deblur.

mkdir deblur
qiime vsearch join-pairs \
--i-demultiplexed-seqs inputs/demux-yoga.qza \
--p-allowmergestagger \
--o-joined-sequences deblur/demux-yoga-merged.qza

It is recommended to run quality assessment and quality control by filtering low-quality

reads using “quality-filter q-score” before denoising with deblur.

qiime quality-filter q-score \

--i-demux inputs/demux-yoga-merged.qza \
--o-filtered-sequences deblur/demux-yoga-merged-filtered.qza \
--o-filter-stats deblur/demux-yoga-merged-filtered-stat.qza

Denoising with deblur is carried out with “q2-deblur” plugin, which has two methods:
“denoise-16S” for denoising 16S rRNA gene sequences and “denoise-other” for denoising
other types of targeted gene sequences. When you use “denoise-16S”, deblur will perform
an initial positive filtering step by discarding any reads that do not have a minimum 60%
identity similarity to database sequences from the 85% GreenGenes.
You can use “deblur denoise-16S” method, which has “--p-trim-length” parameter that
can be set to truncate the reads at specific position for quality filtering or you can set that
parameter to -1 to disable trimming. The following script will perform deblur denoising:

mkdir deblur
qiime deblur denoise-16S \
--i-demultiplexed-seqs deblur/demux-yoga-merged.qza \
--p-trim-length -1 \
--p-jobs-to-start 4 \
--p-sample-stats \
--o-representative-sequences deblur/rep-seqs_yoga_deblur.qza \
 Targeted Gene Metagenomic Data Analysis   ◾    285

--o-table deblur/table_yoga_deblur.qza \
--o-stats deblur/stats_yoga_deblur.qza

The “--p-jobs-to-start” is an optional parameter and you can set it to the number of jobs to
run in parallel. To learn about more options, use “qiime deblur denoise-16S --help”.
The feature table and representative sequences artifacts will be used in the downstream
analysis.
The deblur stats summary artifact contains useful information about the filtering and
denoising. We can use the “deblur visualize-stats” plugin to generate a visualization file
(Figure 7.10).

qiime deblur visualize-stats \

--i-deblur-stats deblur/stats_yoga_deblur.qza \
--o-visualization deblur/stats_yoga_deblur.qzv
qiime tools view deblur/stats_yoga_deblur.qzv

Both clustering and denoising (with DADA2 or deblur) generate feature table and repre-
sentative sequences artifacts that can be used in the following analysis steps. You may need
to visualize these feature table and the sequence data artifacts. The “q2-feature-table” plu-
gin is used just for that. The “summarize” and “tabulate-seqs” methods are used to create a
visualization file for the feature table and representative sequences, respectively. As exam-
ples, the following script creates visualization files for the feature tables produced by the
de novo clustering and DADA2 denoising, respectively. The “summarize” method of the
“feature-table” plugin takes a feature table artifact and the sample metadata file as input
and it outputs a visualization file. We created the sample metadata file “sample-metadata.
tsv” earlier and we saved it in the “data” subdirectory.

FIGURE 7.10 Deblur denoising stats summary.

286   ◾    Bioinformatics

qiime feature-table summarize \

--i-table denovo/table-yoga-denovo.qza \
--m-sample-metadata-file data/sample-metadata.tsv \
--o-visualization denovo/table-yoga-denovo.qzv
qiime tools view denovo/table-yoga-denovo.qzv

qiime feature-table summarize \

--i-table dada2/table_yoga_dada2.qza \
--m-sample-metadata-file data/sample-metadata.tsv \
--o-visualization dada2/table_yoga_dada2.qzv
qiime tools view dada2/table_yoga_dada2.qzv

The html report of the feature table visualization, as displayed on the Internet browser
(Figure 7.11), has three tabs: Overview, Interactive Sample Details, and Feature Details
tab. Overview tab contains Table summary, Frequency per sample with a histogram, and
Frequency per feature with a histogram. Interactive Sample Detail tab (Figure 7.12) displays
a plot, a table containing sample IDs and feature counts, and Plot controls to control the
plot interactively by changing Metadata Category using the dropdown list and sampling
depth using Sampling Depth sliding bar. The metadata category list is obtained from the
sample metadata file which describes the sample groups. Feature Details tab (Figure 7.13)
shows the list of the features found in the samples. The table includes feature id, feature
frequency (abundance), and number of samples in which that feature is observed.
The “feature-table summarize” methods are used to visualize any feature tables gener-
ated by any of the clustering or denoising methods.
We can also use “tabulate-seqs” to create a visualization file for the representative
sequence artifact generated by any of the clustering or denoising methods.

FIGURE 7.11 The Stats summary of the feature table.

 Targeted Gene Metagenomic Data Analysis   ◾    287

FIGURE 7.12 Interactive sample detail of the feature table.

FIGURE 7.13 Feature detail of feature table.

The following script creates a visualization file for the representative sequence artifact
produced with deblur denoising:

qiime feature-table tabulate-seqs \

--i-data deblur/rep-seqs_yoga_deblur.qza \
--o-visualization deblur/rep-seqs_yoga_deblur.qzv
qiime tools view deblur/rep-seqs_yoga_deblur.qzv

The tabular report of the representative sequences (Figure 7.14), as displayed on the Internet
browser, contains Sequence Length Statistics, Seven-Number Summary of Sequence
Lengths, and Sequence Table that includes ID, sequence length, and sequence for each fea-
ture. Moreover, each feature sequence is linked to a BLAST report showing the sequences
with significant alignments to the feature sequence.
288   ◾    Bioinformatics

FIGURE 7.14 The tabular report of the amplicon sequence variants (ASVs).

After you study the feature table visualization on the Internet browser, you may decide
to remove some samples or features from the feature table because of outliers or low abun-
dance. The “q2-feature-table” plugin has the “filter-samples” and “filter-features” methods
for these purposes. For example, you can remove samples based on their minimum total
frequency. The following script removes the samples with a total frequency less than 1000
from feature table created with DADA2 denoising and then it creates a visualization and
views it on the Internet browser:

qiime feature-table filter-samples \

--i-table dada2/table_yoga_dada2.qza \
--p-min-frequency 1000 \
--o-filtered-table dada2/table_sample_freq_filtered_yoga_dada2.
qza
qiime feature-table summarize \
--i-table dada2/table_sample_freq_filtered_yoga_dada2.qza \
--m-sample-metadata-file data/sample-metadata.tsv \
--o-visualization dada2/table_sample_freq_filtered_yoga_dada2.qzv
qiime tools view dada2/table_sample_freq_filtered_yoga_dada2.qzv

Try to study the feature table report after the above filtering.
The “filter-features” method is used to remove low abundance features from a feature
table. The following script removes features with a total abundance (across all samples) of
less than 20:

qiime feature-table filter-features \

--i-table dada2/table_sample_freq_filtered_yoga_dada2.qza \
--p-min-frequency 20 \
--o-filtered-table dada2/table_feat_sample_freq_filtered_yoga_
dada2.qza
 Targeted Gene Metagenomic Data Analysis   ◾    289

qiime feature-table summarize \

--i-table dada2/table_feat_sample_freq_filtered_yoga_dada2.qza \
--m-sample-metadata-file data/sample-metadata.tsv \
--o-visualization dada2/table_feat_sample_freq_filtered_yoga_
dada2.qzv
qiime tools view dada2/table_feat_sample_freq_filtered_yoga_dada2.
qzv

You can use “qiime feature-table filter-samples --help” and “qiime feature-table filter-­
features --help” to learn more about filtering parameters.
The steps of the preprocessing of raw data that we followed above will end up with
creation of feature tables and features (OTUs/ASVs) that we will rely on to move to the
downstream analysis (taxonomic classification, phylogenetic relationship, alpha and beta
diversity analysis, and differential abundance). There are always questions that come up at
this point: Which method is better (clustering or denoising)? Which clustering method is
the best (de novo, closed-reference, or open-reference clustering)? And, which denoising
method is better (DADA2 or deblur)? The right answer from many experts is that: try all of
them and adopt the one that works for you. There are number of articles that discussed the
pros and cons of each of these methods.

7.3.5 Taxonomic Assignment with QIIME2

As discussed above, one of the main goals of metagenomic studies is to identify the micro-
bial organisms present in a sample using any of the alignment-based classifiers or machine
learning classifiers. QIIME2 has alignment-based classifiers, machine learning classifiers,
and hybrid classifier methods contained in the “q2-feature-classifier” plugin. The alignment-
based classifier methods are “classify-consensus-blast” and “classify-consensus-vsearch”.
The machine learning classifier method is “classify-sklearn”, which is used for pre-fitted
sklearn-based taxonomy classifiers (any of the classifiers available in scikit-learn python
package). You can download a shared pre-fitted classifier; however, it is safer to train yours
and use it for taxonomy assignment. Some pre-fitted naïve bayes classifiers and weighted
taxonomic classifiers are available at the QIIME2 data resources web page at “https://docs.
qiime2.org/2022.2/data-resources/” or any newer release. To train your own taxonomy
classifier, you can use any of the two training methods provided by “q2-feature-classifer”
or “fit-classifier-naive-bayes” to train a naïve bayes classifier or “fit-classifier-sklearn” to
train any arbitrary scikit-learn classifier. An alpha hybrid classifier (VSEARCH + sklearn
classifier) is provided with “classify-hybrid-vsearch-sklearn” method.

7.3.5.1 Using Alignment-Based Classifiers

The alignment-based classifiers use each of representative sequences generated with
clustering or denoising as query sequence to search against the database representative
sequences whose taxa are known. The taxonomy assignment is based on the consensus of
the BLAST hits at percent identity greater than a predetermined identity threshold. The
taxonomy ranks, from the highest to the lowest, are kingdom, phylum, class, order, family,
genus, and species. The confidence on the assignment is the fraction of top hits that match
290   ◾    Bioinformatics

the consensus taxonomy at a given taxonomic level. Consensus is determined at each tax-
onomic level, beginning from kingdom and stopping when consensus is no longer met
above a specified minimum consensus value. Then, the taxonomy is trimmed at this point.
To use an alignment-based classifier, you need to download, extract, and import a rep-
resentative sequence database and reference taxonomy database as shown above in the de
novo clustering section. The following script imports “99_otus.fasta” and “99_otu_taxon-
omy.txt” reference database from Greensgenes into QIIME2 artifacts, which will be saved
in the “input” subdirectory. To keep the files organized, we will create the “taxonomy”
subdirectory to store the taxonomy files.

mkdir taxonomy
qiime tools import \
--type ‘FeatureData[Sequence]’ \
--input-path gg_13_8_otus/rep_set/99_otus.fasta \
--output-path taxonomy/99_otus.qza
qiime tools import \
--type ‘FeatureData[Taxonomy]’ \
--input-format HeaderlessTSVTaxonomyFormat \
--input-path gg_13_8_otus/taxonomy/99_otu_taxonomy.txt \
--output-path taxonomy/99_otu_taxonomy.qza

Now, you can run taxonomy classification using the BLAST-based classifier (classify-
consensus-blast). The inputs are any representative sequence artifact (generated by any
of the clustering methods or denoising methods), the reference sequence artifact, and
the reference taxonomy artifact imported in the previous step. The following script uses
BLAST-based classifier to assign taxa to the representative sequences generated by DADA2
denoising:

qiime feature-classifier classify-consensus-blast \

--i-query dada2/rep-seqs_yoga_dada2.qza \
--i-reference-reads taxonomy/99_otus.qza \
--i-reference-taxonomy taxonomy/99_otu_taxonomy.qza \
--p-perc-identity 0.97 \
--o-classification taxonomy/blast_tax_yoga_dada2.qza \
--verbose

If you wish to use the VSEARCH-based method instead, you can use “classify-consensus-
vsearch” methods.

qiime feature-classifier classify-consensus-vsearch \

--i-query dada2/rep-seqs_yoga_dada2.qza \
--i-reference-reads taxonomy/99_otus.qza \
--i-reference-taxonomy taxonomy/99_otu_taxonomy.qza \
--p-perc-identity 0.97 \
--o-classification taxonomy/vsearch_tax_yoga_dada2.qza \
--verbose
 Targeted Gene Metagenomic Data Analysis   ◾    291

7.3.5.2 Using Machine Learning Classifiers

The machine learning taxonomy classifiers use trained model to assign taxa to the
­representative sequences rather than using the alignment approach. A classifier requires a
benchmark training dataset with known taxa for model training. With QIIME2, any of the
machine learning methods available in scikit-learn can be used to train a classifier for tax-
onomy assignment. However, there are also pre-fitted classifiers that can be used instead.
In the following, we will use a pre-fitted classifier for the taxonomy assignment, and later,
we will train a new model and use it as well.
For pre-fitted classifier, we can use “classify-sklearn” method of “q2-feature-classifier”
to assign taxa to the representative sequences that have been obtained from clustering or
denoising. First, we need to download a pre-fitted classifier. We can use a classifier pre-
trained on GreenGenes database with 99% OTUs using Naive Bayes machine learning
method. The pre-fitted classifiers are available at QIIME2 website at “https://docs.qiime2.
org/2022.2/data-resources/”. Create the subdirectory “classifiers” and download the classi-
fier artifact into it as follows:

mkdir classifiers
wget -O “classifiers/gg-nb-99-classifier.qza” \
“https://data.qiime2.org/2021.11/common/gg-13-8-99-nb-
classifier.qza”

Once the download has been completed, use that classifier artifact as an input for “clas-
sify-sklearn” method together with the representative sequence artifact generated in the
clustering or denoising step. In the following, we will assign taxa to the representative
sequences generated by DADA2:

qiime feature-classifier classify-sklearn \

--i-classifier classifiers/gg-nb-99-classifier.qza \
--i-reads dada2/rep-seqs_yoga_dada2.qza \
--o-classification taxonomy/nb_tax_yoga_dada2.qza

Instead of using a pre-fitted one, we can train a classifier using “feature-classifier” plu-
gin, which has two methods for model fitting: “fit-classifier-naive-bayes” for the training
of a naïve bayes classifier and “fit-classifier-sklearn” for the training of any scikit-learn
classifier.
Next, we will train a Naive Bayes classifier using GreenGenes reference sequences and
then we will use the fitted classifier to assign taxa to the representative sequences generated
by a previous clustering or denoising step.
For training any classifier, we need a training dataset with known labels. In the case
of taxonomy classification, we need representative sequences with known taxa. For our
example, we can use GreenGenes 13_8 97% OTU dataset. Remember that we downloaded
GreenGenes database before and stored it in the “gg_13_8_otus” subdirectory. We will use
the representative sequences “gg_13_8_otus/rep_set/97_otus.fasta” and their correspond-
ing taxonomic classifications “gg_13_8_otus/taxonomy/97_otu_taxonomy.txt”. Since the
292   ◾    Bioinformatics

representative sequences are in a FASTA file and taxonomy in text file, we can import both
of them as follows:

qiime tools import \

--type ‘FeatureData[Sequence]’ \
--input-path gg_13_8_otus/rep_set/97_otus.fasta \
--output-path inputs/97_otus.qza

qiime tools import \

--type ‘FeatureData[Taxonomy]’ \
--input-format HeaderlessTSVTaxonomyFormat \
--input-path gg_13_8_otus/taxonomy/97_otu_taxonomy.txt \
--output-path inputs/ref-gg-97-taxonomy.qza

After importing the training datasets (sequences and taxonomy) into the “inputs” subdi-
rectory, we can then use the “fit-classifier-naive-bayes” method of the “feature-classifier”
plugin to train the naïve bayes classifier and to save it in the “classifiers” subdirectory that
we created earlier.

qiime feature-classifier fit-classifier-naive-bayes \

--i-reference-reads inputs/97_otus.qza \
--i-reference-taxonomy inputs/ref-gg-97-taxonomy.qza \
--o-classifier classifiers/nb-gg-97-classifier.qza

After fitting, you can use the classifier artifact “nb-gg-97-classifier.qza” with the “classify-
sklearn” method to assign taxa to our unclassified representative sequences.

qiime feature-classifier classify-sklearn \

--i-classifier classifiers/nb-gg-97-classifier.qza \
--i-reads dada2/rep-seqs_yoga_dada2.qza \
--o-classification dada2/nb2_tax_yoga_dada2.qza

In the above, we used both alignment-based classifiers (BLAST, VSEARCH) and machine
learning classifiers (pre-fitted and fitted) to assign taxa to the unclassified representative
sequences. The output of any of these classification steps is an artifact for the classified
sequences. Whatever classifier is used for taxonomy assignment, the following is applied
to view the taxonomy results. A visualization file can be created from the resulted artifact
using “q2-metadata” plugin with “tabulate” method as follows:
Visualizing the BLAST-based taxonomy assignment:

qiime metadata tabulate \

--m-input-file dada2/blast_tax_yoga_dada2.qza \
--o-visualization dada2/blast_tax_yoga_dada2.qzv
qiime tools view dada2/blast_tax_yoga_dada2.qzv

Visualizing the VSEARCH taxonomy assignment:

 Targeted Gene Metagenomic Data Analysis   ◾    293

qiime metadata tabulate \

--m-input-file dada2/vsearch_tax_yoga_dada2.qza \
--o-visualization dada2/vsearch_tax_yoga_dada2.qzv
qiime tools view dada2/vsearch_tax_yoga_dada2.qzv

Visualizing the pre-fitted naïve bayes classifier taxonomy assignment:

qiime metadata tabulate \

--m-input-file dada2/nb_tax_yoga_dada2.qza \
--o-visualization dada2/nb_tax_yoga_dada2.qzv
qiime tools view dada2/nb_tax_yoga_dada2.qzv

Visualizing the trained naïve bayes classifier taxonomy assignment:

qiime metadata tabulate \

--m-input-file dada2/nb2_tax_yoga_dada2.qza \
--o-visualization dada2/nb2_tax_yoga_dada2.qzv
qiime tools view dada2/nb2_tax_yoga_dada2.qzv

You can compare between the taxonomy classification of the different classifiers. The fea-
ture table displayed on the Internet browser has three columns: Feature ID, Taxon, and
consensus (for alignment-based classifier) or confidence (for machine learning classifier).
The taxon column indicates the taxonomy assignment for each feature (k__ for kingdom,
p__ for phylum, c__ for class, o__ for order, f__ for family, g__ for genus, and s__ for spe-
cies). For example, in Figure 7.15, the naïve bayes classifier predicted the taxa of the first
feature up to the family level “f__Coriobacteriaceae;” with a confidence of 0.994, but it did
not assign a genus or a species to that feature. However, for the second feature, the classifier
predicted taxa up to the species level with confidence of 0.76. The confidence reflects the
probability that the taxonomy is correct.
Instead of confidence, alignment-based classifiers provide consensus that is based on
the agreement of the alignment hits. For example, a consensus of 1 means that all hits
aligned to the feature agreed on the taxa.
A taxonomy bar plot visualization can also be created with “taxa barplot” using the
taxonomy predicted by the classifiers and the filtered feature table artifact generated in the
clustering or denoising step as input.

FIGURE 7.15 Taxa assigned to features using naïve bayes classifier.

294   ◾    Bioinformatics

qiime taxa barplot \

--i-table dada2/table_feat_sample_freq_filtered_yoga_dada2.qza \
--i-taxonomy dada2/nb_tax_yoga_dada2.qza \
--m-metadata-file data/sample-metadata.tsv \
--o-visualization dada2/nx_tax_yoga_dada2_barplot.qzv
qiime tools view dada2/nx_tax_yoga_dada2_barplot.qzv

The taxonomy barplot (Figure 7.16) provides an interactive graphic interface to view the
bacterial taxa distribution in each sample as shown by the color keys. The bar width can
be changed by the slider bar on the top left. You can also use the dropdown menus on
the top (taxonomic levels, color palette, and Sort Sample by) to change the view and view
taxa ­distribution at specific taxonomic level or sort sample by a taxon or an experimental
group.
Earlier, we discussed how to filter the feature table to remove samples or features (with
low frequency). However, after viewing the taxonomy report, you may decide to focus on
a specific taxon or taxa across all samples. To achieve that you can use “taxa filter-table”
to create a new feature table for taxa that you wish to study. Filtering can be applied using
that method to retain only specific taxa using “--p-include” or to remove specific taxa using
“--p-exclude” or can be applied together. Multiple taxa are provided to the argument as
comma-separated list. By default, the terms provided for “--p-include” or “--p-exclude”
will match if they are contained in a taxonomic annotation. However, if you need the terms
match exactly the complete taxonomic annotation, then you can use “--p-mode exact”
parameter as well. Run “qiime taxa filter-table --help” to read more about the usage and
parameters. The “taxa” plugin output is a new feature table with the selected taxa. For
example, assume that you wish to retain only features that contain a phylum level classifi-
cation. So, you can use “--p-include p__” as follows:

FIGURE 7.16 Taxonomy bar plot (the samples are sorted by group).
 Targeted Gene Metagenomic Data Analysis   ◾    295

qiime taxa filter-table \

--i-table dada2/table_feat_sample_freq_filtered_yoga_dada2.qza \
--i-taxonomy taxonomy/nb_tax_yoga_dada2.qza \
--p-include p__\
--o-filtered-table dada2/table_phylum_yoga_dada2.qza
qiime taxa barplot \
--i-table dada2/table_phylum_yoga_dada2.qza \
--i-taxonomy taxonomy/nb_tax_yoga_dada2.qza \
--m-metadata-file data/sample-metadata.tsv \
--o-visualization taxonomy/nx_phylum_yoga_dada2_barplot.qzv
qiime tools view taxonomy/nx_phylum_yoga_dada2_barplot.qzv

Assume that you intend to explore the species under the “p__Bacteroidetes” phylum. Thus,
you can use “p__Bacteroidetes” to retain the features that contain this.

qiime taxa filter-table \

--i-table dada2/table_feat_sample_freq_filtered_yoga_dada2.qza \
--i-taxonomy taxonomy/nb_tax_yoga_dada2.qza \
--p-include “p__Bacteroidetes”\
--o-filtered-table taxonomy/table_Bacteroidetes_yoga_dada2.qza
qiime taxa barplot \
--i-table taxonomy/table_Bacteroidetes_yoga_dada2.qza \
--i-taxonomy taxonomy/nb_tax_yoga_dada2.qza \
--m-metadata-file data/sample-metadata.tsv \
--o-visualization taxonomy/nx_Bacteriodetes_yoga_dada2_barplot.
qzv
qiime tools view taxonomy/nx_Bacteriodetes_yoga_dada2_barplot.qzv

A metagenomic study may have a specific design such as groups, treatment, and facto-
rial design. The groups are to be specified in the sample metadata file as discussed above.
Grouping samples in an analysis by a group or multiple groups can be achieved by “feature-
table group”. The following script creates a feature table in which the samples are grouped
by the “group” column of the sample metadata file:

qiime feature-table group \

--i-table dada2/table_feat_sample_freq_filtered_yoga_dada2.qza \
--p-axis sample \
--m-metadata-file data/sample-metadata.tsv \
--m-metadata-column group \
--p-mode sum \
--o-grouped-table taxonomy/groupedby-group-yoga-table.qza

You can then create a barplot visualization for the new feature table; however, you need to
create a new metadata file in which the group levels (the metadata column which was used
for grouping) will be as sample ID. The metadata “taxonomy/group_metadata.tsv” will be
as in Figure 7.17.
296   ◾    Bioinformatics

FIGURE 7.17 Metadata file for sample grouping (group levels as sample ID).

FIGURE 7.18 Taxonomy bar plot for the distribution of taxa in each group.

The barplot can be then created and viewed as follows (Figure 7.18):

qiime taxa barplot \

--i-table taxonomy/groupedby-group-yoga-table.qza \
--i-taxonomy taxonomy/nb_tax_yoga_dada2.qza \
 Targeted Gene Metagenomic Data Analysis   ◾    297

--m-metadata-file taxonomy/group_metadata.tsv \
--o-visualization taxonomy/groupedby-group-yoga-barplot.qzv
qiime tools view taxonomy/groupedby-group-yoga-barplot.qzv

7.3.6 Construction of Phylogenetic Tree

The phylogenetic tree is used to visualize the evolutionary relationship between sequences.
In the amplicon-based sequence data analysis, a phylogenetic tree is also required for gen-
erating some diversity indexes such as Faith’s phylogenetic diversity index and UniFrac
index.
QIIME2 provides several methods for constructing phylogenetic trees based on the
representative sequences generated from clustering or denoising. The methods include
alignment-based approaches for the de novo tree construction and a fragment-insertion
method that aligns features against a reference tree. It is recommended that the de novo
phylogenetic tree is constructed from long sequences only.

7.3.6.1 De Novo Phylogenetic Tree

In QIIME2, the de novo phylogenetic tree is constructed by using the following steps:

7.3.6.1.1 Multiple Sequence Alignment

The representative sequences obtained from clustering or denoising step are aligned using
“q2-align” plugin with “mafft” method. MAFFT is a program for multiple sequence align-
ment. For the tree files, we will create the “trees” subdirectory in the project directory.

mkdir trees
qiime alignment mafft \
--i-sequences dada2/rep-seqs_yoga_dada2.qza \
--o-alignment trees/aligned-rep-seqs_yoga_dada2.qza

7.3.6.1.2 Masking Sites of Low Information

The positions of low phylogenetic information on the sequences are masked to be avoided.

qiime alignment mask \

--i-alignment trees/aligned-rep-seqs_yoga_dada2.qza \
--o-masked-alignment trees/masked-aligned-rep-seqs_yoga_dada2.
qza

7.3.6.1.3 Creating a Tree

An unrooted tree is created from the aligned sequences using Fasttree that infers approxi-
mately maximum-likelihood unrooted phylogenetic trees from the sequence alignments.

qiime phylogeny fasttree \

--i-alignment trees/masked-aligned-rep-seqs_yoga_dada2.qza \
--o-tree trees/unrooted-denovoTr-tree.qza
298   ◾    Bioinformatics

7.3.6.1.4 Midpoint Rooting

Then, we can define the root of the phylogenetic tree as its midpoint.

qiime phylogeny midpoint-root \

--i-tree trees/unrooted-denovoTr-tree.qza \
--o-rooted-tree trees/rooted-denovoTr-tree.qza

Now, we have created both unrooted and rooted phylogenetic trees using de novo approach.
The tree artifact can be exported into NEWICK tree file format which can be viewed on any
tree viewing program such as ETE toolkit. In the following, we will use “export” method of
“q2-tools” plugin to export phylogenetic tree data into NEWICK tree files:

qiime tools export \

--input-path trees/unrooted-denovoTr-tree.qza \
--output-path tree/unrooted
qiime tools export \
--input-path trees/rooted-denovoTr-tree.qza \
--output-path trees/rooted

7.3.6.2 Fragment-Insertion Phylogenetic Tree

A phylogenetic tree based on aligning the representative sequences to reference sequences
can be constructed with “q2-phylogeny” using “align-to-tree-mafft-fasttree” method, which
takes representative sequences as an input and outputs four artifacts: aligned sequences,
masked alignments, unrooted tree, and rooted tree. The following script creates unrooted
tree and rooted tree and export them to NEWICK files:

mkdir trees2
qiime phylogeny align-to-tree-mafft-fasttree \
--i-sequences dada2/rep-seqs_yoga_dada2.qza \
--o-alignment trees2/rep-seqs_yoga_dada2_alignedTr.qza \
--o-masked-alignment trees2/rep-seqs_yoga_dada2_maskedTr.qza \
--o-tree trees2/unrooted-tree-yoga_dada2.qza \
--o-rooted-tree trees2/rooted-tree-yoga_dada2.qza
qiime tools export \
--input-path trees2/unrooted-tree-yoga_dada2.qza \
--output-path trees2/unrooted
qiime tools export \
--input-path trees2/rooted-tree-yoga_dada2.qza \
--output-path trees2/rooted

7.3.7 Alpha and Beta Diversity Analysis

Above, we discussed the alpha and beta diversity and the metrics used for each type. In the
following, we will show you how to use QIIME2 to compute these metrics.
We will use the “q2-diversity” plugin to generate the phylogenetic and non-phylogenetic
diversity metrics. This plugin creates an artifact that contains both alpha and beta diversity
 Targeted Gene Metagenomic Data Analysis   ◾    299

metrics. First, we need to normalize the read counts across sample to adjust for any bias
arising from the different sequence depths and to make the comparison meaningful. The
normalization is performed by rarefying the count of feature table to a user-specified depth.
The lowest read count can be chosen as the user-defined depth. The lowest number of reads
is determined from a summary created from the feature table. The lowest count number
is then provided to the “--p-sampling-depth” parameter of the “diversity” plugin as a sam-
pling depth for all samples. Once the plugin command is executed, samples are drawn
without replacement so that each sample in the resulting table will have a total count equal
to that of sampling depth. Then, the alpha and beta metrics are computed. The following
script creates summary from the feature table to determine the lowest read number:

qiime feature-table summarize \

--i-table dada2/table_feat_sample_freq_filtered_yoga_dada2.qza \
--o-visualization dada2/table_feat_sample_freq_filtered_yoga_
dada2.qzv \
--m-sample-metadata-file data/sample-metadata.tsv
qiime tools view dada2/table_feat_sample_freq_filtered_yoga_dada2.qzv

When we study the summary, we can observe that the lowest read number for the samples
is 955 sequences. So, we can set the --p-sampling-depth parameter to 955. This step will
sub-sample the counts in each sample without replacement so that each sample in the
resulting table will have a total count of 955.
The “diversity” plugin requires a phylogenetic tree and feature table artifacts and the
sample metadata file as inputs and it outputs the alpha and beta diversity metrics saved into
the specified output directory.

qiime diversity core-metrics-phylogenetic \

--i-phylogeny trees2/rooted-tree-yoga_dada2.qza \
--i-table dada2/table_feat_sample_freq_filtered_yoga_dada2.qza \
--p-sampling-depth 955 \
--m-metadata-file data/sample-metadata.tsv \
--output-dir diversity-indices

The metrics would be saved to the output directory. We can use that metric to explore
the microbial composition of sample in the context of the grouping defined in the sample
metadata.
We will test for associations between categorical metadata columns and alpha diversity
data. We will do that here for the Faith Phylogenetic Diversity (a measure of community
richness) and Shannon diversity. The following commands will test for significant differ-
ences in the alpha diversity measures of samples:

qiime diversity alpha-group-significance \

--i-alpha-diversity diversity-indices/faith_pd_vector.qza \
--m-metadata-file data/sample-metadata.tsv \
--o-visualization diversity-indices/faith-pd-group-significance.qzv
300   ◾    Bioinformatics

qiime tools view diversity-indices/faith-pd-group-significance.qzv

qiime diversity alpha-group-significance \
--i-alpha-diversity diversity-indices/shannon_vector.qza \
--m-metadata-file data/sample-metadata.tsv \
--o-visualization diversity-indices/shannon-group-significance.qzv
qiime tools view diversity-indices/shannon-group-significance.qzv

These commands will run all-group and pairwise Kruskal–Wallis tests (non-parametric
analysis of variance). The visualization files show boxplots and test statistics for each meta-
data grouping.
We will analyze sample composition (beta diversity group distances) in the context
of categorical metadata using PERMANOVA. Note: The qiime diversity beta-group-­
significance command computes only one metadata grouping at a time, so to test the
differences between groups we have to indicate the appropriate column name from the
metadata file. In addition, if we call this command with --p-pairwise parameter, it will
perform pairwise tests that will allow us to determine which specific pairs of groups are
different from one another in terms of dispersion. We will apply a PERMANOVA to test
for significant differences of the weighted UniFrac metrics between the samples.

qiime diversity beta-group-significance \

--i-distance-matrix diversity-indices/weighted_unifrac_distance_
matrix.qza \
--m-metadata-file data/sample-metadata.tsv \
--m-metadata-column group \
--o-visualization \
diversity-indices/weighted-unifrac-life-stage-significance.qzv \
--p-pairwise
qiime tools view \
diversity-indices/weighted-unifrac-life-stage-significance.qzv

Finally, we will use the Emperor tool to explore the microbial community composition
using principal coordinate analysis (PCoA) plots in the context of sample metadata.

qiime emperor plot \

--i-pcoa diversity-indices/weighted_unifrac_pcoa_results.qza \
--m-metadata-file data/sample-metadata.tsv \
--o-visualization diversity-indices/weighted-unifrac-emperor-life-
stage.qzv
qiime tools view diversity-indices/weighted-unifrac-emperor-life-
stage.qzv

7.4 SUMMARY
The amplicon-based sequencing is targeting a specific marker gene that is able to distinguish
species. Hence, it is used to identify species in a sample that contains multiple microbes
such as environmental and clinical samples. The 16S rRNA gene is usually targeted in the
 Targeted Gene Metagenomic Data Analysis   ◾    301

amplicon-based metagenomic analysis because it can identify bacterial species. A region

or regions of the gene are amplified using PCR and the amplicon then is sequenced with
the high-throughput technologies. The reads are usually for the targeted gene but for sev-
eral species. The analysis is then focused on identifying the taxonomic groups and their
abundance in the sample. After quality control, features unique sequences representing
taxonomic groups are obtained either by clustering or denoising. There are three kinds of
clustering: de novo clustering, open-reference clustering, and closed-reference clustering.
Any of these clustering methods will generate OTUs or operational taxonomic units. On
the other hand, denoising attempts to remove base call errors and classification error and
it produces ASVs, which are unique features that represent species in the sample. There
are three common algorithms for denoising: DADA2, Deblur, and UNOISE2. The most
commonly used program for amplicon-based metagenomic data analysis is QIIME2,
which implements both clustering methods and denoising methods. To analyze data with
QIMME2, raw data must be imported into QIIME2 artifacts. Several analyses can be con-
ducted with QIIME2 including taxonomic group identification and abundance, phyloge-
netic analysis, and diversity analysis.

REFERENCES
1. Coughlan L, Cotter P, Hill C, Alvarez-Ordóñez A: Biotechnological applications of functional
metagenomics in the food and pharmaceutical industries. Front Microbiol 2015, 6.
2. Schwartsmann G, Brondani da Rocha A, Berlinck RG, Jimeno J: Marine organisms as a source
of new anticancer agents. Lancet Oncol 2001, 2(4):221–225.
3. Xiong ZQ, Wang JF, Hao YY, Wang Y: Recent advances in the discovery and development of
marine microbial natural products. Mar Drugs 2013, 11(3):700–717.
4. Sun Z, Li J, Dai Y, Wang W, Shi R, Wang Z, Ding P, Lu Q, Jiang H, Pei W et al: Indigo Naturalis
Alleviates Dextran Sulfate Sodium-Induced Colitis in Rats via Altering Gut Microbiota. Front
Microbiol 2020, 11: 731.
5. Blaxter M, Mann J, Chapman T, Thomas F, Whitton C, Floyd R, Abebe E: Defining opera-
tional taxonomic units using DNA barcode data. Philos Trans R Soc Lond B Biol Sci 2005,
360(1462):1935–1943.
6. Westcott SL, Schloss PD: De novo clustering methods outperform reference-based meth-
ods for assigning 16S rRNA gene sequences to operational taxonomic units. PeerJ 2015,
3:e1487.
7. Rideout JR, He Y, Navas-Molina JA, Walters WA, Ursell LK, Gibbons SM, Chase J, McDonald
D, Gonzalez A, Robbins-Pianka A et al: Subsampled open-reference clustering creates
­consistent, comprehensive OTU definitions and scales to billions of sequences. PeerJ 2014,
2:e545.
8. Callahan BJ, McMurdie PJ, Rosen MJ, Han AW, Johnson AJA, Holmes SP: DADA2:
High-resolution sample inference from Illumina amplicon data. Nat Methods 2016,
­
13(7):581–583.
9. Nearing JT, Douglas GM, Comeau AM, Langille MGI: Denoising the Denoisers: an
­independent evaluation of microbiome sequence error-correction approaches. PeerJ 2018,
6:e5364.
10. Edgar RC: UNOISE2: improved error-correction for Illumina 16S and ITS amplicon
­sequencing. bioRxiv 2016:081257.
11. Altschul SF, Gish W, Miller W, Myers EW, Lipman DJ: Basic local alignment search tool. J Mol
Biol 1990, 215(3):403–410.
302   ◾    Bioinformatics

12. Rognes T, Flouri T, Nichols B, Quince C, Mahé F: VSEARCH: a versatile open source tool for
metagenomics. PeerJ 2016, 4:e2584.
13. Cole JR, Wang Q, Fish JA, Chai B, McGarrell DM, Sun Y, Brown CT, Porras-Alfaro A, Kuske
CR, Tiedje JM: Ribosomal Database Project: data and tools for high throughput rRNA analy-
sis. Nucleic Acids Res 2014, 42(Database issue):D633–642.
14. Olsen GJ: [53] Phylogenetic analysis using ribosomal RNA. In: Methods in Enzymology. vol.
164: Academic Press; 1988: 793–812.
15. Bolyen E, Rideout JR, Dillon MR, Bokulich NA, Abnet CC, Al-Ghalith GA, Alexander H,
Alm EJ, Arumugam M, Asnicar F et al: Reproducible, interactive, scalable and extensible
microbiome data science using QIIME 2. Nat Biotechnol 2019, 37(8):852–857.
16. Thompson LR, Sanders JG, McDonald D, Amir A, Ladau J, Locey KJ, Prill RJ, Tripathi A,
Gibbons SM, Ackermann G et al: A communal catalogue reveals Earth’s multiscale microbial
diversity. Nature 2017, 551(7681):457–463.
 Chapter 8

Shotgun Metagenomic
Data Analysis

8.1 INTRODUCTION
In the previous chapter, we discussed the amplicon-based metagenomic data analysis which
is based on the profiling of a single targeted gene, usually 16S rRNA gene in environmental
or clinical samples. Many researchers debate that approach is not metagenomic in nature
because it focuses only on a single gene rather than the entire genomes of the microbes
in the samples. In this chapter, we will discuss the shotgun sequencing metagenomic
approach which involves the sequencing of the entire genomes of the microbes in the sam-
ples, and therefore, it provides more insights onto the microbial communities, their genetic
profiling, and their impact on hosts and association to the host phenotype. The shotgun
sequencing for the metagenomes is rather new but it also emerged as a consequence of the
progress in the high-throughput sequencing technologies, which was also followed by the
progress in the development of the computational resources and tools that are capable to
handle the massiveness and complexity of the metagenomic sequencing data. The shotgun
whole-genome metagenomic sequencing and data analysis are now used to quantify the
microbial communities and diversity, to assemble novel microbial genomes, to identify
new microbial taxa and genes, and to determine the metabolic pathways orchestrated by
the microbial community and more.
The metagenomic raw data produced by a high-throughput sequencer is originated
from either environmental or clinical samples that contain multiple microbial organisms,
including bacteria, fungi, and viruses. Data originated from samples recovered from a
host may be contaminated with the host genomic sequences. Multiple samples can also
be sequenced in a single run (multiplexing). In the multiplex sequencing, unique barcode
sequences identifying each sample are ligated to the DNA fragments in the DNA library
preparation step. Some library preparation kits allow multiplexing of hundreds of samples.
Illumina has multiple kits for library preparation, including Illumina DNA Prep, (M) tag-
mentation, which uses bead-linked transposomes in the tagmentation process to randomly

DOI: 10.1201/9781003355205-8 303

304   ◾    Bioinformatics

insert transposomes into the metagenomic DNA. The tagmentation is followed by PCR
using index primers which enables amplification and subsequent indexing of the sample
libraries (barcoding) to allow multiplexing. The library preparation is followed by sequenc-
ing and production of the raw data in FASTQ files. The steps of read quality assessment and
processing are, to some extent, similar to the steps discussed in Chapter 1. The purpose of
the quality control is to reduce sequence biases or artifacts by removing sequencing adap-
tors, trimming low-quality ends of the reads, and removing duplicate reads. If the DNA is
extracted from a clinical sample, an additional quality control step is required which is to
remove the contaminating host DNA or non-target DNA sequences. If we need to perform
between-sample differential diversity analysis, we may also need to draw a random sub-
sample from the original sample to normalize read counts.
After the step of quality control, there are two strategies that can be followed for the
metagenomic raw data. The first one is to assemble the metagenomes using de novo genome
assembly method and the second one is an assembly-free approach similar to amplicon-
based method. Each of these strategies may address different kinds of questions. The types
and algorithms of the de novo assembly were discussed in Chapter 3. However, in the
shotgun metagenomics, a new step is introduced. This step is called metagenomic bin-
ning, which aims to separate the assembled sequences by species so that the assembled
contigs in a metagenomic sample will be assigned to different bins in FASTA files. A bin
will correspond to only one genome. A genome built with the process of binning is called
Metagenome-Assembled Genome (MAG). Binning algorithms adopt several ways to per-
form binning. Some algorithms use taxonomic assignment and others use properties of
contigs like GC-content of the contigs, nucleotide composition, or the abundance. Binning
algorithms use two approaches for assigning contigs to species: supervised machine learn-
ing and unsupervised machine learning. Both approaches use similarity scores to assign
a contig to a bin. Since many of the microbial species have not been sequenced and hence
some of the reads may not map to reference genomes, it is good practice to not rely on
mapping to reference genomes. Binning-based nucleotide composition of a contig has been
found useful in separating genomes into possible species. The nucleotide composition of
a contig is the frequency of k-mers in the contig, where k can be any reasonable integer
(e.g., 3, 4, 5, …). It has been found that different genomes of microbial species have dif-
ferent frequencies that may discriminate the genomes into potential taxonomic groups.
A machine learning algorithm like naïve Bayes and other machine learning algorithms
are used for taxonomic group assignment. However, features, more powerful than the
sequence composition, are required to deal with the complexity in the sequences of contigs.
The unsupervised machine learning tools cluster contigs into bins without requiring prior
information. There are several binning programs using different algorithms. MetaBAT 2
[1] uses an adaptive binning algorithm that does not require manual parameter tuning as
the case with its previous version. Its algorithm consists of multiple aspects, including nor-
malized tetranucleotide frequency (TNF) scores, clustering, and steps to recruit smaller
contigs. Moreover, the computational efficiency has been increased compared to the previ-
ous version. MaxBin [2] uses nucleotide composition and contig abundance information
to group metagenomic contigs into different bins; each bin represents one species. MaxBin
algorithm uses tetranucleotide frequencies and scaffold coverage levels to estimate the
 Shotgun Metagenomic Data Analysis   ◾    305

Reads
Bin 1
Contigs
Bin 2
Sequencing Assembly Binning
Bin 3

Bin 4

FIGURE 8.1 Reads processing from sequencing to bin formation.

probability that a scaffold belongs to a bin using an expectation-maximization (EM) algo-

rithm. The program also provides statistics, including genome completeness, GC-content,
and genome size. Figure 8.1 shows the steps from sequencing to binning.

8.2 SHOTGUN METAGENOMIC ANALYSIS WORKFLOW

The first two steps of the metagenomic data analysis workflow are raw data acquisition and
quality control. After the quality control, the raw data can pass through two different steps:
(i) de novo assembly and subsequent analysis and (ii) assembly-free analysis. In the follow-
ing, we discuss these steps with a worked example.

8.2.1 Data Acquisition

The raw shotgun metagenomic data is sequences of the metagenomic DNA extracted from
either environmental or clinical samples which usually contain several species of microbes.
Depending on the sequencing technology, data can be short reads produced by Illumina
and other short-read sequencing technologies or long reads produced either by Pacific
Bioscience (PacBio) or by Oxford Nanopore Technology (ONT). The read layout can also be
single end or paired end. The raw data is usually provided in FASTQ files. Many research-
ers uploaded their raw data to a database like NCBI SRA and make it available for public.
We will download FASTQ files from the NCBI SRA data for the purpose of demonstrat-
ing how analysis is conducted. The run numbers are “ERR1823587”, “ERR1823601”, and
“ERR1823608” which contain shotgun metagenomic data of human stool samples from a
healthy, a moderate, and a severe sickle cell disease patient, respectively. We will create the
directory “shotgun” as the project working directory; then, we will use the SRA-toolkits
“fasterq-dump” utility to download the paired-end files in a directory called “fastqdir”.

mkdir shotgun; cd shotgun

fasterq-dump --threads 4 --verbose --outdir fastqdir ERR1823587
fasterq-dump --threads 4 --verbose --outdir fastqdir ERR1823601
fasterq-dump --threads 4 --verbose --outdir fastqdir ERR1823608

Six FASTQ files will be saved in the “fastqdir”. Use “ls fastqdir/” to display the content of
that directory to make sure the files are there.

8.2.2 Quality Assessment and Processing

If you obtained these files directly from the sequencer, it is likely that they may need quality
control, which includes both quality assessments using one of the quality assessment pro-
grams like FastQC and processing to filter out the low-quality reads, to trim low-quality
306   ◾    Bioinformatics

ends of the reads, and to remove adaptors and duplicates. Refer to Chapter 1 for detailed
information about this step. For multiplexing data, you need to perform demultiplexing
before you do the quality control. The multiplexing and demultiplexing are discussed in
Chapter 7. The FASTQ files, which we have downloaded, had already been processed and
they contain reads of good quality. You can check their quality with FastQC as follows:

fqs=$(ls fastqdir/*.fastq)
fastqc $fqs
htmls=$(ls fastqdir/*.html)
firefox $htmls

The above commands will display the quality control report of the six FASTQ files on the
Firefox browser. Check on the six tabs to study the reports.

8.2.3 Removing Host DNA Reads

Metagenomic data recovered from clinical samples is usually mixed with the host genomic
DNA sequences. These sequences, which represent untargeted fraction of data, must be
filtered out before the subsequent step of the analysis. Any other untargeted sequences can
also be removed following the step of removing host sequences. The process of removing
the host sequences begins by aligning raw data to the reference genome of the host. The
host sequences will map to the reference genome, whereas the metagenomic reads will not
map. Thus, after the mapping process, we can extract the unmapped sequences and store
them in separate FASTQ files. For paired-end reads, we will have two FASTQ files repre-
senting the raw metagenomic data without the host sequences.
Since the host of our data is human, we will align reads to the human reference genome.
We have already discussed read mapping in Chapter 2 and other chapters as well. This
time we will use Bowtie2 aligner. We can walk you through the steps without repeating
the discussion. The following are the steps to remove the human host sequences from the
genomic data.

8.2.3.1 Download Human Reference Genome

You did this step before in Chapter 6. So, if you have the human reference genome and
Bowtie2 index saved in your drive, you can use them instead since building the Bowtie2
index may take some time. If you do not have those files stored on your computer, run the
following command in your project working directory to download the FASTA sequence of
the human reference genome, decompress it, and index it with both Samtools and Bowtie2:

mkdir ref; cd ref

wget https://hgdownload.soe.ucsc.edu/goldenPath/hg19/bigZips/hg19.
fa.gz
gunzip -d hg19.fa.gz
samtools faidx hg19.fa
bowtie2-build hg19.fa hg19
cd ..
 Shotgun Metagenomic Data Analysis   ◾    307

8.2.3.2 Mapping Reads to the Reference Genome

The following Bowtie2 commands will map the files of the samples to the human genome.
The SAM files, which contain both mapped and unmapped reads, are saved in “sam”
directory.

mkdir sam
bowtie2 -p 4 \
-x ref/hg19 \
-1 fastqdir/ERR1823587_1.fastq \
-2 fastqdir/ERR1823587_2.fastq \
-S sam/ERR1823587.sam
bowtie2 -p 4 \
-x ref/hg19 \
-1 fastqdir/ERR1823601_1.fastq \
-2 fastqdir/ERR1823601_2.fastq \
-S sam/ERR1823601.sam
bowtie2 -p 4 \
-x ref/hg19 \
-1 fastqdir/ERR1823608_1.fastq \
-2 fastqdir/ERR1823608_2.fastq \
-S sam/ERR1823608.sam

8.2.3.3 Converting SAM to BAM Format

samtools view \
-bS sam/ERR1823587.sam \
> sam/ERR1823587.bam
samtools view \
-bS sam/ERR1823601.sam \
> sam/ERR1823601.bam
samtools view \
-bS sam/ERR1823608.sam \
> sam/ERR1823608.bam

Remember that the BAM files contain mapped and unmapped reads.

8.2.3.4 Separating Metagenomic Reads in BAM Files

The next step is to remove the unmapped reads in separate files. We will use the “samtools
view” and FLAG to separate the unmapped reads.

samtools view \
-b -f 12 \
-F 256 \
sam/ERR1823587.bam \
> sam/ERR1823587_unmapped.bam
samtools view \
-b -f 12 \
308   ◾    Bioinformatics

-F 256 sam/ERR1823601.bam \
> sam/ERR1823601_unmapped.bam
samtools view \
-b -f 12 \
-F 256 sam/ERR1823608.bam \
> sam/ERR1823608_unmapped.bam

The “-f 12” option is used to extract only the unmapped forward and reverse reads and “-F
256” option is used to exclude secondary alignments. Refer to Chapter 2 for FLAG field of
the SAM/BAM file.
The above Samtools commands separate unmapped reads, which represent the
pure metagenomic data, in the separate BAM files “ERR1823587_unmapped.bam”,
“ERR1823601_unmapped.bam”, and “ERR1823608_unmapped.bam”.

8.2.3.5 Creating Paired-End FASTQ Files from BAM Files

Now, we can extract the FASTQ files from the above BAM files; we will extract two FASTQ
files from each BAM file. However, before doing that, we need to sort the BAM files by read
name using the “samtools sort” command with “-n” option, which sort the paired reads to
be next to each other.

samtools sort \
-n -m 5G \
-@ 2 sam/ERR1823587_unmapped.bam \
-o sam/ERR1823587_unmapped_sorted.bam
samtools sort \
-n -m 5G \
-@ 2 sam/ERR1823601_unmapped.bam \
-o sam/ERR1823601_unmapped_sorted.bam
samtools sort \
-n -m 5G \
-@ 2 sam/ERR1823608_unmapped.bam \
-o sam/ERR1823608_unmapped_sorted.bam

Then, we create FASTQ files from the BAM files and store them in a new directory “fastq_
pure” so that we can use them in the next steps of the downstream analysis.

Mkdir fastq_pure
samtools fastq -@ 4 sam/ERR1823587_unmapped_sorted.bam \
-1 fastq_pure/ERR1823587_pure_R1.fastq.gz \
-2 fastq_pure/ERR1823587_pure_R2.fastq.gz \
-0 /dev/null -s /dev/null -n
samtools fastq -@ 4 sam/ERR1823601_unmapped_sorted.bam \
-1 fastq_pure/ERR1823601_pure_R1.fastq.gz \
-2 fastq_pure/ERR1823601_pure_R2.fastq.gz \
-0 /dev/null -s /dev/null -n
samtools fastq -@ 4 sam/ERR1823608_unmapped_sorted.bam \
 Shotgun Metagenomic Data Analysis   ◾    309

-1 fastq_pure/ERR1823608_pure_R1.fastq.gz \
-2 fastq_pure/ERR1823608_pure_R2.fastq.gz \
-0 /dev/null -s /dev/null -n

The saved FASTQ files contain the reads of the metagenomic data after removing the con-
taminating human sequences.
If you run FastQC on those files, you will find that the reads length varies between 35
and 151 bp. We can remove any paired reads of length less than 50 bp. Removing such
reads from paired-end FASTQ files requires a program or a script that removes reads from
both pair of files without leaving singletons that may be rejected by some programs used in
the downstream analysis. Thus, we need to filter reads of certain length by using the right
program. There may be some programs that can do this but we use a bash script for this
purpose. First, change into “fastq_pure” and run the following script to decompress the
FASTQ files:

cd fastq_pure
for i in $(ls *.gz);
do
gzip -d ${i}
done

Then, open a text editor of your choice such as “vim or nano” and save the following bash
script in a file “remove_PE.sh”:

vim remove_PE.sh

Then, copy the bash script into the file, save it, and exit:

#!/bin/sh
#Use: remove_PE.sh R1.fastq R2.fastq 80
#1. Start with inputs
fq_r1=$1
fq_r2=$2
minLength=$3
#2. Find all entries with read length less than minimum length and
print line numbers, for both R1 and R2
awk -v min=$minLength \
‘{if(NR%4==2) \
if(length($0)<min) \
print NR”\n”NR-1”\n”NR+1”\n”NR+2}’ \
$fq_r1 > temp.lines1
awk -v min=$minLength \
‘{if(NR%4==2) \
if(length($0)<min) \
print NR”\n”NR-1”\n”NR+1”\n”NR+2}’ \
$fq_r2@ temp.lines1
310   ◾    Bioinformatics

#3. Combine files into one, sort them numerically, and collapse
redundant entries
sort -n temp.lines1 | uniq > temp.lines
rm temp.lines1
outfq1=$(echo $fq_r1| cut -d’.’ -f 1)
outfq2=$(echo $fq_r2| cut -d’.’ -f 1)
#4. Remove the line numbers recorded in “lines” from both fastqs
awk ‘NR==FNR{l[$0];next;} !(FNR in l)’ \
temp.lines $fq_r1 \
> $outfq1-$minLength.fastq
awk ‘NR==FNR{l[$0];next;} !(FNR in l)’ \
temp.lines $fq_r2 \
> $outfq2-$minLength.fastq
gzip $outfq1-$minLength.fastq
gzip $outfq2-$minLength.fastq
rm temp.lines

Once you have saved the file, you may need to make the file executable by using the Linux
command “chmod”:

chmod +x remove_PE.sh

Then, run the following commands:

./remove_PE.sh ERR1823587_pure_R1.fastq ERR1823587_pure_R2.fastq 50

./remove_PE.sh ERR1823601_pure_R1.fastq ERR1823601_pure_R2.fastq 50
./remove_PE.sh ERR1823608_pure_R1.fastq ERR1823608_pure_R2.fastq 50

Up to this step, we would have removed the host sequences from metagenomic data which are
stored in “ERR1823587_pure_R1-50.fastq.gz” and “ERR1823587_pure_R2-50.fastq.gz” for
the sample of the healthy person, “ERR1823601_pure_R1-50.fastq.gz” and “ERR1823601_
pure_R2-50.fastq.gz” for the moderate sickle cell patient, and “ERR1823608_pure_R1-50.
fastq.gz” and “ERR1823608_pure_R2-50.fastq.gz” for severe sickle cell patient. To save
some storage space, you can delete the other FASTQ files using “rm *.fastq” and also delete
all files in “fastqdir”.
The metagenomic FASTQ files are stored in “fastq_pure” as shown in Figure 8.2. Above,
we have deleted the original FASTQ files from “fastqdir” directory and also the intermedi-
ate FASTQ files from “fastq_pure” directory. You can also delete the SAM and BAM files
from “sam” directory and the reference sequences and indexes from “ref” directory if you
want to save storage space. However, you are advised to keep reference genome files in “ref”
as you may need to repeat all the steps and indexing usually takes a long time.

8.2.4 Assembly-Free Taxonomic Profiling

We can use the FASTQ files to perform taxonomic profiling without metagenome assem-
bly. This approach employs NGS short or long reads present in the metagenomic samples to
assign taxonomic groups by identifying unique genomic regions in the reads. Long reads
 Shotgun Metagenomic Data Analysis   ◾    311

allow more accurate taxonomic group assignment. There are several programs for assem-
bly-free classification and profiling of microbial communities in metagenomic samples.
Kaiju [3] uses taxonomy and NCBI refseq databases to find maximum matches to the reads
on the protein-level using the Burrows–Wheeler transform (BWT). CLARK [4] (CLAssifier
based on Reduced K-mers) creates a large index of k-mers of all target sequences and then
it removes the common ones among targets so that each target is described by unique
k-mers, which are used for taxonomic classification. Kraken [5] creates k-mers from the
reads and then it builds taxonomy trees that help discriminate closely related microbes
using classification tree and path. Those programs are just examples and there are others
with different algorithms. Centrifuge [6] is a rapid classifier that requires a little memory
and a relatively smaller index (only 5.8 GB for bacterial, viral, and human genomes) on
desktop computers compared to others. Centrifuge uses an indexing system that is based
on BWT and the Ferragina–Manzini (FM) index.
Most taxonomy classifiers of the metagenomic data use genomic database of known spe-
cies to construct an index and then use that index to assign taxa to the metagenomic reads.
The majority of the classifiers require a large storage space for database files and a large
memory for indexing and classification process. Kaiju and Kraken require a lot of memory
(around 128GB–512GB). Therefore, we recommend using these classifiers only if you have
enough computational resources. To use any of these classifiers, you need to download and
build an index and then to perform the classification.
Kaiju installation instructions are available at “https://github.com/bioinformatics-cen-
tre/kaiju”. You can install it by running the following command:

git clone https://github.com/bioinformatics-centre/kaiju.git

cd kaiju/src
make

Then, you need to add its path by adding the following to the “.bashrc” file. You need to
replace YOUR_PATH with the program path.

export PATH=”YOUR_PATH/kaiju/bin”:$PATH

You must restart the terminal or use “source ~/.bashrc” to make the change active. Run
“kaiju” command to check if it has been installed.
Before using kaiju, you need to download the refseq database from the NCBI or you
can download it from the kaiju website at “https://kaiju.binf.ku.dk/server”. To download it
from the NCBI database, use the following:

mkdir kaijudb
cd kaijudb
kaiju-makedb -s refseq

The download will take a long time and a large storage space. When the database has
been downloaded, make sure that “nodes.dmp”, “kaiju_db_refseq.fmi”, and “names.
dmp” files are present in the “kaijudb” directory. You may need to decompress
312   ◾    Bioinformatics

“kaiju_db_refseq_xxxx-xx-xx.tgz”. To classify the short reads in our FASTQ files, you need
to run the following:

mkdir kaiju_output
kaiju -t kaijudb/nodes.dmp \
-f kaijudb/kaiju_db_refseq.fmi \
-i fastq_pure/ERR1823587_pure_R1-80.fastq.gz \
-j fastq_pure/ERR1823587_pure_R2-80.fastq.gz \
-o kaiju_output/ERR1823587.out \
-a greedy \
-z 4 -v
kaiju -t kaijudb/nodes.dmp \
-f kaijudb/kaiju_db_refseq.fmi \
-i fastq_pure/ERR1823601_pure_R1-80.fastq.gz \
-j fastq_pure/ERR1823601_pure_R2-80.fastq.gz \
-o kaiju_output/ERR1823601.out \
-a greedy \
-z 4 -v
kaiju -t kaijudb/nodes.dmp \
-f kaijudb/kaiju_db_refseq.fmi \
-i fastq_pure/ERR1823608_pure_R1-80.fastq.gz \
-j fastq_pure/ERR1823608_pure_R2-80.fastq.gz \
-o kaiju_output/ERR1823608.out \
-a greedy \
-z 4 -v

To learn more about these options, run “kaiju”. The indexing and classification require
around 128 GB RAM. We do not recommend using kaiju unless you have enough memory
and storage space.
After running the program successfully, you will need to convert the kaiju output file
into a summary table using “kaiju2table” command as follows:

kaiju2table -t kaijudb/nodes.dmp \
-n kaijudb/names.dmp \
-r taxonomic_level \
-o kaiju_output/ERR1823587_table.tsv \
kaiju_output/ERR1823587.out \
-l taxonomic,levels,separated,by,commas
kaiju2table -t kaijudb/nodes.dmp \
-n kaijudb/names.dmp \
-r taxonomic_level \
-o kaiju_output/ERR1823601_table.tsv \
kaiju_output/ERR1823601.out \
-l taxonomic,levels,separated,by,commas
kaiju2table -t kaijudb/nodes.dmp \
-n kaijudb/names.dmp \
 Shotgun Metagenomic Data Analysis   ◾    313

-r taxonomic_level \
-o kaiju_output/ERR1823608_table.tsv \
kaiju_output/ERR1823608.out \
-l taxonomic,levels,separated,by,commas

Run “kaiju2table” to learn about the usage and options of this command.
Most taxonomy classifiers of the metagenomic data follow the same steps: the database
downloading and classification. For almost all of them, these steps require large storage
space and memory that may not be available on the regular desktop computers. However,
if we do not have enough computational resources, we can use Centrifuge which requires
relatively small storage space and memory that fits personal computers.
Centrifuge classifier is available at “https://github.com/infphilo/centrifuge”. For the
updated installation instructions, visit that site. Up to this day, you can install it on Linux
using the following commands:

git clone https://github.com/infphilo/centrifuge

cd centrifuge
make
sudo make install prefix=/usr/local

If it has been installed successfully, no need to do anything else but to use it from any
directory. Run “centrifuge -h” to display the usage and options.
As usual, to use Centrifuge classifier, we will begin by building an index. There are
several ready-to-use indexes available at http://www.ccb.jhu.edu/software/centrifuge.
However, Centrifuge also needs sequence and taxonomy files and sequence ID. That can be
simplified by using “make” command that can build several standard and custom indices.
To do that, find the Centrifuge directory and change into “indices” directory and then run
the “make” command as follows:

cd indices
make p+h+v # bacterial, human, and viral genomes [~12G]
make p_compressed # bacterial genomes compressed at the species
level [~4.2G]
make p_compressed+h+v # combination of the two above [~8G]

This command will download the reference taxonomy files and reference genome at assem-
bly levels. The download may take a while depending on the speed of the Internet connec-
tion. It is also easier to download a database from Centrifuge homepage, which is available
at “https://ccb.jhu.edu/software/centrifuge/manual.shtml”. Centrifuge is used to assign
taxa to the short reads in the FASTQ files. For the “-x” option, make sure that you provide
the database name with the path if it is not in the current path.

mkdir centrifuge_out
centrifuge -x p+h+v \
314   ◾    Bioinformatics

-1 fastq_pure/ERR1823587_pure_R1-50.fastq.gz \
-2 fastq_pure/ERR1823587_pure_R2-50.fastq.gz \
--report-file centrifuge_out/ERR1823587-report.txt \
-S centrifuge_out/ERR1823587-results.txt
centrifuge -x p+h+v \
-1 fastq_pure/ERR1823601_pure_R1-50.fastq.gz \
-2 fastq_pure/ERR1823601_pure_R2-50.fastq.gz \
--report-file centrifuge_out/ERR1823601-report.txt \
-S centrifuge_out/ERR1823601-results.txt
centrifuge -x p+h+v \
-1 fastq_pure/ERR1823608_pure_R1-50.fastq.gz \
-2 fastq_pure/ERR1823608_pure_R2-50.fastq.gz \
--report-file centrifuge_out/ERR1823608-report.txt \
-S centrifuge_out/ERR1823608-results.txt

The results are saved in “*-results.txt” files. Each read classified by Centrifuge results in
a single line of output. The output lines consist of eight tab-delimited fields: (1) the read
ID (from FASTQ file); (2) sequence ID (from the database sequence); (3) taxonomic ID of
the database sequence; (4) classification score (weighted sum of hits); (5) score for the next
best classification; (6) two numbers: (i) a number of base pairs of the read that match the
database sequence and (ii) the length of a read or the combined length of mate pairs; (7)
two numbers: (i) a number of base pairs of the read that match the database sequence and
(ii) the length of a read or the combined length of mate pairs; and (8) the number of clas-
sifications for this read.
The “*-report.txt” files contain summaries of the identified taxa and their abun-
dances. Each line in the file consists of seven tab-delimited fields: The name of a genome,

FIGURE 8.2 Partial centrifuge report for the healthy sample.

 Shotgun Metagenomic Data Analysis   ◾    315

taxonomic ID, taxonomic rank (kingdom, genus, family, etc.), genome size in bp, number
of reads classified to this genomic sequence including multi-classified reads, number of
reads uniquely classified to this genomic sequence, and abundance proportion as shown in
Figure 8.2. The Centrifuge report shows that the metagenomic reads have been assigned to
taxonomic group in different ranks. This report can be further analyzed to filter the most
significant taxa based on their summary statistics.

8.2.4 Assembly of Metagenomes

In the free-assembly microbial profiling, we could assign a taxonomic group to the metage-
nomic sequences and we could also obtain some useful statistics. In contrast, the assembly-
based profiling requires metagenome assembly, which is faced by challenges not present in
the assembly of a single genome discussed in Chapter 2. The metagenomic data includes
reads for several microbes with different sequence coverage rather than a uniform cover-
age, which is assumed by typical assemblers to assemble a single genome. Assuming a
uniform sequence coverage will allow to distinguish true sequences from errors, identify
repeat sequences, and identify allelic variation. That assumption is invalid in metagenome
assembly because the coverage of the genome of each species in the sample depends on the
abundance of that species in the sample. Since metagenome assembly is performed with
de novo approach that uses de Bruijn graph, low sequence coverage may lead to incontigu-
ous path. Assemblers can overcome that challenge by using a short k-mer size but that will
also increase the frequencies of identical k-mers in the graph, compromising the assembly
quality.
One more challenging problem faced by metagenome assembly is the presence of sub-
strains of the same bacterial species and that makes the graph more complex. The metage-
nome assemblers attempt to overcome these challenges using different strategies. As an
example, we will use metaSPAdes [7], which uses de Bruijn graph to form an assembly
graph and it also works across a wide range of coverage depths and attempts to main-
tain the trade-off between the accuracy and continuity of the metagenome assemblies.
metaSPAdes is one of the SPAdes programs that we discussed in Chapter 3. Refer to that
chapter for SPAdes installation. If you followed the installation instructions of SPAdes in
Chapter 3, you would have added its path to the “.bashrc” file. To check if the program is
installed and it is on the path, run the following:

metaspades.py

This will display the usage and options of metaSPAdes program. Otherwise, you may need
to install the program following the installation instructions.
The following metaSPAdes command will perform de novo metagenome assembly using
the metagenomic FASTQ files as input:

mkdir metag_healthy
metaspades.py \
-o metag_healthy \
-1 fastq_pure/ERR1823587_pure_R1-50.fastq.gz \
316   ◾    Bioinformatics

-2 fastq_pure/ERR1823587_pure_R2-50.fastq.gz \
--only-assembler \
--threads 4 \
--memory 16 \
--phred-offset 33 \
-k 51
mkdir metag_moderate
metaspades.py \
-o metag_moderate \
-1 fastq_pure/ERR1823601_pure_R1-50.fastq.gz \
-2 fastq_pure/ERR1823601_pure_R2-50.fastq.gz \
--only-assembler \
--threads 4 \
--memory 16 \
--phred-offset 33 \
-k 51
mkdir metag_severe
metaspades.py \
-o metag_severe \
-1 fastq_pure/ERR1823608_pure_R1-50.fastq.gz \
-2 fastq_pure/ERR1823608_pure_R2-50.fastq.gz \
--only-assembler \
--threads 4 \
--memory 16 \
--phred-offset 33 \
-k 51

Run “metaspades.py --help” to read about the usage and options of this program.
Several files are produced in the output directories: “metag_healthy”, “metag_moder-
ate”, and “metag_severe”. The files that contain the assembly sequences are the “contigs.
fasta” and the “scaffolds.fasta”. Contigs are made from read overlaps. The contigs are then
ordered, oriented, and connected with gaps filled with Ns to form the scaffolds. The K51
directory contains the individual result files for an assembly with 51-mers. However, when
multiple K directories are found, the best assembled sequences are the ones that are stored
outside these K directories. The directory “misc” contains broken scaffolds.
The file with the “.gfa” extension is in Graphic Fragment Assembly (GFA) file format in
which the sequences are represented by lines starting with “S” and the overlaps between
sequences are represented by lines starting with “L” as shown in Figure 8.3. The plus (+)
and minus (−) signs indicate whether the overlapping sequence is the original or its reverse
complement. The value in the form “XM” in a link indicates overlap length.
Thus, the file “assembly_graph_with_scaffolds.gfa” generated by metaSPAdes is the
GFA file that represents the final assembly of metagenomes in the sample. SPADes built
this assembly graph based on k-mers formed from the reads (vertices) and their overlaps
(edges). Then, the assembler resolves paths across the assembly vertices and outputs non-
branching paths as contigs.
 Shotgun Metagenomic Data Analysis   ◾    317

FIGURE 8.3 Graphic Fragment Assembly (GFA) file format.

This GFA file can be visualized by graph visualization programs like Bandage [8]. The
Bandage program is available at “http://rrwick.github.io/Bandage/” and can be down-
loaded and installed on Linux, Windows, and Mac OX. Visualizing a graph file will give
you an idea about the assembly quality. You can zoom in and out and do many operations
on the graph file. Moreover, there are good graph visualization packages in Python and R
such as igraph [9], which is available in both programming platforms. To read more about
Bandage and igraph, refer to their user manuals which are available on their web pages.

8.2.5 Assembly Evaluation

MetaSPAdes produced several assemblies for different species in a single scaffolds file.
Those assemblies can be evaluated by using an evaluation program like “metaquast.py”
[10], which is a Python-based program that can generate report showing the quality of the
assemblies. For the installation instructions of this program, visit “http://cab.cc.spbu.ru/
quast/manual.html”. To generate report for the scaffolds for each of the three samples, you
can run the following:

metaquast.py -t 4 \
-m 500 \
metag_healthy/scaffolds.fasta \
metag_moderate/scaffolds.fasta \
metag_severe/scaffolds.fasta \
-o output

This will generate an HTML report in the “output” directory. The other directories and
files are linked to this report when it is displayed on the Internet browser. You can display
“report.html” file by using “firefox report.html”.
318   ◾    Bioinformatics

FIGURE 8.4 An assemblies’ evaluation report generated with metaquast.py.

The assemblies’ evaluation report contains important statistics that reflect the quality of
the assemblies in the file. Figure 8.4 shows partial evaluation report of the three samples.
The colored heatmap indicates the quality from the worst (red) to the best (blue). On the
top, there are links that can take to each sample graph as shown in Figure 8.5. The graphs
show the key identified bacterial taxa and their abundance. Refer to the program users’
manual, which is available at “http://cab.cc.spbu.ru/quast/manual.html”, to read more
about the program use and the different report sections, refer to Chapter 3 to read more
about the de novo assembly evaluation metrics.

8.2.6 Mapping Reads to the Assemblies

We have already created the metagenomic FASTQ files, which are produced from the reads
unmapped to the host reference genome. Then, we used a de novo assembler to produce an
assembly (scaffolds.fasta) for each sample that may contain genomic sequences of several
microbes. In this step, we will use an aligner to map the reads in the FASTQ files to the
respective assembly. For this purpose, we can use Bowtie2 aligner. First, we need to build
an index for the “scaffolds.fasta” for each sample and then we will use it to align the reads
in FASTQ files. Now, let us create a directory named “assemblies” in the main project
directory and copy scaffolds FASTA files from the three sample directories into this new
directory with new file names as follows:
 Shotgun Metagenomic Data Analysis   ◾    319

FIGURE 8.5 The metagenomic composition of the healthy individual.

mkdir assemblies
cp metag_healthy/scaffolds.fasta assemblies/healthy_scaffolds.
fasta
cp metag_moderate/scaffolds.fasta assemblies/moderate_scaffolds.
fasta
cp metag_severe/scaffolds.fasta assemblies/severe_scaffolds.fasta

Then, we can use Samtools and Bowtie2 to build an index for each “scaffolds.fasta” file of
each sample.

cd assemblies
cd assemblies
for i in $(ls *.fasta);
320   ◾    Bioinformatics

do
samtools faidx ${i}
done
bowtie2-build healthy_scaffolds.fasta healthy
bowtie2-build moderate_scaffolds.fasta moderate
bowtie2-build severe_scaffolds.fasta severe

Once the index has been built, we can use Bowtie2 to align the FASTQ reads to their
respective “scaffolds.fasta”.

mkdir sam_assemblies
bowtie2 --sensitive-local \
-p 4 \
-x assemblies/healthy \
-1 fastq_pure/ERR1823587_pure_R1-50.fastq.gz \
-2 fastq_pure/ERR1823587_pure_R2-50.fastq.gz \
-S sam_assemblies/ERR1823587_healthy.sam
bowtie2 --sensitive-local \
-p 4 \
-x assemblies/moderate \
-1 fastq_pure/ERR1823601_pure_R1-50.fastq.gz \
-2 fastq_pure/ERR1823601_pure_R2-50.fastq.gz \
-S sam_assemblies/ERR1823601_moderate.sam
bowtie2 --sensitive-local \
-p 4 \
-x assemblies/severe \
-1 fastq_pure/ERR1823608_pure_R1-50.fastq.gz \
-2 fastq_pure/ERR1823608_pure_R2-50.fastq.gz \
-S sam_assemblies/ERR1823608_severe.sam

We can notice that there are some statistics when the alignment process finishes for each
sample.
We will convert SAM files to bam files and then we will sort the alignments in the BAM
files.

cd sam_assemblies
samtools view -S -b ERR1823587_healthy.sam > ERR1823587_healthy.
bam
samtools view -S -b ERR1823601_moderate.sam > ERR1823601_moderate.
bam
samtools view -S -b ERR1823608_severe.sam > ERR1823608_severe.bam

for i in $(ls *.bam);

do
samtools sort -@ 4 ${i} -o ${i}.sorted
done
 Shotgun Metagenomic Data Analysis   ◾    321

We need to index the sorted BAM files using “samtools index” command.

for i in $(ls *.sorted);

do
samtools index -@ 4 ${i}
done

Then, we will use “samtools idxstats” to generate some statistics from the sorted BAM files.

samtools idxstats ERR1823587_healthy.bam.sorted > ERR1823587_

healthy_stat.txt
samtools idxstats ERR1823601_moderate.bam.sorted > ERR1823601_
moderate_stat.txt
samtools idxstats ERR1823608_severe.bam.sorted > ERR1823608_
severe_stat.txt

The output of the “samtools idxstats” command is a TAB-delimited file with each line con-
sisting of the reference sequence name, sequence length, number of mapped read-­segments,
and number of unmapped read-segments. From those files, we can generate abundance
table similar to the OTU (operation taxonomic units) generated from clustering of the
amplicon-based reads in Chapter 7. For this purpose, we can use “get_count_table.py”
script, which can be cloned from GitHub using the following command:

git clone https://github.com/metajinomics/mapping_tools.git

Then, we can use that Python 2 script to generate an abundance table for each sample. So,
if you do not have Python 2 installed on your computer, you may need to install it.

python2 mapping_tools/get_count_table.py ERR1823587_healthy_stat.

txt > ERR1823587_healthy_count.txt
python2 mapping_tools/get_count_table.py ERR1823601_moderate_stat.
txt > ERR1823601_moderate_count.txt
python2 mapping_tools/get_count_table.py ERR1823608_severe_stat.
txt > ERR1823608_severe_count.txt
cd ..

We will use the output of this script for binning in the next step.

8.2.7 Binning
Above, we discussed binning as the process of separating the sequences into bins that
represent the most likely taxa. There are many programs that can do this job including
metabat2, CONCOCT, and MaxBin. Here, we will use metabat2 as an example. Metabat2
is easier to install on Anaconda or Miniconda.

conda install -c biconda metablat2

conda install -c bioconda/label/cf201901 metabat2
322   ◾    Bioinformatics

Before we perform the taxonomic binning, we need to generate sequence depth from the
sorted BAM files produced by mapping the metagenomic FASTQ reads to the de novo
assemblies. For this purpose, we will also need the tables produced by “get_count_table.
py” script above as an input with the sorted BAM file for the “jgi_summarize_bam_con-
tig_depths” function to produce a file of five columns: contig name, contig length, total
average depth, mean depth, and variance.

mkdir stats_metabat
jgi_summarize_bam_contig_depths \
--outputDepth stats_metabat/healthy_depth.txt \
sam_assemblies/ERR1823587_healthy.bam.sorted
jgi_summarize_bam_contig_depths \
--outputDepth stats_metabat/moderate_depth.txt \
sam_assemblies/ERR1823601_moderate.bam.sorted
jgi_summarize_bam_contig_depths \
--outputDepth stats_metabat/severe_depth.txt \
sam_assemblies/ERR1823608_severe.bam.sorted

Then, we can perform binning on the contigs.fasta produced by de novo assembly above.
We can copy these files in a new directory “binning” with new names.

mkdir binning
cp metag_healthy/contigs.fasta binning/healthy_contigs.fasta
cp metag_moderate/contigs.fasta binning/moderate_contigs.fasta
cp metag_severe/contigs.fasta binning/severe_contigs.fasta

The next step is to separate the contigs in the contigs files into bins; each bin represents a
species. The bins of the three samples are saved in different subdirectories inside “binning”
directory.

mkdir binning/healthy
metabat2 -i binning/healthy_contigs.fasta \
-a stats_metabat/healthy_depth.txt \
-o binning/healthy/healthy \
-t 4 -v --seed 123
mkdir binning/moderate
metabat2 -i binning/moderate_contigs.fasta \
-a stats_metabat/moderate_depth.txt \
-o binning/moderate/moderate \
-v --seed 123
mkdir binning/severe
metabat2 -i binning/severe_contigs.fasta \
-a stats_metabat/severe_depth.txt \
-o binning/severe/severe \
-v --seed 123
 Shotgun Metagenomic Data Analysis   ◾    323

The “-i” option specifies the input contigs FASTA file, “-a” option specifies the depth
file that contains contig depth averages and variances, “-o” specifies the output path
and ­prefix, “-v” for verbose, and “--seed” specifies a seed integer to replicate the same
results.
Up to this point, we have performed the taxonomic binning successfully and now we
have separated genomes for each potential species in the metagenomic sample. However,
we do not know the qualities of these genomes and to which microbial species they belong.
So, the next step, we must evaluate these genomic sequences and assess their completeness
with regard to protein-coding genes and their annotations.

8.2.8 Bin Evaluation

The binning quality is usually assessed with CheckM [11], which includes a collection of
tools for assessing the quality of the genome sequence separated from metagenomes and
also to assess the quality of genomes recovered from single cells and isolates. CheckM pro-
vides estimates of genome completeness and contamination in addition to plots and other
important reports. For this software installation, visit “https://github.com/Ecogenomics/
CheckM/wiki”. On Linux, it requires HMMER, Prodigal, and Pplacer programs to be
installed and added to the system path.

sudo apt update

sudo apt install hmmer
sudo apt install prodigal

You need to follow the installation instructions at Pplacer home page, which is available
at “https://matsen.fhcrc.org/pplacer/”, and add it to the Linux path. Then, you can install
CheckM with the following commands:

pip3 install numpy

pip3 install matplotlib
pip3 install pysam
pip3 install checkm-genome

You can also install CheckM on Anaconda using:

conda install -c bioconda checkm-genome

conda install -c bioconda/label/cf201901 checkm-genome

Now, we can run CheckM commands to assess the completeness and contamination of
the genome bins by using lineage-specific marker sets. This workflow consists of several
steps that include placing bins in the reference genome tree, assessing phylogenetic mark-
ers found in each bin, and inferring lineage-specific marker sets for each bin. These steps
are done with multiple CheckM commands but they can also be done in a single step by
using “lineage_wf” command.
324   ◾    Bioinformatics

mkdir checkM_out
mkdir checkM_out/healthy
checkm lineage_wf \
-t 4 \
-x fa \
-f checkM_out/healthy_checkm_report.txt \
binning/healthy \
checkM_out/healthy
mkdir checkM_out/moderate
checkm lineage_wf \
-t 4 \
-x fa \
-f checkM_out/moderate_checkm_report.txt \
binning/moderate \
checkM_out/moderate
mkdir checkM_out/severe
checkm lineage_wf \
-t 4 \
-x fa \
-f checkM_out/severe_checkm_report.txt \
binning/severe \
checkM_out/severe

The report (Figure 8.6) shows the bin ID, marker lineage (taxonomic rank), # genome
(number of genomes used to infer marker sets), # marker (number of marker genes), #
marker set (number of sets within the inferred markers), 0–5+ (number of times each
marker gene is identified), completeness (presence/absence of marker genes), and strain
heterogeneity (high heterogeneity indicates the contamination is from one or more closely
related organisms).
In Figure 8.6, for the moderate sample, we can notice that for the bin “moderate.4”, there
are 5449 bacterial genomes that were used to infer 104 markers genes; only 6 genes were
inferred in the bin, the completeness is 10.34, and there was no contamination. For more
details about the use of CheckM and report, refer to the program home page at “https://
ecogenomics.github.io/CheckM/”.

8.2.9 Prediction of Protein-Coding Region

This step is to annotate the single genomes recovered by binning or metagenomic assem-
blies with potential gene locations by predicting the open reading frames (ORFs). The gene

FIGURE 8.6 Genome completeness evaluation report generated by CheckM.

 Shotgun Metagenomic Data Analysis   ◾    325

annotations and polypeptides and ORFs are written to files. Gene annotation of a new
genome assembly is an important step. Since bacteria have no introns, prediction of ORFs
is easier than in the eukaryotic genome. There are many programs for ORF prediction, but
Prodigal [12] is the most commonly used one. We have installed Prodigal above. Prodigal
can predict ORFs in any genomic sequences. Thus, we can predict the ORFs in assemblies
separated by binning. In the following, we will predict the ORFs in one of the assemblies
recovered from the sample of the patient with severe sickle cell disease.

prodigal -a prod_out/healthy.faa \
-d prod_out/healthy.fnt \
-o prod_out/healthy.gbk \
-s prod_out/genes.gff \
-i binning/severe/severe.1.fa \
-p single

The “-a” option specifies the FASTA file name for the polypeptides or proteins translated
from the predicted ORFs. The “-d” option specifies the FASTA file name of the nucleotide
sequences that represent the predicted ORFs. The “-o” option specifies the predicted ORF
as features in GenBank format. The “-s” option specifies the gene annotation in GFF (gen-
eral feature format). The “-i” option specifies the input file which is the assembly. The “-p”
option specifies the procedure, which is either “single” for a single assembly or “meta” for
metagenomic assembly that may include genomes of multiple species.

8.3 SUMMARY
The metagenomic DNA is isolated from environmental samples or clinical samples in
which several microbes are present. Unlike targeted gene sequencing, shotgun metage-
nomic sequencing allows researchers to sequence the whole genomes of all organisms pres-
ent in a sample and to evaluate the microbial diversity and abundance.
Shotgun metagenomic sequencing attempts to sequence the whole genomes of a large
diverse number of microbes, each with a different genome size. Long reads produced by
PacBio and Oxford Nanopore are preferred. However, they usually have higher error rate
than the short reads. Since there are several species in the metagenomic sample, there
must be a sufficient sequencing depth to allow assembling the genomes of all species in the
sample.
Before analysis, we should make sure that we have fixed any quality problem by trim-
ming adaptors, filtering out low-quality reads, and removing technical sequences. In the
case of clinical samples, we should also remove the host DNA by aligning reads to the host
genome and then separate the unaligned reads in new FASTQ files to be used in the analy-
sis. There are two approaches for the shotgun metagenomic data analysis: the assembly-
free and de novo assembly. The assembly-free approach does not require assembling the
genomes of the species in the sample; it uses reads present in the metagenomic samples
to assign taxonomic groups by identifying unique genomic regions in the reads. Most of
the programs used for taxonomy assignment require a large amount of memory and stor-
age space. The second approach uses de novo algorithms to assemble the genomes of the
326   ◾    Bioinformatics

species in the sample. This method provides information on the total genomic DNA from
all species in a sample. After running de novo assembly, binning is used to cluster con-
tigs that apparently originated from the same species population. The recovered genome
sequences can be annotated by identifying protein-coding sequence of the ORFs.
The shotgun metagenomic sequencing is widely used to study unculturable microor-
ganisms that are difficult to culture and analyze.

REFERENCES
1. Kang DD, Li F, Kirton E, Thomas A, Egan R, An H, Wang Z: MetaBAT 2: an adaptive binning
algorithm for robust and efficient genome reconstruction from metagenome assemblies. PeerJ
2019, 7:e7359.
2. Wu Y-W, Tang Y-H, Tringe SG, Simmons BA, Singer SW: MaxBin: an automated binning
method to recover individual genomes from metagenomes using an expectation-maximiza-
tion algorithm. Microbiome 2014, 2(1):26.
3. Menzel P, Ng KL, Krogh A: Fast and sensitive taxonomic classification for metagenomics with
Kaiju. Nat Commun 2016, 7(1):11257.
4. Ounit R, Wanamaker S, Close TJ, Lonardi S: CLARK: fast and accurate classification of
metagenomic and genomic sequences using discriminative k-mers. BMC Genomics 2015,
16(1):236.
5. Wood DE, Salzberg SL: Kraken: ultrafast metagenomic sequence classification using exact
alignments. Genome Biol 2014, 15(3):R46.
6. Kim D, Song L, Breitwieser FP, Salzberg SL: Centrifuge: rapid and sensitive classification of
metagenomic sequences. Genome Res 2016, 26(12):1721–1729.
7. Nurk S, Meleshko D, Korobeynikov A, Pevzner PA: metaSPAdes: a new versatile metagenomic
assembler. Genome Res 2017, 27(5):824–834.
8. Wick RR, Schultz MB, Zobel J, Holt KE: Bandage: interactive visualization of de novo genome
assemblies. Bioinformatics 2015, 31(20):3350–3352.
9. Csárdi G, Nepusz T: The igraph software package for complex network research. In: 2006.
10. Mikheenko A, Prjibelski A, Saveliev V, Antipov D, Gurevich A: Versatile genome assembly
evaluation with QUAST-LG. Bioinformatics 2018, 34(13):i142–i150.
11. Parks DH, Imelfort M, Skennerton CT, Hugenholtz P, Tyson GW: CheckM: assessing the
quality of microbial genomes recovered from isolates, single cells, and metagenomes. Genome
Res 2015, 25(7):1043–1055.
12. Hyatt D, Chen G-L, LoCascio PF, Land ML, Larimer FW, Hauser LJ: Prodigal: prokaryotic
gene recognition and translation initiation site identification. BMC Bioinfor 2010, 11(1):119.
Index

16S rRNA gene 253, 260, 263 CD-HIT 255

centrifuge 311, 313–315, 326
ABySS 92–96 CheckM 323–324, 326
acetylation 213 chimeric 77, 81, 86, 172, 284
adaptor 4–9, 13, 18 chimeric read 81, 172, 284
adaptor dimer 18, 31–33, 46, 218 ChIP-Seq 214–217
adenine 1–2, 13 ChIPseeker 230, 236, 251
adenosine triphosphate 5 chromatin 213–215
alpha diversity 261–262, 299 chromatin immunoprecipitation see ChIP-Seq
alternative splicing 164–165, 172 chromatin remodeling 213
ammonium persulfate 5 chromatography 3
amplicon 253–257 CIGAR string 67–68, 82, 86
ANNOVAR 145, 151–159, 161 circular consensus sequences 9
archaea 49 CLARK 311, 326
ASCII 12, 14, 23, 67–68 closed-reference 255, 278, 280–282, 289, 301
assembly-free 304–305, 310–311, 325 clustering 193–194, 254–255, 257–258
Augustus 104 CNV 109–110
consensus sequence 83, 85, 89–92, 114–115, 160
bacteria 49, 97, 105, 160, 253–254, 303, 325
contig 83, 89–91, 93–100
bacteriophage 3, 46
copy number variation see CNV
base caller 11
coverage 89–91
base calling 6, 9, 11, 23, 85, 135, 256
Crick, Francis 3
Base Quality Score Recalibration see BQSR
Crohn’s disease 144
base substitution 109–110, 114, 160
cytogenetic bands 151, 154
Basic Local Alignment Search Tool see BLAST
cytosine 1–2, 13, 213, 247
BCFTools 84–85
beta diversity 261–263, 289, 298–300 de Bruijn 90–93, 97, 107–108, 126, 137, 315
binning 304–305, 321–326 de novo 89–95, 97–101, 103–105, 297–298
BioProject 20, 271–272, 274 de novo clustering 255, 280, 282, 285, 290, 301
BioPython 104 deep sequencing 89–90, 107
BLAST 55, 104, 146, 258–259, 287 defline 17, 51, 54
Bowtie 58, 70–71, 75, 87, 215 deletion 68–70, 82, 86, 109–112
BQSR 130, 135–139 dereplication 279–281
Bray-Curtis distance 261 DESeq 174, 204, 210, 212
Burrows-Wheeler transform see BWT DGEList 182–184
BUSCO 100, 103–108 differential expression 165–166, 176–177, 180, 185
BWA 58, 70–75 DNA ligase 6
BWA-backtrack 72, 74 DOT language 95
BWA-MEM 72–74, 168 double-stranded DNA see dsDNA
BWA-SW 72, 74 dsDNA 9
BWT 56, 58–63, 65, 71–72, 75, 86 dynamic programing 55

327
328   ◾    Index

EBI 50, 265 Hamiltonian path 90–92, 107

EdgeR 174–175 Hamming distance 65, 255, 257, 260
electrophoresis 3–4 haplotype 83, 104, 112, 114, 125–127
entropy 126 HaplotypeCaller 114, 130, 137
epigenetics 4–5, 9, 11, 45, 49, 64, 86 hard filtering 141–142
eQTL 165–166 hashing xiv 65, 93
escherichia coli 20, 49, 94 hash table 65, 86, 148
ETE toolkit 298 heatmap 25, 193–194, 200–202, 233, 318
Eulerian path 90, 92, 107 helicase 3
evenness 261–262 heuristic clustering 255
exon 2, 52, 63, 77, 144 hierarchical clustering 193, 255
high-throughput sequencing see HTS
Faith’s phylogenetic diversity 261–262, 297 histone 213, 216–217, 235–236, 243
Fastp 44, 46–47 HMMER 104, 323
Fastqc 18–23, 32, 36–44 homology 144, 251
Fasttree 297–298 HTS 13, 34, 45, 69–70
feature table 258, 261, 278, 280–283 HTSeq-count 173–174, 210
FeatureCounts 173, 212 hydroxyl 5, 7
Ferragina-Manzini 311
flow cell 4, 7–10, 14, 18, 25–27, 66, 134 Icarus 102, 104
FM-index 56, 62, 71, 75, 86–87 iGenomes 169
FPKM 174–175, 206–207, 210 IGV 142–143
frameshift 110, 143–144, 148, 150, 159 InDel 114, 116, 121
FreeBayes 114, 127, 129, 143 insertion 63, 68–70
Full-text Minute-space see FM-index insertion-deletion see InDel
fungi 49, 97, 303 Integrated Genomics Viewer see IGV
fusion 165–166, 211 Ion Torrent 6–7, 97–98
IQR 24
GATK 66, 114, 127, 129–132,
134–143, 160 Jaccard distance 261–262
GC content 20, 23–24, 28–29, 46–47 Jukes-Cantor 260
GenBank 50–52, 325
gene expression 163–167 k-mers 33, 65, 91–93, 97, 107
gene ontology see GO Kaiju 311–312, 326
gene transcription 163–164, 218, 241 KEGG 165, 202–203, 211–212, 239, 241–242, 250
gene transfer format see GTF Kraken 311, 326
General Feature Format see GFF Kruskal-Wallis 300
genetic drift 114 Kyoto Encyclopedia of Genes and Genomes see
genetic mosaicism 113 KEGG
Genome Reference Consortium see GRC
L50 96, 102–103, 107
Genome Variant Call Format see gVCF
lactose operon 164
GFF 51–52
Levenshtein distance 65, 257
ggplot2 104
LG50 96
Gilbert, Walter 3
local sequence alignment 55, 258
global sequence alignment 55
locus 90, 172
GO 66, 165, 202–203, 211–212, 239–241, 250
logo 245, 249
graphviz 95
lookup table 62–63
GRC 50
greedy 90–91, 107, 255, 312 MAFFT 297–298
greengenes database 255, 281, 291 MAG 304
GTF 51–53 manifest file 268–269, 274–275
guanine 1–2, 13, 213, 247 Maxam, Allan M. 3
gVCF 130, 137–138 MaxBin 304, 321, 326
 Index   ◾    329

maximum likelihood 261 peak calling 215–217, 223

Mendel, Gregor 109 phosphorylation 213
message passing interface see MPI phred 12–14, 23–24
MetaBAT 304, 326 PICARD 79, 86, 130, 132, 134–135, 210
Metaeuk 104 Pielou’s Evenness 262
Metagenome-Assembled Genome see MAG plants 49, 253
metagenomics 253–254, 265, 301–302, 304, 326 polymerase chain reaction 4
methylation 213–214 position-specific scoring matrix see PSSM
microscopy 9 prefix tree 56
missense 109, 114, 143–145, 148, 150 principal coordinate analysis see PCoA
molecular clock 261 prodigal 104, 323, 325–326
motif 215–217, 225, 243–250 PSI-BLAST 146
MPI 93 PSSM 146
mpileup 79, 84, 114–116, 121, 123 purine 1, 109, 247
MSA 146 pyrimidine 1, 109, 247
multiple sequence alignment see MSA pyrophosphate 5–6
multiplexing 5, 254, 270, 303–304, 306 pyrosequencing 5–6

N50 96, 102–103, 107 Q score 12–13

naïve Bayes 289, 291–293, 304 quasi-negative binomial 178–179, 196
nanopore sequencing 9, 90 QUAST 100–103, 107–108, 317–318, 326
National Center for Biotechnology Information 50
Neanderthal 49 rank table 62–63
Needleman-Wunsch 55, 259 RDP 258–259
negative binomial 178–180, 185–189 RefSeq 50, 228, 311
NEWICK 298 regular expression 81
NJ 177, 257, 261 representative sequence 255, 258–261
nodes 56–57, 91–93 ribonucleic acid see RNA
nonsense 110, 144, 148, 150 Ribosomal Database Project 258–259, 302
nonsynonymous codon 109, 146 ribosome 2, 254
normalization 174, 176, 182–183, 186–187 richness 261–262, 299
novoalign 65 RIN 18
nucleosome 213, 216 RNA integrity see RIN
RNA-Seq 163, 165–169
oligo 6, 8, 249 Robert Holley 3
ONT 9–10, 90, 203, 239–240, 305 Roche 454 5–7
open reading frame see ORF RPKM 174–175, 210, 212
open-reference clustering 255, 278, 280, 282, 289,
301 Samtools 53–54
operon 2, 164 Sanger, Frederick 3
ORF 2, 164, 325 Sanger sequencing 4, 90, 254
ortholog 103–105, 108 SARS-CoV-2 97, 99, 116, 118, 127, 155–157, 159–160
overlap 81, 89–93, 107 scaffold 89, 94–100, 102–103, 304–305, 316–320
overlap-consensus 90–91, 107 seed-and-extend 65
Oxford Nanopore Technologies see ONT seeding 73
SEPP 104
PacBio 9–10 sequence alignment 53, 55, 63, 65, 70, 87, 90
Pacific Bioscience see PacBio sequencing depth 41, 45, 90, 114, 160, 175, 190
PairHMM 126 Shannon’s diversity index 261–262
pairwise alignment 90–91, 107, 126, 259 sickle 305, 310, 325
Pandas 104 SIFT 145–148, 152, 161
parsimony 261 singleton 42, 46, 309
PCoA 300 SMRT 9, 90, 98, 166
330   ◾    Index

SNP 135–136, 139–143 transition 109, 256–257

SnpEff 145, 148–151, 161 translocation 109–110
SNV 110, 114, 116 transversion 109
SOAP 65, 87 trie 11, 56, 65, 87
SOLiD 6, 8, 135, 156, 253 trimmomatic 42–47
sonication 4, 214 triphosphate 4–5
SPAdes 92, 97–100, 102, 105–108, TruSeq 97
315–316 TSS 144, 164, 233–238
splicing site 148, 164
SRA-toolkit 14–15, 70, 215, 265, 305 ubiquitylation 213
ssDNA 3, 5, 7, 9–10 UniFrac 261–263, 297, 300
STAR 58, 70–71, 76–78 untranslated region see UTR
stop codon 110, 143 UPGMA 261
stop-gain 143, 148 Uracil 1
structural variant 113 USEARCH 255
suffix array 55, 57–60, 65, 72, 77–78, 86–87 UTR 144, 148
suffix tree 55–58, 65, 87, 246
variant call format see VCF
sulfurylase 5
variant calling 113–116
SV 110
Variant Quality Score Recalibration see VQSR
SW 46, 55, 65, 72–74, 326
VCF 110–113, 115–116
tag 69–70, 80, 134, 215–216 vertexes 91
tandem repeats 89 virion 1–2
TATA box 164 virus 99, 116, 253
TGS 8–9, 45, 85 volcano 199–200, 208–209
third-generation sequencing 8 VQSR 139–141
thymine 1–2, 13, 213, 247 VSEARCH 258–259, 278, 280–282
TPM 175
Watson, James 3
transcription start site see TSS
WPGMA 261
transcripts per million see TPM