User Manual for

Genomic Association and Prediction Integrated Tool

(Version 3)

Last updated on May 4, 2022

Zhiwu Zhang Laboratory

 GAPIT User Manual

Disclaimer: While extensive testing has been performed by the Zhiwu Zhang Lab at (2014 to present) at
Washington State University and Edward Buckler Lab (2012-2014) at Cornell University, respectively.
Results are, in general, reliable, correct, and appropriate. However, results are not guaranteed for any
specific set of data. We strongly recommend that users validate GAPIT results with other software
packages, such as SAS and TASSEL.

Support documents: Extensive support documents, including this user manual, source code,
demonstration scripts, data, and results, are available at GAPIT website hosted by Zhiwu Zhang
Laboratory: http://zzlab.net/GAPIT

Questions and comments: To benefit GAPIT community, questions and comments should be addressed
to GAPIT forum: https://groups.google.com/forum/#!forum/gapit-forum. The GAPIT team members will
periodically go through these questions and comments and address them accordingly. For countries with
restriction on Google, questions and comments are welcome to Jiabo Wang by email:
[email protected].

Citation: Multiple statistical methods are implemented in GAPIT version 1, 2 and 3. Citations of GAPIT
vary depending on methods and versions used in the analysis:
Method Method paper Version 11 Version 22 Version 33

General Linear Model (GLM) Price et al, 2006, Nature Genetics4 ✓ ✓ ✓

Mixed Linear Model (MLM) Yu et al, 2005, Nature Genetics5 ✓ ✓ ✓
Compression MLM (CMLM) Zhang et al, 2010, Nature Genetics6 ✓ ✓ ✓
gBLUP Zhang et al, 2007, J. Anim. Science7 ✓ ✓ ✓
Enriched CMLM Li et al, 2014, BMC Biology8 ✓ ✓
SUPER Wang et al, 2014, PLoS One9 ✓ ✓
MLMM Segura et al, 2012, Nature Genetics10 ✓
FarmCPU Liu et al, 2016, PloS Genetics11 ✓
cBLUP and sBLUP Wang et al, 2019, Heredity12 ✓
BLINK Huang et al, 2019, GigaScience13 ✓
Note: These references are listed in section of Reference.

The GAPIT project is partially supported by USDA, DOE, NSF, the Agricultural Research Center at
Washington State University, and Washington Grain Commission

1
 GAPIT User Manual

Table of Contents

1 INTRODUCTION .......................................................................................................................................... 5

1.1 WHY GAPIT? ......................................................................................................................................... 5

1.2 GETTING STARTED ................................................................................................................................. 6
1.3 HOW TO USE THE GAPIT USER MANUAL? ............................................................................................... 7
1.4 HOW TO CITE GAPIT? .................................................................................................................................. 7

2 INPUT DATA ........................................................................................................................................... 8

2.1 PHENOTYPIC DATA ................................................................................................................................. 8

2.2 GENOTYPIC DATA ................................................................................................................................... 9
2.2.1 HAPMAP FORMAT .................................................................................................................................................9
2.2.2 NUMERIC FORMAT ..............................................................................................................................................10
2.3 KINSHIP ................................................................................................................................................ 11
2.4 COVARIATE VARIABLES ........................................................................................................................ 11

3 GWAS ...................................................................................................................................................... 13

3.1 GWAS MODEL OVERVIEW .................................................................................................................... 13

3.2 MODEL SELECTION ............................................................................................................................... 14
3.3 MODEL DESCRIPTION............................................................................................................................ 15
3.4 MODEL JUSTIFICATION ......................................................................................................................... 15
3.5 GAPIT SYNTAX .................................................................................................................................... 15
3.6 MIXED LINEAR MODEL (MLM) ............................................................................................................ 16
3.7 COMPRESSED MLM (CMLM) .............................................................................................................. 17
3.8 GENERAL LINEAR MODEL (GLM) ........................................................................................................ 17
3.9 P3D/EMMAX ....................................................................................................................................... 17
3.10 SUPER ................................................................................................................................................. 17
3.11 MULTIPLE LOCUS MIXED LINEAR MODEL (MLMM) .......................................................................... 18
3.12 FARMCPU .......................................................................................................................................... 18
3.13 BLINK ............................................................................................................................................... 18

4 GENOMIC SELECTION ............................................................................................................................... 19

4.1 GENOMIC BLUP ................................................................................................................................... 19

4.2 COMPRESSED GBLUP ................................................................................................................................. 20
4.3 SUPER GBLUP......................................................................................................................................... 20

5 OUTPUT RESULTS ............................................................................................................................... 21

5.1 PHENOTYPE DIAGNOSIS ............................................................................................................................... 22

5.2 MARKER DENSITY ...................................................................................................................................... 22
5.3 LINKAGE DISEQUILIBRIUM DECAY .................................................................................................................. 23
5.4 HETEROZYGOSIS ........................................................................................................................................ 23

2
 GAPIT User Manual

5.5 PRINCIPAL COMPONENT (PC) PLOT ................................................................................................................ 24

5.6 KINSHIP PLOT............................................................................................................................................ 24
5.7 NEIGHBOR-JOINING (NJ)-TREE...................................................................................................................... 25
5.8 QQ-PLOT................................................................................................................................................. 25
5.9 MANHATTAN PLOT .................................................................................................................................... 26
5.10 ASSOCIATION TABLE ................................................................................................................................. 27
5.11 ALLELIC EFFECTS TABLE.............................................................................................................................. 27
5.12 COMPRESSION PROFILE ............................................................................................................................. 28
5.13 THE OPTIMUM COMPRESSION .................................................................................................................... 29
5.14 MODEL SELECTION RESULTS ....................................................................................................................... 30
5.15 MULTIPLE TRAITS, ENVIRONMENTS, OR MODELS .............................................................................................. 30
5.16 GENOMIC PREDICTION .............................................................................................................................. 31
5.17 DISTRIBUTION OF BLUPS AND THEIR PEV. ..................................................................................................... 32
5.18 INTERACTIVE GWAS PLOT ......................................................................................................................... 33

6 TUTORIALS ............................................................................................................................................... 34

6.1 A BASIC SCENARIO ..................................................................................................................................... 34

6.2 ENHANCED COMPRESSION ........................................................................................................................... 34
6.3 USER-INPUTTED KINSHIP MATRIX AND COVARIATES........................................................................................... 35
6.4 MULTIPLE GENOTYPE FILES .......................................................................................................................... 35
6.5 NUMERIC GENOTYPE FORMAT ...................................................................................................................... 36
6.6 NUMERIC GENOTYPE FORMAT IN MULTIPLE FILES.............................................................................................. 36
6.7 FRACTIONAL SNPS FOR KINSHIP AND PCS ....................................................................................................... 37
6.8 MEMORY SAVING ...................................................................................................................................... 37
6.9 MODEL SELECTION ..................................................................................................................................... 38
6.10 SUPER ................................................................................................................................................. 38
6.11 MLMM ................................................................................................................................................ 38
6.12 FARM-CPU ............................................................................................................................................ 39
6.13 BLINK .................................................................................................................................................. 39
6.14 MULTIPLE MODEL .................................................................................................................................... 39
6.15 GBLUP ................................................................................................................................................. 40
6.16 CBLUP .................................................................................................................................................. 40
6.17 SBLUP .................................................................................................................................................. 40

7 PROTOTYPE.............................................................................................................................................. 41

7.1 STATISTICAL POWER COMPARISON AMONG METHODS ........................................................................................ 41

7.2 GENOMIC SELECTION .................................................................................................................................. 42
7.3 CROSS VALIDATION WITH REPLACEMENT.......................................................................................................... 42
7.4 CROSS VALIDATION WITHOUT REPLACEMENT .................................................................................................... 43
7.5 CONVERT HAPMAP FORMAT TO NUMERICAL .................................................................................................... 44

8 APPENDIX ................................................................................................................................................ 45

8.1 GAPIT BIOGRAPHY .................................................................................................................................... 45

8.2 FREQUENTLY ASKED QUESTIONS ................................................................................................................... 46
1. HOW TO CITE GAPIT? ..........................................................................................................................................46

3
 GAPIT User Manual

2. WHAT DO I DO IF I GET FRUSTRATED? .................................................................................................................46

3. WHY GAPIT HAS DIFFERENT RESULTS FROM OTHER SOFTWARE? ......................................................................46
4. THERE ARE MANY METHODS IMPLEMENTED IN GAPIT, WHICH ONE SHOULD I USE? .........................................46
5. HOW MANY PCS TO INCLUDE? .............................................................................................................................46
6. IS IT FEASIBLE I COMPARE DIFFERENT MODELS ON MY DATA? ............................................................................46
7. HOW DO I REPORT AN ERROR? .............................................................................................................................46
8. WHAT SHOULD I DO WITH “ERROR IN FILE (FILE, "RT") : CANNOT OPEN THE CONNECTION”? ............................46
9. WHAT SHOULD I DO WITH “ERROR IN GAPIT (... : UNUSED ARGUMENT(S) ...”? .................................................47
10. HOW DEAL WITH “ERROR IN SOLVE.DEFAULT(CROSSPROD(X, X)) : SYSTEM IS COMPUTATIONALLY
SINGULAR”? .................................................................................................................................................................47
11. HOW TO FIX THE ERROR OF USING COVARIATES FROM STRUCTURE AS FIXED EFFECTS? ..............................47
12. SHOULD I REMOVE SNPS WITH MAF BELOW 5%? ............................................................................................47
13. MY TRAIT WAS MEASURED IN MULTIPLE ENVIRONMENTS, HOW DO I USE THEM SIMULTANEOUSLY? ..............47
14. IS IT OK TO ANALYZE BINARY TRAITS (CASE-CONTROL) WITH GAPIT? ...........................................................47
15. DOES NORMALITY TRANSFORMATION HELP?.....................................................................................................47
16. SHOULD I USE PCS OR Q MATRIX? .....................................................................................................................47
REFERENCES ............................................................................................................................................. 48

4
 GAPIT User Manual

1 INTRODUCTION
1.1 Why GAPIT?
GAPIT implemented a series of methods for Genome Wide Association (GWAS) and Genomic Selection
(GS). The GWAS models include General Linear Model (GLM), Mixed Linear Model (MLM or Q+K),
Compressed MLM (CMLM), Enriched CMLM, SUPPER, Multiple Loci Mixed Model (MLMM),
FarmCPU and BLINK. The GS models include gBLUP, Compressed BLUP, and SUPER BLUP.

Figure 1.1. Methods implemented in GAPIT for GWAS and genomic selection. All the methods support
GWAS, including General Linear Model (GLM), Mixed Linear Model (MLM), Compressed MLM
(CMLM), Enriched CMLM (ECMLM), Settlement of MLM Under Progressively Exclusive Relationship
(SUPER), Fixed and random model Circulating Probability Unification (FarmCPU), and Bayesian-
information and Linkage-disequilibrium Iteratively Nested Keyway (BLINK). Some of these methods
support genomic selection, including MLM, CMLM, ECMLM, SUPER, and FarmCPU. The remaining
(GLM and BLINK) can be used for breeding through marker assisted selection (MAS).

5
 GAPIT User Manual

1.2 Getting Started

GAPIT is a package that is run in the R software environment, which can be freely downloaded from
http://www.r-project.org or http://www.rstudio.com. There are two sources to install GAPIT package.
Zhiwu Zhang Lab website: Or GitHub:
source("http://zzlab.net/GAPIT/GAPIT.library.R")
source("http://zzlab.net/GAPIT/gapit_functions.txt") install.packages("devtools")
devtools::install_github("jiabowang/GAPIT3",force=TRUE)
library(GAPIT3)

The easiest way is to COPY/PASTE GAPIT tutorial script. Here are example code and outputs:
#Import data from Zhiwu Zhang Lab
myY <- read.table("http://zzlab.net/GAPIT/data/mdp_traits.txt", head = TRUE)
myGD=read.table(file="http://zzlab.net/GAPIT/data/mdp_numeric.txt",head=T)
myGM=read.table(file="http://zzlab.net/GAPIT/data/mdp_SNP_information.txt",head=T)

#GWAS
myGAPIT=GAPIT(
Y=myY[,c(1,2,3)], #fist column is ID
GD=myGD,
GM=myGM,
PCA.total=3,
model=c("FarmCPU", "Blink"),
Multiple_analysis=TRUE)

As demonstrated above, users can specify any one or multiple models. GAPIT accepts multiple input data
formats, including both numeric, hapmap, and PLINK genotype formats. GAPIT produces comprehensive
reports to interpret data and results in publication ready formats. For examples, the distribution of marker
density and decay of linkage equilibrium inform user if the markers are dense enough. When GWAS were
conducted with multiple traits, environments, or multiple models, GAPIT produces the integrated
Manhattan plots with overlapped associated markers highlighted. The above analysis should be completed
within couple minutes. In your current R working directory, you should find multiples files with three types
of extensions: pdf, csv, and txt. The three types of the Manhattan plots are displayed above.

6
 GAPIT User Manual

1.3 How to use the GAPIT user manual?

The next three chapters (2-5) describe details on the input data, GWAS, GS, and output results. Chapter 6
presents scenarios to demonstrate the applications. Chapter 7 is for users to use GAPIT for prototyping. The
last chapter (8) lists frequently questions and answers. Before reading the next three chapters, we
recommend you go directly to the tutorial chapter and run other tutorials.

1.4 How to cite GAPIT?

Although historical version of GAPIT (1 and 2) are available, the newest version (3) is recommended for
full support from GAPIT team. Citations should specify the version and models used. For an example, a
GWAS run by GAPIT version 3 using BLINK can cite as “The GWAS was conducted by GAPIT (version
3)3 using BLINK model13” . A GS with run by GAPIT version 3 using gBLUP/cBLUP can cite as “GS was
conducted by GAPIT (version 3)3 using gBLUP model7 and cBLUP model12”.

7
 GAPIT User Manual

2 Input Data
There are six types of input data: phenotype (Y), genotype in hapmap format (G), genotype data in
numerical format (GD), genotype map (GM), kinship (K), and covariate variables (CV), see Table 2.1.
Phenotypic data must be provided, and the rest are optional, including genotype data, map, kinship, and
covariate. Kinship can be provided by users or be generated from genotype data, or even omitted by suing
BLINK method. Genotypic data may not be needed for genomic prediction if the kinship matrix is provided.
Covariate variables (fixed effects), such as population structure represented by the Q matrix (subpopulation
proportion) or principal components (PCs), are optional. GAPIT provides the option to calculate PCs from
the genotypic data. All input files should be saved as a “Tab” delimited text file.

Notice: It is important that each taxa name is spelled, punctuated, and capitalized (R is case sensitive) the
same way in each of the input data sets. If this is not done, they will be excluded from the analysis.
Additionally, the taxa names must not be numeric.

Table 2.1 Gallery of GAPIT input data

Param
Default Options Description Tutorial files*
eter
Y NULL User Phenotype mdp_traits.txt
KI NULL User Kinship Matrix KSN.txt
CV NULL User Covariate Variables mdp_PC.txt
G NULL User Genotype Data in Hapmap Format mdp_genotype_test.hmp.txt
GD NULL User Genotype Data in Numeric Format mdp_numeric.txt
Genotype Map for Numeric
GM NULL User mdp_SNP_information.txt
Format

The tutorial file can be downloaded at: http://zzlab.net/GAPIT/GAPIT_Tutorial_Data.zip. These files can
read into R with following commands:
#Phenotypic Data
#myY <- read.table("mdp_traits.txt", head = TRUE)

#HapMap genotype format

myG <- read.delim("mdp_genotype_test.hmp.txt", head = FALSE)

#Numerical genotype format

#--------------------A pair of Genotypic Data and map files-------------------------------
myGD <- read.table("mdp_numeric.txt", head = TRUE)
myGM <- read.table(“mdp_SNP_information.txt", head = TRUE)

#Kinship matrix
myKI <- read.table("KSN.txt", head = FALSE)

#covaraite variables (such as population structure represented by Q matrix or PC)

myCV <- read.table("mdp_PC", head = TRUE)

2.1 Phenotypic Data

The user has the option of performing GWAS on multiple phenotypes in GAPIT. This is achieved by
including multiple phenotype columns in phenotypic file. Taxa names should be in the first column of the
phenotypic data file and the remaining columns should contain the observed phenotype from each
individual. Missing data should be indicated by either “NaN” or “NA”. The first ten observations in the
tutorial data (mdp_traits.txt) are displayed as follows:

8
 GAPIT User Manual

The file is “Tab” delimited. The first row consists of column labels (i.e., headers). The column labels
indicate the phenotype name, which is used for the remainder of the analysis.

The phenotype file can be input to R by typing command line:

myY <- read.table("mdp_traits.txt", head = TRUE)

2.2 Genotypic Data

Genotypic data are required for GWAS, but are optional for GS. In the later case, genomic prediction is
performed using a kinship matrix provided by the user. GAPIT accepts genotypic data in either HapMap
format or in numeric format.

2.2.1 Hapmap Format

Hapmap is a commonly used format for storing sequence data where SNP information is stored in the rows
and taxa information is stored in the columns. This format allows the SNP information (chromosome and
position) and genotypes of each taxa to be stored in a single file.

The first 11 columns display attributes of the SNPs and the remaining columns show the nucleotides
observed at each SNP for each taxa. The first row contains the header labels and each remaining row
contains all the information for a single SNP. The first five individuals on the first seven SNPs from the
tutorial data (mdp_genotype.hmp.txt) are presented below.

This file can be read into R by typing the following command line:
myG <- read.table("mdp_genotype_test.hmp.txt", head = FALSE)

9
 GAPIT User Manual

Although all of the first 11 columns are required, GAPIT uses only 3 of these: the “rs” column, which is
the SNP name (e.g. “PZB00859.1”); the “chrom” column, which is the SNP’s chromosome; and the “pos”,
which is the SNP’s base pair (bp) position. It is sufficient to fill in the requested information in the remaining
eight columns with “NA”s. To be consistent with HapMap naming conventions, missing genotypic data are
indicated by either “NN” (double bit) or “N” (single bit).

For genotypic data in HapMap format, GAPIT accepts genotypes in either double bit or in the standard
IUPAC code (single bit) as following:

Genotype AA CC GG TT AG CT CG AT GT AC
Code A C G T R Y S W K M

By default, the HapMap numericalization is performed so that the sign of the allelic effect estimate (in the
GAPIT output) is with respect to the nucleotide that is second in alphabetical order. For example, if the
nucleotides at a SNP are “A” and “T”, then a positive allelic effect indicates that “T” is favorable. Selecting
“Major.allele.zero = TRUE” in the GAPIT() function will result in the sign of the allelic effect estimate
being with respect to the minor allele. In this scenario, a positive allelic effect estimate will indicate that the
minor allele is favorable.

2.2.2 Numeric format

GAPIT also accepts the numeric format. The order of taxa and SNPs is reversed from the HapMap format.
Columns are used for SNPs and rows are used for taxa in the numeric format. This format is problematic in
Excel because the number of SNPs used in a typical analysis exceeds the Excel column limit. Additionally,
this format does not contain the chromosome and position of the SNPs. Therefore, two separate files must
be provided to GAPIT. One file contains the numeric genotypic data (called the “GD” file), and the other
contains the position of each SNP along the genome (called the “GM” file).
Note: The SNPs in the “GD” and “GM” files NEED to be in the same order.

Homozygotes are denoted by “0” and “2” and heterozygotes are denoted by “1” in the “GD” file. Any
numeric value between “0” and “2” can represent imputed SNP genotypes. The first row is a header file
with SNP names, and the first column is the taxa name. The example file (mdp_numeric.txt from tutorial data set)
is as following:

taxa PZB00859.1 PZA01271.1 PZA03613.2 PZA03613.1

33-16 2 0 0 2
38-11 2 2 0 2
4226 2 0 0 2
4722 2 2 0 2
A188 0 0 0 2
…

This file can be read into R by typing the following command line:
myGD <- read.table("mdp_numeric.txt", head = TRUE)

The genetic map (“GM”) file contains the name and location of each SNP. The first column is the SNP id,
the second column is the chromosome, and the third column is the base pair position. As seen in the example,
the first row is a header file. The example file (mdp_SNP_information.txt from tutorial data set) is as following:
10
 GAPIT User Manual

Name Chromosome Position

PZB00859.1 1 157104
PZA01271.1 1 1947984
PZA03613.2 1 2914066
PZA03613.1 1 2914171
PZA03614.2 1 2915078
…

This file is read into R by typing the following command line:

myGM <- read.table("mdp_SNP_information.txt", head = TRUE)

2.3 Kinship
The kinship matrix file (called “KI” in GAPIT) is formatted as an n by n+1 matrix where the first column
is the taxa name, and the rest is a square symmetric matrix. Unlike the other input data files, the first row
of the kinship matrix file does not consist of headers. The example (KSN.txt from tutorial data set) is as following:

This file is read into R by typing the following command line:

myKI <- read.table("KSN.txt", head = FALSE)

2.4 Covariate variables

A file containing covariates (called “CV” in GAPIT) can include information such as population structure
(commonly called the “Q matrix”), which are fitted into the GWAS and GS models as fixed effects. These
files are formatted similarly to the phenotypic files. Specifically, the first column consists of taxa names,
and the remaining columns contain covariate values. The first row consists of column labels. The first
column can be labeled “Taxa”, and the remaining columns should be covariate names. The example file
(mdp_population_structure.txt from tutorial data set) is as following

11
 GAPIT User Manual

This file is read into R by typing the following command line:

myCV <- read.table("mdp_population_structure.txt", head = TRUE)

12
 GAPIT User Manual

3 GWAS
3.1 GWAS model overview
Currently, GAPIT has implemented more than ten models. The similarity and difference among seven
milestone models are summarized in the figure below. The simplest model (t test) is to directly detect the
association between a phenotype (y) and markers (Si) one at a time, where i=1 to m, and m is number of
markers. When a cofactor, such as population structure (Q) is introduced through a general linear model
(GLM), the cofactor may not only account residuals (e) partially, but also adjust some effect that does not
belong to the testing markers and consequently reduce false positives. The mixed linear model (MLM)
applies the same principle by adding individuals’ genetic effects as random cofactor effects with variance
structure defined by the kinship (K) among individuals. In both Q or Q+K models, Q and K stay the same.
There are no cofactors that are adjusted by the marker tests.

Inclusion of cofactors benefits the

reduction of false positives for testing
markers in GLM and MLM. The
disadvantage is these cofactors are also
confounded with the testing markers. In
MLM particularly, the kinship defines the
genetic effect of individuals which equal
the sum of causal genes. Many know
genes identified by GLM had signals
below threshold using MLM14.

The compressed MLM (CMLM) was

proposed to reduce the confounding
problem of MLM6. Individuals are
compressed into groups. The individual
genetic effects are replaced with the
group genetic effects. Correspondingly,
kinship among individuals is replaced
with kinship among groups with
grouping maximized using maximum
likelihood method. The optimization of
kinship among groups further improve
statistical power8 in the enriched CMLM
(ECMLM).

GLM and MLM are the special cases of

CMLM which is a general format. When
number of groups is forced to be one in
CMLM, CMLM becomes GLM.
Similarly, when number of groups is
forced to be the number of individuals in
CMLM, CMLM becomes MLM. The
optimization of groping improves
statistical power15.

13
 GAPIT User Manual

The optimization of groping in CMLM and optimization of kinship among groups in ECMLM are
thoroughly based on phenotypes. There is no impact from marker tests. The situation was changed by the
multiple loci mixed model (MLMM). Through marker association tests, the associated markers are fitted
as the cofactors for marker test. The cofactors are adjusted through forward and backward stepwise
regression of mixed model. However, both Q and K remain unchanged.

In the SUPER method, K is derived from the associated markers and is adjected accordingly by the marker
tests. As the K is derived from a smaller number of markers than the K in MLM and MLMM that are derived
from all the markers, the confounding between K and some of the markers becomes more severe. SUPER
eliminate the confounding by using the complimentary kinship derived from associated markers except for
the ones that are in strong linkage disequilibrium (LD) with the testing markers under a user-defined
threshold.

To eliminate the ambiguity of determining associated markers are in LD with a testing marker, FarmCPU
completely removes the confounding from kinship by using a fixed-effect model without a kinship derived
either from all markers, or associated markers. Instead, the kinship derived from the associated markers is
used to select the associated markers using the maximum likelihood method. This process overcomes the
model overfitting problems of stepwise regression. FarmCPU uses both the fixed effect model and the
random effect model iteratively.

In both SUPER and FarmCPU models, the bin approach is used to avoid selecting markers from the same
locations with bin size and the number of bins optimized using the maximum likelihood method. The
underlying assumption is that causal genes are distributed equally across the genome. BLINK eliminates
the assumption to improve statistical power by using the linkage disequilibrium (LD) method. Markers are
sorted with the most significantly associated maker on the top as reference. The remaining markers are
removed if they are in LD with the most associated marker. Among the remaining makers, the most
significantly associated maker is selected as the reference. The process is repeated until no markers can be
removed. The random effect model in FarmCPU to select associated markers using the maximum likelihood
method remains a high computing cost for a large number of individuals. BLINK approximate the
maximum likelihood using Bayesian Information Content (BIC) in a fixed-effect model to eliminate the
computational burden.

3.2 Model selection

With the multiple models implemented in GAPIT, a common question is which to choose. Many people
make the selection based on their trust gained over experience. For example, some researchers must choose
GLM implemented in PLINK16 because it is the only software accepted by the reviewers and editors in their
fields. In general, computing efficiency and statistical power should be the criteria for the selection.

Two models use the fixed-effect model only which is the most computing efficient, including GLM and
BLINK. FarmCPU is a hybrid that uses both the fixed-effect model and the random effect model. The rest
use a fixed and random effects mixed model which is computationally expensive, including MLM, CMLM,
ECMLM, SUPER, and MLMM. CMLM uses groups and is cubic time faster than MLM. Due to additional
optimizations, ECMLM and SUPER are slower than CMLM. For a trial analysis, GLM and BLINK are
good to start with.

Regarding statistical power, multiple loci models (e.g. MLMM, FarmCPU, and BLINK) are superior to the
rest. Within multiple loci model category, FarmCPU is superior to MLMM11 and BLINK is superior to
FarmCPU13. Within the single locus model category, MLM is superior to GLM5, CMLM is superior to
MLM6, ECMLM is superior to ECMLM8, SUPER and MLM are superior to MLM9,10. These relationships
14
 GAPIT User Manual

are summarized by the model stairs in the first chapter. The method on a higher stair has higher statistical
power than the one on a lower stair. The magnitude of the differences among models may change from case
to case, however, their order stays the same. The inversion of the order has not been found. Therefore,
BLINK is selected as the default GAPIT model because of its high computing efficiency and statistical
power. Users are welcome to use the following statement to justify the usage of BLINK.
“In addition to the capability to incorporate principal components as covariates to reduce false positives
due to population stratification, BLINK iteratively incorporates associated markers as covariates for
testing markers to eliminate their connection to the cryptic relationship among individuals. The associated
markers are selected according to linkage disequilibrium, optimized for Bayesian information content, and
reexamined across multiple tests to reduce false negatives”.

3.3 Model description

The detailed model description is critical for readers to understand exactly how the analyses were performed
or to replicate the analyses. As all the implemented models are well described elsewhere, the model
description should focus on the covariates that are specific to the analyses. All covariates should be
described in detail, including the levels for the category covariates. Here is an example:
“GAWAS was conducted by GAPIT (version 3)3 using BLINK model17. The covariate variables include the
first three principal components derived from all the markers and the origin-group. The origin group was
coded as indicators (0/1) for each of the origin groups except the last one to avoid linear dependency”.

3.4 Model justification

It was found during the development of FarmCPU that causal genes can be detected even when they are
confounded with population structure and population structure such as the first three principal components
were fitted as covariates for testing markers. As an anonymous FarmCPU reviewer suggested, fitting several
PCs does not hurt the degree of freedom very much, however, it helps in situations there are non-genetic
effects associated with population structure during phenotyping. Otherwise, a false positive marker would
appear to capture the non-genetic effect. Therefore, fitting several PCs is recommended for all analyses.
The related justification is as follows.

“Principal component analysis was performed with GAPIT (version 3)3 using all available SNPs. GAWAS
was conducted by GAPIT (version 3)3 using BLINK model17. The first principal components were fitted as
covariate variables to reduce the false positives due to population stratification”.

3.5 GAPIT Syntax

GAPIT can be executed by calling “GAPIT()” with inputs and parameters included in “()”. The inputs
include phenotypes, genotype data, genetic map, covariate variables. The general parameters include
number of PCs as covariates and models. More general parameters can be found in Table 3.5.1.

15
 GAPIT User Manual

There are also parameters specific to models. For example, the CMLM model involves number of groups.
These model specific parameters will be described within the sections of specific models.
Table 3.5.1. GAPIT input parameters.
Parameter Default Options Description
GLM, MLM, CMLM, SUPER,
model Blink Choose one or multiple models to conduct GWAS
MLMM, FarmCPU, and Blink
kinship.algorithm VanRaden Zhang, Loiselle and EMMA Algorithm to Derive Kinship from Genotype
complete, ward, single,
kinship.cluster average Clustering algorithm to group individuals based on their kinship
mcquitty, median, and centroid
kinship.group Mean Max, Min, and Median Method to derive kinship among groups
LD.chromosome NULL User Chromosome for LD analysis
LD.location NULL User Location (center) of SNPs for LD analysis
LD.range NULL User Range around the Central Location of SNPs for LD Analysis
PCA.total 0 >0 Total Number of PCs as Covariates
PCA.scaling None Scaled, Centered.and.scaled Scale And/Or Center And Scale The SNPs Before Conducting PCA
SNP.FDR 1 >0 and <1 Threshold to Filter SNP on FDR
SNP.MAF 0 >0 and <1 Minor Allele Frequency to Filter SNPs in GWAS Reports
SNP.effect Add Dom Genetic Model
SNP.P3D TRUE FALSE Logic Variable to Use P3D or Not for Testing SNPs
SNP.fraction 1 >0 and <1 Fraction of SNPs Sampled to Estimate Kinship and PCs
SNP.test TRUE FALSE Logic Variable to Test SNPs or Not

3.6 Mixed Linear Model (MLM)

MLM includes both fixed and random effects. Including individuals as random effects gives an MLM the
ability to incorporate information about relationships among individuals. This information about
relationships is conveyed through the kinship (K) matrix, which is used in an MLM as the variance-
covariance matrix between the individuals. When a genetic marker-based kinship matrix (K) is used jointly
with population structure (commonly called the “Q” matrix, and can be obtained through STRUCTURE18
or conducting a principal component analysis19), the “Q+K” approach improves statistical power compared
to “Q” only20. An MLM can be described using Henderson’s matrix notation as follows:

Y = Xβ + Zu + e, (1)

where Y is the vector of observed phenotypes; β is an unknown vector containing fixed effects, including
the genetic marker, population structure (Q), and the intercept; u is an unknown vector of random additive
genetic effects from multiple background QTL for individuals/lines; X and Z are the known design matrices;
and e is the unobserved vector of residuals. The u and e vectors are assumed to be normally distributed with
a null mean and a variance of:

u G 0 
Var   =  
e  0 R (2)

where G = σ2aK with σ2a as the additive genetic variance and K as the kinship matrix. Homogeneous
variance is assumed for the residual effect; i.e., R= σ2eI, where σ2e is the residual variance. The proportion
of the total variance explained by the genetic variance is defined as heritability (h2).

16
 GAPIT User Manual

σ a2
h2 =
σ a2 + σ e2 , (3)

3.7 Compressed MLM (CMLM)

As kinship is derived from all the markers, incorporating with the kinship for testing markers in a MLM
causes the confounding between the testing markers and the individuals’ genetic effects with variance
structure defined by the kinship. To reduce the confounding, individuals are replaced by their corresponding
groups in the compressed MLM developed by Zhang et al in 2010 21. Cluster analysis is used to assign
similar individuals into groups. The elements of the kinship matrix are used as similarity measures in the
clustering analysis. Various linkage criteria (e.g., unweighted pair group method with arithmetic mean,
UPGMA) can be used to group the lines together. The number of groups is specified by the user. Once the
lines are assigned into groups, summary statistics of the kinship between and within groups are used as the
elements of a reduced kinship matrix. This procedure is used to create a reduced kinship matrix for each
compression level.
A series of mixed models are fitted to determine the optimal compression level. The value of the log
likelihood function is obtained for each model, and the optimal compression level is defined as the one
whose fitted mixed model yields the largest log likelihood function value. There are three parameters to
determine the range and interval of groups for examination: group.from, group.to and group.by. Their
defaults are 0, n and 10, where n is the total number of individuals.

3.8 General Linear Model (GLM)

Regular MLM22 is an extreme case of CMLM where each individual is considered as a group. It can be
simply performed by setting the number of groups equal to the total number of individuals, e.g. group.from
= n and group.to = n, where n is total number of individuals shared in both the genotype and phenotype
files. Similarly, general linear model (GLM) is another extreme case of CMLM where all individuals are
considered as one group. It can be simply performed by setting the number of groups equal to one, i.e.
group.from = 1 and group.to = 1. GLM is the working model in PLINK23, a primary software for studies in
human genetics.

3.9 P3D/EMMAx
In addition to implementing compression, GAPIT uses EMMAx/P3D6,24 to reduce computing time for
MLM, CMLM, ECMLM, and SUPER. If specified, the additive genetic (σ2a) and residual (σ2e) variance
components will be estimated prior to conducting GWAS. These estimates are then used for each SNP
where a mixed model is fitted.

3.10 SUPER
SUPER is an advanced version of FaST-Select, developed Wang et al. in 2016. The major difference
between SUPER and FaST-Select is that SUPER uses bin approach to select associated markers. The entire
genome is divided into equal sized bins and each bin is represented by the most significant marker on the
bin. The bin size and number of bins selected are optimized using maximum likelihood method in a random
model with the kinship derived from the selected bins. Consequently, the confounding between the kinship
and some of markers become more severe than the kinship derived from all markers. SUPER eliminate the
confounding by using the complementary kinship derived from associated markers except the ones that are
in strong linkage disequilibrium (LD) with the testing markers under a user defined threshold. Both
simulation and real data demonstrated that SUPER had higher statistical power than regular MLM.

To run SUPER in GAPIT, simply specify model= "SUPER".

17
 GAPIT User Manual

3.11 Multiple Locus Mixed Linear Model (MLMM)

GAPIT implemented a Multiple Loci Mixed Linear Model (MLMM) which use forward-backward stepwise
linear mixed-model regression to include associated markers as covariates.

To run MLMM in GAPIT, simply specify model= "MLMM".

3.12 FarmCPU
To solve the problem of false positive control and confounding between testing markers and cofactors
simultaneously, an iterative method, named Fixed and random model Circulating Probability Unification
(FarmCPU), was developed in 201611. The associated markers detected from the iterations are fitted as the
cofactors to control false positives for testing the rest markers in a fixed effect model. To avoid the over
model fitting problem in stepwise regression, a random effect model is used to select the associated markers
using maximum likelihood method11.

In the cycle of fixed effect model of iterations, markers are tested against the associated markers, not the
confounded kinship used by MLM, CMLM, ECMLM, SUPER, and MLMM. In the cycle of random effect
model of iterations, markers are selected among a small number of associated markers using maximum
likelihood method to avoid the over model fitting problem in stepwise regression used by MLMM, which
select marker among all available markers. Consequently, FarmCPU exhibits higher statistical power than
MLMM11. As FarmCPU tests markers in a fixed effect model, it is computational efficient than the methods
that test markers in random effect model, such as MLM, CMLM, ECMLM, SUPER, and MLMM 11.

To run FarmCPU in GAPIT, simply specify model= "FarmCPU".

3.13 BLINK
BLINK method was designed to have both high statistical power and computational efficiency 13. It was
inspired by FarmCPU method with two major changes to achieve the objectives. One is to eliminate the
assumption that causal genes are evenly distributed across genome that required by FarmCPU. As the
assumption cause either inclusion of non causal genes, or missing the causal genes that are in the same bin
with another causal genes with stronger signal. BLINK works directly on markers instead of bins. Markers
that are in linkage disequilibrium (LD) with the most significant marker are excluded. For the second
remaining marker, the exclusion is conducted in the same way as the most significant marker, so on and so
forth until no marker can be excluded.

The other change is to use Bayesian Information Content (BIC) of a fixed effect model to approximate the
maximum likelihood of a random effect model to select the associated markers among the markers remained
the exclusion based on LD. As both the models of testing markers and selecting associated markers as
cofactors are fixed effect model, the computation complexity reach the maximum. A dataset with one
million individuals and one million markers can be solved in hours by using BLINK C version. The BLINK
R version can be run as standard alone, or through GAPIT. To run BLINK in GAPIT, simply specify model=
"Blink". The performances of the two versions were documented by the BLINK article on GigaScience 13.

18
 GAPIT User Manual

4 Genomic Selection
Genomic selection, or genomic prediction termed in human genetics, is to use genetic markers across the
whole genome to predict individual performances if formats of phenotypes or predicted genetic merit. In
contrast to GWAS, there is a strong interaction between prediction methods and the traits measured in a
particular condition. The reversion of method superiorities has been found in many cases. The genomic
selection based on SUPER, named SUPER BLUP, has higher prediction accuracy than the genomic
selection based on MLM known as genomic BLUP (gBLUP) for traits controlled with a smaller number of
genes. The prediction accuracies are reversed for traits controlled by a large number of genes. The genomic
selection based on CMLM, named Compressed BLUP, has higher accuracy for traits with low heritability
than gBLUP. GAPIT implemented a series of methods for GWAS and genomic selection toward high
statistical power.

4.1 Genomic BLUP

Genomic prediction is performed with the method based on genomic best linear unbiased prediction,
(gBLUP)7. The method was extended to compressed best linear unbiased prediction (cBLUP) by using the
CMLM approach that was proposed for GWAS6. The genetic potential for a group, which is derived from
the BLUPs of group effects in the compressed mixed model, is used as a prediction for all individuals in
the group.

The groups created from compression belong to either a reference (R) or an inference (I) panel. All groups
in the reference panel have at least one individual with phenotypic data, and all groups in the inference
panel have no individuals with phenotypic data. Genomic prediction for groups in the inference panel is
based on phenotypic ties with corresponding groups in the reference panel.

The group kinship matrix is then partitioned into to R and I groups as follows:
k k RI 
k=  RR
k II 
(4)
 k IR ,

where kRR is the variance-covariance matrix for all groups in the reference panel, kRI is the covariance
matrix between the groups in the reference and inference panels, kIR = (kRI)’ is the covariance matrix
between the groups inference and reference panels, and kII is the variance-covariance matrix between the
groups in the inference panels.
Solving of mixed linear model is performed on the reference individuals.
yR =X R  + ZR u R + eR , (5)
where all terms are as defined in Equation (1), and the “R” subscript denotes that only individuals in the
reference panel are considered.

The genomic prediction of the inference groups is derived Henderson’s formula (1984) as follows:
u I =K IR K -1RR uR , (6)

where kIR, kRR , and uR are as previously defined, and uI is the predicted genomic values of the individuals
in the inference group.
The reliability of genomic prediction is calculated as follows:
PEV
Reliability=1-
 a2 19
 GAPIT User Manual

(7)

where PEV is the prediction error variance which is the diagonal element in the inverse left-hand side of
the mixed model equation, and  a is the genetic variance.
2

4.2 Compressed gBLUP

The compressed MLM substitute individuals with their corresponding groups that were clustered based on
the kinship among individuals. Research demonstrates that the compressed MLM had higher statistical
power for GWAS. Research also demonstrated that compressed MLM also had higher prediction accuracy
than the regular MLM, especially for traits with low heritability. As the regular MLM is an extreme case
of compressed MLM, the compressed MLM has higher, or at least equal prediction accuracy as the
regular MLM. When a compressed MLM is specified in GAPIT, individuals’ breeding values are
predicted by the breeding values of their corresponding groups.

4.3 SUPER gBLUP

The regular MLM uses the kinship derived from all the markers while SUPER uses the kinship derived
from the associated markers. As the associated markers are selected from all the markers using the
maximum likelihood method, the kinship used by SUPER has a better likelihood than the kinship used by
the regular MLM. Research demonstrated that the estimated breeding values from SUPER had higher
prediction accuracy than the estimated breeding values from the regular MLM.

20
 GAPIT User Manual

5 Output Results
GAPIT produces a series of output files that are saved in two formats. All tabular results are saved as comma
separated value (.csv) files, and all graphs are stored as printable document format (.pdf) files. This section
provides descriptions of these output files.

File name Description Type

Allelic_Effect_Estimates Estimate allelic effect with method CSV

Df.tValue.StdErr Estimate allelic t-value CSV
GWAS.Results SNP information and P-Value CSV
Log Log of whole model CSV

PRED Genomic Prediction CSV

ROC Table for power and FDR CSV
Kin.VanRaden kinship with VanRanden method CSV
PCA Principle components analysis CSV

PCA.eigenvalues Eigenvalues of PCA CSV

PCA.loadings Rotation of PCA CSV
Compression.multiple.group Compress likelihood, heritability and variance. PDF

MAF Minimum Allelic Frequency PDF

Manhattan.Plot.Chromosomewise Chromosome Manhattan PDF

Manhattan.Plot.Genomewise Genome Manhattan PDF

Optimum Heritability and Variance components PDF

phenotype_view Phenotype analysis PDF

QQ-Plot QQ plot PDF

ROC Power and FDR in ROC PDF

Heterozygosity Heterozygosity of genotype PDF
Kin.VanRaden Heat map of kinship PDF
Marker.Density Marker Density PDF
Marker.LD LD of first 1000 markers PDF

PCA.2D 2D PCA plot PDF

PCA.3D 3D PCA plot PDF
PCA.eigenValue Eigenvalue and variance of PCA PDF
NJtree.fan Fan type NJ tree PDF
NJtree.unrooted Unrooted NJ tree PDF
Manhattan.Mutiple.Plot Manhattan plot for multiple traits or methods PDF
Circular.Manhattan.Plot. Circular Manhattan plot PDF
Multraits.QQplot QQ plot for multiple trait or method PDF

Interactive.PCA Interactive PCA plot HTML

Interactive.Manhattan Interactive Manhattan plot HTML
Interactive.QQ Interactive QQ plot HTML

21
 GAPIT User Manual

5.1 Phenotype diagnosis

GAPIT diagnosis phenotype in several ways, including scatter plot, histogram, box plot and accumulative
distribution.

5.2 Marker density

Marker density is critical to establish Linkage Disequilibrium (LD) between markers and causal
mutations. Comparison between the marker density and the LD decade over distance provides the
indication if markers are dense enough to have good coverage of LD.

Figure 5.2 Frequency and accumulative frequency of marker density. Distribution of marker density is
displayed as a histogram and an accumulative distribution.
22
 GAPIT User Manual

5.3 Linkage Disequilibrium Decay

Linkage disequilibrium are measured as R square for pair wise markers and plotted against their distance.
The moving average of adjacent markers were calculated by using a sliding windows with ten markers.

Figure 5.3 Linkage disequilibrium (LD) decay over distance. LDs were calculated on sliding windows with
100 adjacent genetic markers. Each dot represents a pair of distances between two markers on the window
and their squared correlation coefficient. The red line is the moving average of the 10 adjacent markers.

5.4 Heterozygosis
The frequency of heterozygous were calculated for both individuals and markers. High level of
heterozygosis indicated low quality. For example, over 50% of heterozygosis on inbred lines for some of
markers suggested they problematic (see bottom right).

23
 GAPIT User Manual

5.5 Principal Component (PC) plot

For each PC included in the GWAS and GPS models, the observed PC values are plotted.

Figure 5.5 Pair-wise plots and 3D plots of principal component (PC).

5.6 Kinship plot

The kinship matrix used in GWAS and GPS is visualized through a heat map. To reduce computational
burden, this graph is not made when the sample size exceeds 1,000.

24
 GAPIT User Manual

Figure 5.6 Kinship plot. A heat map of the kinship matrix is created to indicate the relationship between
individuals.

5.7 Neighbor-Joining (NJ)-tree

Following the compress group, we classify individuals to explain population structure. We also can plot
group PCA plot with previous group.

Figure 5.7 Neighbor-Joining (NJ)-tree The whole population was divided into 5 clusters with each colors.

5.8 QQ-plot
The quantile-quantile (QQ) –plot is a useful tool for assessing how well the model used in GWAS accounts
for population structure and familial relatedness. In this plot, the negative logarithms of the P-values from
the models fitted in GWAS are plotted against their expected value under the null hypothesis of no
association with the trait. Because most of the SNPs tested are probably not associated with the trait, the
majority of the points in the QQ-plot should lie on the diagonal line. Deviations from this line suggest the
presence of spurious associations due to population structure and familial relatedness, and that the GWAS
model does not sufficiently account for these spurious associations. It is expected that the SNPs on the
upper right section of the graph deviate from the diagonal. These SNPs are most likely associated with the
trait under study. By default, the QQ-plots in GAPIT show only a subset of the larger P-values (i.e., less
significant P-values) to reduce the file size of the graph.

25
 GAPIT User Manual

Figure 5.8 Quantile-quantile (QQ) –plot of P-values. The Y-axis is the observed negative base 10 logarithm
of the P-values, and the X-axis is the expected observed negative base 10 logarithm of the P-values under
the assumption that the P-values follow a uniform[0,1] distribution. The dotted lines show the 95%
confidence interval for the QQ-plot under the null hypothesis of no association between the SNP and the
trait.

5.9 Manhattan Plot

The Manhattan plot is a scatter plot that summarizes GWAS results. The X-axis is the genomic position
of each SNP, and the Y-axis is the negative logarithm of the P-value obtained from the GWAS model
(specifically from the F-test for testing H0: No association between the SNP and trait). Large peaks in the
Manhattan plot (i.e., “skyscrapers”) suggest that the surrounding genomic region has a strong association
with the trait. GAPIT produces one Manhattan plot for the entire genome (Figure 3.4) and individual
Manhattan plots for each chromosome.

26
 GAPIT User Manual

Figure 5.9 Manhattan plot. The X-axis is the genomic position of the SNPs in the genome, and the Y-axis
is the negative log base 10 of the P-values. Each chromosome is colored differently. SNPs with stronger
associations with the trait will have a larger Y-coordinate value.

5.10 Association Table

The GWAS result table provides a detailed summary of appropriate GWAS results. The rows display the
results for each SNP above the user-specified minor allele frequency threshold. The SNPs sorted by their P
values (from smallest to largest).

Table 5.10 GWAS results for all SNPs that were analyzed.

This table provides the SNP id, chromosome, bp position, P-value, minor allele frequency (maf), sample
size (nobs), R2 of the model without the SNP, R2 of the model with the SNP, and adjusted P-value following
a false discovery rate (FDR)-controlling procedure25.

5.11 Allelic Effects Table

A separate table showing allelic effect estimates is also included in the suite of GAPIT output files. The
SNPs, presented in the rows, are sorted by their position in the genome.

27
 GAPIT User Manual

Table 5.11 Information of associated SNPs.

This table provides the SNP id, chromosome, bp position, and the allelic effect estimate of each SNP
analyzed.

5.12 Compression Profile

There are seven algorithms available to cluster individuals into groups for the compressed mixed linear
model. There are also four summary statistics available for calculating the group kinship matrix. When only
one group number (i.e., one dimension for the group kinship matrix) is specified, a column chart is created
to illustrate the compression profile for 2*log likelihood function (the smaller the better), genetic variance,
residual variance and the estimated heritability.

Figure 5.12.1. Compression profile with single group. The X-axis on all graphs display the summary
statistic method used to obtain the group kinship matrix. The rectangles with different colors indicate the
clustering algorithm used to group individuals.

Note: This graph is not created when multiple groups are specified.

When a range of groups (i.e., a range of dimensions for the group kinship matrix) is specified, a different
series of graphs are created. In this situation, the X-axis displays the group number. Lines with different
style and colors are used to present the combinations between clustering algorithm and the algorithm to
calculate kinship among groups.

28
 GAPIT User Manual

Figure 5.12.2. Compression profile over multiple groups. The X-axis on each graph is the number of
groups considered, and the Y-axes on the graphs are the -2*log likelihood function, the estimated genetic
variance components, the estimated residual variance component, the estimated total variance, and the
heritability estimate. Each clustering method and group kinship type is represented as a line on each graph.

Notice: This graph is not created when only one group is specified.

5.13 The Optimum Compression

Once the optimal compression settings are determined, GAPIT produces a PDF file containing relevant
detailed information. This information includes the optimal algorithm to calculate the group kinship matrix,
the optimal clustering algorithm, the optimal number of groups, -2*log likelihood function and the
estimated heritability.

Figure 5.13 The profile for the optimum compression. The optimal method to calculate group kinship is
“Mean”, the optimal clustering method is “average”, the number of groups (ie., the dimension of the
group kinship matrix) is 251, the value of -2*log likelihood function is 1800.11, and the heritability is
41.6%.

29
 GAPIT User Manual

5.14 Model Selection Results

By selecting “Model.selection = TRUE”, forward model selection using the Bayesian information
criterion (BIC) will be conducted to determine the optimal number of PCs/Covariates to include for each
phenotype. The results summary (below) for model selection are stored in a .csv file called
“.BIC.Model.Selection.Results”.

Table 5.14.1 Summary for Bayesian information criterion (BIC) model selection results.

The number of PCs/Covariates, the BIC value, and the log Likelihood function value are presented. In this
table, the optimal number of PCs to include in the GWAS model is 2.

5.15 Multiple traits, environments, or models

There are several new method integrated in. Furthermore, more than single method or trait result, we
propose a multiple method or traits Manhattan and QQ plots for comparison with methods or traits, this
part is based on MVP library to exhibit. There are two types of Manhattan plots (Orthogon and
Roundness).

Figure 5.15.1 Multiple traits or methods Manhattan and QQ plot. The dash line in the Figure A indicated
the common significant markers were detected by two methods or traits. The solid line in the Figure A
indicated the common significant markers were detected by more than two methods or traits.

30
 GAPIT User Manual

Another multiple Manhattan plot with multiple symphysis could be created in the GAPIT now. The
square, triangle, circle, diamond, and inverted triangle indicated each GWAS results of multiple methods
or traits. Users also can define the type of points by “allpch” parameter.

Figure 5.15.2 Multiple traits or methods symphysis Manhattan plot. The dash line indicated the common
significant markers were detected by two methods or traits. The solid line indicated the common
significant markers were detected by more than two methods or traits.

5.16 Genomic Prediction

The genomic prediction results are saved in a .csv file.
Table 4.16.1 Genomic Breeding values and prediction error variance.

The individual id (taxa), group, RefInf which indicates whether the individual is in the reference group (1)
or not (2), the group ID number, the BLUP, and the PEV of the BLUP.

31
 GAPIT User Manual

Figure 5.16 Interactive GS plot for gBLUP, cBLUP and sBLUP.

5.17 Distribution of BLUPs and their PEV.

A graph is provided to show the joint distribution of GBV and PEV. The correlation between them is an
indicator of selection among the sampled individuals 26.

Figure 5.17 Joint distribution of genomic breeding value and prediction error variance.

32
 GAPIT User Manual

5.18 Interactive GWAS plot

A graph is provided to show the Interactive Manhattan and QQ plot with “Inter.Plot=T”. These files are
HTML and supported to act with mouse. The more details of SNP will be showed when the mouse moved
on the point.

Figure 5.18 Interactive Manhattan and QQ plot.

33
 GAPIT User Manual

6 Tutorials
The “Getting started” section should be reviewed before running these tutorials, and it is assumed that the
GAPIT package and its required libraries have been installed. These tutorials begin with a scenario requiring
minimal user input. Subsequent scenarios require a greater amount of user input. Each scenario involves
two steps: reading in the data and then running the GAPIT() function. All tutorials are available on the
GAPIT home page, which also contains the R source code and results for all the scenarios.

The GAPIT maize demonstration data (described at www.panzea.org) are from a maize association panel
consisting of 281 diverse lines27. The genotypic data consist of 3,093 SNPs distributed across the maize
genome, and are available in HapMap and numeric format. The three phenotypes included are ear height,
days to pollination, and ear diameter. The kinship matrix was calculated using the method of Loiselle et
al.28 and the fixed effects used to account for population structure were obtained from STRUCTURE 29.

Notice: It is important that the correct paths to the directories are specified. Please note that two backward
slashes (“\\”) are necessary when specifying these paths.
6.1 A Basic Scenario
The user needs to provide two data sets (phenotype and genotype) and one input parameter. This parameter,
“PCA.total”, specifies the number of principal components (PCs) to include in the GWAS model. GAPIT
will automatically calculate the kinship matrix using the VanRaden method30, perform GWAS and genomic
prediction with the optimum compression level using the default clustering algorithm (average) and group
kinship type (Mean). The scenario assumes that the genotype data are saved in a single file in HapMap
format. If the working directory contains the tutorial data, the analysis can be performed by typing these
command lines:

#Step 1: Set data directory and import files

myY <- read.table("mdp_traits.txt", head = TRUE)
myG <- read.table("mdp_genotype_test.hmp.txt", head = FALSE)

#Step 2: Run GAPIT

myGAPIT <- GAPIT(
Y=myY,
G=myG,
PCA.total=3
)

6.2 Enhanced Compression

In this scenario, the user can specify additional clustering algorithms (controlled by the “kinship.cluster”
parameter) and kinship summary statistic (controlled by the “kinship.group” parameter). The default is
kinship.cluster="average", and kinship.group="Mean". Their expansion, the Enriched CMLM 8 improves
statistical power. Additionally, a specific range group numbers (i.e., dimension of the kinship matrix) can
be specified. This range is controlled by the “group.from”, “group.to”, and “group.by” parameters. The
analysis can be performed by typing these command lines:

34
 GAPIT User Manual

#Step 1: Set data directory and import files

myY <- read.table("mdp_traits.txt", head = TRUE)
myG <- read.table("mdp_genotype_test.hmp.txt", head = FALSE)

#Step 2: Run GAPIT

myGAPIT <- GAPIT(
Y=myY,
G=myG,
PCA.total=3,
kinship.cluster=c("average", "complete", "ward"),
kinship.group=c("Mean", "Max"),
model="CMLM"
)

6.3 User-inputted Kinship Matrix and Covariates

This scenario assumes that the user provides a kinship matrix and covariate file. The kinship matrix or
covariates (e.g., PCs) may be calculated previously or from third party software. When the PCs are input in
this way, the parameter “PCA.total” should be set to 0 (default). Otherwise, PCs will be calculated within
GAPIT, resulting in a singular design matrix in all model fitted for GWAS. The analysis can be performed
by typing these command lines:

#Step 1: Set data directory and import files

myY <- read.table("mdp_traits.txt", head = TRUE)
myG <- read.table("mdp_genotype_test.hmp.txt", head = FALSE)
myKI <- read.table("KSN.txt", head = FALSE)
myCV <- read.table("mdp_PC", head = TRUE)

#Step 2: Run GAPIT

myGAPIT <- GAPIT(
Y=myY,
G=myG,
KI=myKI,
CV=myCV
)

6.4 Multiple Genotype Files

In this scenario, the HapMap genotypic data set from Scenario 1 is subdivided into multiple genotype
files , one for each chromosome. This scenario mimics the situation where the genotype file is too large to
be handled in R. When this situation arises, all genotype files need to have a common name and
extensions, as well as a sequential number (e.g., “mdp_genotype_chr1.hmp.txt”,
“mdp_genotype_chr2.hmp.txt”, …). The starting and ending file are indicated by the “file.from” and
“file.to” parameters. The common file name (e.g., “mdp_genotype_chr”) and file name extension (e.g.,
“hmp.txt”) are passed to GAPIT through the “file.G”, “file.Ext.G” parameters, respectively. When
“file.path” is not provided, GAPIT try to get the data from the current working directory. The analysis can
be performed by typing these command lines:

35
 GAPIT User Manual

#Step 1: Set data directory and import files

myY <- read.table("mdp_traits.txt", head = TRUE)

#Step 2: Run GAPIT

myGAPIT <- GAPIT(
Y=myY,
PCA.total=3,
file.G="mdp_genotype_chr",
file.Ext.G="hmp.txt",
file.from=1,
file.to=10,
file.path="C:\\myGAPIT\\"
)

The three genotype file used in these scenario are from the file used in Tutorial 5.1. Their results should
be identical.
6.5 Numeric Genotype Format
In this scenario, the genotype data set from Scenario 1 is formatted differently, specifically in numerical
format. Two genotype files are required. One file contains the genotypic data, and the other contains the
chromosome and base pair position of each SNP. These are passed to GAPIT through the “GD” and “GM”
parameters, respectively. The analysis can be performed by typing these command lines:

#Step 1: Set data directory and import files

myY <- read.table("mdp_traits.txt", head = TRUE)
myGD <- read.table("mdp_numeric.txt", head = TRUE)
myGM <- read.table("mdp_SNP_information.txt" , head = TRUE)

#Step 2: Run GAPIT

myGAPIT <- GAPIT(
Y=myY,
GD=myGD,
GM=myGM,
PCA.total=3
)

6.6 Numeric Genotype Format in Multiple Files

In this scenario, the numeric genotype data set from Scenario 6 is subdivided into multiple genotype files.
The common name and extension of genotype data file are passed to GAPIT through “file.GD” and
“file.Ext.GD” parameters, respectively. Similarly, the common name and extension of genotype map file
are passed to GAPIT through the “file.GM” and “file.Ext.GM” parameters, respectively. The analysis can
be performed by typing these command lines:

#Step 1: Set data directory and import files

myY <- read.table("mdp_traits.txt", head = TRUE)

#Step 2: Run GAPIT

myGAPIT <- GAPIT(
Y=myY,
PCA.total=3,
file.GD="mdp_numeric",
file.GM="mdp_SNP_information",
file.Ext.GD="txt",
file.Ext.GM="txt",
file.from=1,
file.to=3,
)

36
 GAPIT User Manual

The three genotype file used in these scenario are the splits from the file used in the previous scenario.
Their results should be identical.

6.7 Fractional SNPs for Kinship and PCs

The computations of kinship and PCs are extensive with large number of SNPs. Sampling a fraction of it
would reduce computing time. More importantly, it would give very similar result with appropriate number
of SNPs sampled. The fraction can be controlled by “Ratio” parameter in GAPIT. The sampling scheme is
random. A line of “SNP.fraction=0.6” is added to the previous scenario which has 3,093 SNPs:

#Step 1: Set data directory and import files

myY <- read.table("mdp_traits.txt", head = TRUE)

#Step 2: Run GAPIT

myGAPIT <- GAPIT(
Y=myY,
PCA.total=3,
file.GD="mdp_numeric",
file.GM="mdp_SNP_information",
file.Ext.GD="txt",
file.Ext.GM="txt",
file.from=1,
file.to=3,
SNP.fraction=0.6
)

6.8 Memory saving

With large amount of individuals, loading a entire large genotype dataset could be difficult. GAPIT load a
fragment of it each time. The default of the fragment size is 512 SNPs. This number can be changed with
“file.fragment” parameter in GAPIT. Here is an example of using “file.fragment =128”.

#Step 1: Set data directory and import files

myY <- read.table("mdp_traits.txt", head = TRUE)

#Step 2: Run GAPIT

myGAPIT <- GAPIT(
Y=myY,
PCA.total=3,
file.GD="mdp_numeric",
file.GM="mdp_SNP_information",
file.Ext.GD="txt",
file.Ext.GM="txt",
file.from=1,
file.to=3,
SNP.fraction=0.6,
file.fragment = 128
)

This scenario is the same as previous scenario except changing “file.fragment” from default (512) to 128.
As SNPs (minimume of two) are sampled withing each fragment, the final SNPs sampled would be
different for different length of fragment when the SNP sample fraction is less than 100%. The results in
this scenario would be different from the previous one.

37
 GAPIT User Manual

6.9 Model selection

The degree of correlation with population structure varies from trait to trait. Therefore, the full set of PCs
selected to account for population structure in the GWAS model are not necessary for all traits. As such,
GAPIT has the capability to conduct Bayesian information criterion (BIC)-based model selection to find
the optimal number of PCs for inclusion in the GWAS models. Model selection is activated by selecting
“Model.selection = TRUE”. The results for the BIC model selection procedure are summarized in the
“.BIC.Model.Selection.Results.csv” output file.
myY <- read.table("mdp_traits.txt", head = TRUE)
myG <- read.table("mdp_genotype_test.hmp.txt", head = FALSE)

#Step 2: Run GAPIT

myGAPIT <- GAPIT(
Y=myY,
G=myG,
PCA.total=3,
Model.selection = TRUE
)

6.10 SUPER
GAPIT also implements the SUPER GWAS method 9, which extracts a small subset of SNPs and uses
them in FaST-LMM. This method not only retains the computational advantage of FaST-LMM, but also
increases statistical power.

#Step 1: Set data directory and import files

myY <- read.table("mdp_traits.txt", head = TRUE)
myG <- read.table("mdp_genotype_test.hmp.txt" , head = FALSE)

#Step 2: Run GAPIT

myGAPIT_SUPER <- GAPIT(
Y=myY[,c(1,2)],
G=myG,
PCA.total=3,
model=”SUPER”
)

6.11 MLMM
Multiple Loci Mixied linear Model is published by Segura in 2012. The code of MLMM in GAPIT is:
myY <- read.table("mdp_traits.txt", head = TRUE)
myG <- read.table("mdp_genotype_test.hmp.txt", head = FALSE)

#Step 2: Run GAPIT

myGAPIT <- GAPIT(
Y=myY,
G=myG,
PCA.total=3,
model="MLMM"
)

38
 GAPIT User Manual

6.12 Farm-CPU
Fixed and random model Circulating Probability Unification (FarmCPU) is published by Xiaolei in 2016.
The code of Farm-CPU in GAPIT is:

myY <- read.table("mdp_traits.txt", head = TRUE)

myG <- read.table("mdp_genotype_test.hmp.txt", head = FALSE)

#Step 2: Run GAPIT

myGAPIT <- GAPIT(
Y=myY,
G=myG,
PCA.total=3,
model="FarmCPU"
)

6.13 BLINK
Bayesian-information and Linkage-disequilibrium Iteratively Nested Keyway (BLINK) is published by
Meng in 2018. The code of BLINK used in GAPIT is:

myY <- read.table("mdp_traits.txt", head = TRUE)

myG <- read.table("mdp_genotype_test.hmp.txt", head = FALSE)

#Step 2: Run GAPIT

myGAPIT <- GAPIT(
Y=myY,
G=myG,
PCA.total=3,
model="Blink"
)

The BLINK C version execute file should be in the working directory and change mod 777.
(chmod 777 blink_versions)
The execute file can be downloaded at https://github.com/Menggg/BLINK/blob/master/

6.14 Multiple model

The GAPIT provide an approach for comparison of multiple methods in GWAS. All GWAS methods in
the GAPIT can be used in here:

myY <- read.table("mdp_traits.txt", head = TRUE)

myGD=read.table("http://zzlab.net/GAPIT/data/mdp_numeric.txt",head=T)
myGM=read.table("http://zzlab.net/GAPIT/data/mdp_SNP_information.txt",head=T)

#Step 2: Run GAPIT

myGAPIT <- GAPIT(
Y=myY,
GD=myGD,
GM=myGM,
PCA.total=3,
Multiple_analysis=TRUE,
model=c("GLM","MLM","MLMM","FarmCPU","Blink")
)

39
 GAPIT User Manual

6.15 gBLUP
The gBLUP is based on the Mixed linear Model to predict phenotype, BLUP, and BLUE value.

myY <- read.table("mdp_traits.txt", head = TRUE)

myGD=read.table("http://zzlab.net/GAPIT/data/mdp_numeric.txt",head=T)
myGM=read.table("http://zzlab.net/GAPIT/data/mdp_SNP_information.txt",head=T)

#Step 2: Run GAPIT

myGAPIT <- GAPIT(
Y=myY[,c(1,2)],
GD=myGD,
GM=myGM,
PCA.total=3,
model=c("gBLUP")
)

6.16 cBLUP
The cBLUP is based on the compressed Mixed linear Model. It used optimum compression group kinship
instead of individual kinship.
myY <- read.table("mdp_traits.txt", head = TRUE)
myGD=read.table("http://zzlab.net/GAPIT/data/mdp_numeric.txt",head=T)
myGM=read.table("http://zzlab.net/GAPIT/data/mdp_SNP_information.txt",head=T)

#Step 2: Run GAPIT

myGAPIT <- GAPIT(
Y=myY[,c(1,2)],
GD=myGD,
GM=myGM,
PCA.total=3,
model=c("cBLUP")
)

6.17 sBLUP
The sBLUP is based on the SUPER Model. It used pseudo QTNs to build individual kinship.

myY <- read.table("mdp_traits.txt", head = TRUE)

myGD=read.table("http://zzlab.net/GAPIT/data/mdp_numeric.txt",head=T)
myGM=read.table("http://zzlab.net/GAPIT/data/mdp_SNP_information.txt",head=T)

#Step 2: Run GAPIT

myGAPIT <- GAPIT(
Y=myY[,c(1,2)],
GD=myGD,
GM=myGM,
PCA.total=3,
model=c("sBLUP")
)

40
 GAPIT User Manual

7 Prototype
The usage of GAPIT described in previous chapters barely require knowledge of R. Users can simply
copy/paste the command lines from the user manual with a minimal keyboard typing such as changing
file names and path. All the results are saved in the format of text files and PDF files. GAPIT also output
R objects which can be used for advance purposes, including: 1) developing new statistical approaches or
software package using GAPIT output as a starting point; 2) comparing GAPIT with other new or existing
statistical methods or software packages; 3) studying a specific result from GAPIT. Using these objects
require knowledge of R. This chapter give examples to use GAPIT output R objects.
7.1 Statistical power comparison among methods
GAPIT.Power.compare() is an example function use multiple functions in GAPIT and the output R
objects to perform comparison of statistical power using different models. The deteils can be found in
GAPIT source code. The following is an example to use the function. Note: running 100 replicates may
take more than a day to finsh.

myGD <-read.table("mdp_numeric.txt", head = TRUE)

myGM <-read.table("mdp_SNP_information.txt", head = TRUE)
GAPIT.Power.compare(
myGD=myGD,
myGM=myGM,
nrep=100,
h2=0.9,
all.method=c("GLM","MLM","MLMM","FarmCPU","BLINK"),
NQTN=5)

41
 GAPIT User Manual

7.2 Genomic selection

GAPIT.Prediction() is another example of using GAPIT subfunctions and R objects outputs for GS

myY<-read.table("mdp_traits.txt", head = TRUE)

myGD <-read.table("mdp_numeric.txt", head = TRUE
myGM <-read.table("mdp_SNP_information.txt", head = TRUE)
set.seed(99163)
GAPIT.Validation(
Y=myY[,1:2],
model=c("gBLUP"),
GD=myGD,
GM=myGM,
PCA.total=3,
file.output=T,
nfold=5
)

7.3 Cross validation with replacement

Here we demonstrate an example of studying a specific output of GAPIT, specifically to investigate the
accuracy of genome prediction through cross validation.

First, we randomly set 25% of original phenotype (Y) as missing (NA) and generate a genomic prediction
model by using their kinship. Then we record the correlation between the predicted phenotypic values and
the original phenotype. We repeat this process for 1000 times. The average correlation is used as the
criteria of genome prediction accuracy. The corresponding R code is displayed in the following box. The
accuracy (correlation coefficient) over the 100 replicates (for demonstration purposes only; we suggest

42
 GAPIT User Manual

using 1000 replications) were 0.9203 and 0.6749 in the reference and inference (cross validation),
respectively. The standard deviations were 0.078 and 0.0054 in the reference and inference, respectively.

R code for cross validation with replacement

#Import files
#######################################################################################
myY <- read.table("mdp_traits.txt", head = TRUE)
myKI <- read.table("KSN.txt", head = FALSE)
myCV <- read.table("mdp_PC", head = TRUE)

#Initial
#######################################################################################
t=100 #total replicates
s=1/5 #sample of inference, e.g. set it to 1/5 for five fold cross validation
Y.raw=myY[,c(1,3)]#choos a trait
Y.raw=Y.raw[!is.na(Y.raw[,2]),] #Remove missing data
n=nrow(Y.raw)
n.missing=round(n*s)
storage.ref=matrix(NA,t,1)
storage.inf=matrix(NA,t,1)

#Loop on replicates
for(rep in 1:t){

#Set missing data

sample.missing=sample(1:n,n.missing)
if(n.missing>0){ Y0=Y.raw[-sample.missing,]
}else{Y0=Y.raw}

#Prediction
myGAPIT <- GAPIT(
Y=Y0,
KI=myKI,
CV=myCV,
model="gBLUP"
)
prediction=myGAPIT$Pred

#Separate reference (with phenotype) and inference (without phenotype)

prediction.ref=prediction[prediction[,3]==1,]
prediction.inf=prediction[prediction[,3]==2,]

#Merge prediction with original Y

YP.ref <- merge(Y.raw, prediction.ref, by.x = "Taxa", by.y = "Taxa")
YP.inf <- merge(Y.raw, prediction.inf, by.x = "Taxa", by.y = "Taxa")

#Calculate correlation and store them

r.ref=cor(as.numeric(as.vector(YP.ref[,2])),as.numeric(as.vector(YP.ref[,6]) ))
r.inf=cor(as.numeric(as.vector(YP.inf[,2])),as.numeric(as.vector(YP.inf[,6]) ))
storage.ref[rep,1]=r.ref
storage.inf[rep,1]=r.inf
}#End of for (rep in 1:t)

storage=cbind(storage.ref,storage.inf)
colnames(storage)=c("Reference","Inference")
write.table(storage, "GAPIT.Cross.Validation.txt", quote = FALSE, sep = "\t", row.names = TRUE,col.names = NA)

7.4 Cross validation without replacement

Cross validation can also be performed by excluding one or a set of individuals in the reference to derive
predicted phenotypic values from the genomic prediction model. Using this approach, this process can be
43
 GAPIT User Manual

repeated until all the individuals have been excluded at least once. The correlation between the originals
and the prediction (might be more than once) is used as the accuracy of prediction. The following
demonstrate the process with the same data in previous section.

#Initial
#######################################################################################
nj= 200 # number of Jack Knifes, nj>0, nj!=1

Y.raw=myY[,c(1,3)]#choos a trait
Y.raw=Y.raw[!is.na(Y.raw[,2]),] #Remove missing data
n=nrow(Y.raw)

if(nj>=1){nLoop=nj
}else{
nLoop=1/nj
}
assignment=ceiling((1:n)/(n/nLoop))
randomization=sample(1:n,n)
assignment=assignment[randomization]
nLoop=ceiling(nLoop)

#Loop on replicates
for(rep in 1:nLoop){

#Set missing data

if(nj>=1){Y0=Y.raw[assignment!=rep,]
}else{
Y0=Y.raw[assignment==rep,]
}
#Prediction
myGAPIT <- GAPIT(
Y=Y0,
KI=myKI,
CV=myCV,
model=" gBLUP "
)
prediction=myGAPIT$Pred

#Separate reference (with phenotype) and inference (without phenotype)

if(rep==1){
prediction.inf=prediction[prediction[,3]==2,]
}else{
prediction.inf=rbind(prediction.inf,prediction[prediction[,3]==2,] )
}
}#End of for (rep in 1:t)

#Merge prediction with original Y

YP.inf <- merge(Y.raw, prediction.inf, by.x = "Taxa", by.y = "Taxa")

#Calculate correlation and store them

r.inf=cor(as.numeric(as.vector(YP.inf[,2])),as.numeric(as.vector(YP.inf[,6]) ))
write.table(YP.inf, "GAPIT.Jack.Knife.txt", quote = FALSE, sep = "\t", row.names = TRUE,col.names = NA)
print(r.inf)

7.5 Convert HapMap format to numerical

Many software require genotype data in the numerical format. GAPIT can perform such conversion with a
few lines of code as follows.
myG <- read.table("mdp_genotype_test.hmp.txt", head = FALSE)
myGAPIT <- GAPIT(G=myG, output.numerical=TRUE)
myGD= myGAPIT$GD
myGM= myGAPIT$GM

44
 GAPIT User Manual

8 Appendix

8.1 GAPIT Biography

Date Version Event
First public release with following method implemented:
• Principal component method to encounter population structure (Price).
• Unified mixed model to encounter both population structure and kinship.
11-May-11 1.2
• EMMA method to improve the speed to estimate variance components (ratio).
• Compressed mixed model to improve statistical power and speed
• P3D or EMMAx to improve speed by estimating population parameters (e.g.
variances and grouping) only once.
13-Jun-11 1.22 Genotype in numerical format in addition to hapmap format
2-Sep-11 1.31 Reading fragment within single genotype file to save memory
17-Sep-11 1.36 Interface change for prototyping
24-Oct-11 1.41 Option to impute missing genotypes as middle, major, minor, present/absent.
1-Nov-11 1.42 Matrix partitioning to improve speed (5-10 fold faster)
7-Dec-11 2.01 FaST-LMM method implemented
19-May-12 2.19 Splitting big genotype file into small ones
8-Jul-12 NA GAPIT Bioinformatics paper accepted for publication
8-Nov-12 2.2 Automatic sorting taxa in kinship for regular MLM
14-Feb-13 2.25 Confidence interval on QQ plot
26-Apr-13 2.26 Labeling QTNs on Manhattan plot
15-Jul-14 2.27 SUPER implemented
20-Aug-14 2.28 ECMLM implemented
25-Oct-14 2.29 VanRaden kinship algorithm with centralization
28-Feb-15 2.3.41 Enrichment on output figures and tables
5-Sep-15 by date Uniform output across models (Version 3 initiation)
12-Oct-15 Simulation of category phenotype
22-Oct-15 Compression BLUP (cBLUP) for genomic prediction
1-Apr-16 GAPIT version 2 paper published by Plant Genome
4-Apr-16 Pedigree-like marker based kinship
4-Apr-16 SUPER BLUP (sBLUP) for genomic prediction
31-Mar-17 Indicating LD to the strongest associated marker above threshold
2-May-18 Add cBLUP and sBLUP demo script into user manual
Jul 17,2019 Add command for loading GAPIT from GitHub
Sep 4, 2021 GAPIT version 3 paper published on Genomics, Proteomics & Bioinformatics

45
 GAPIT User Manual

8.2 Frequently Asked Questions

1. How to cite GAPIT?
A: Citation may vary based on the usage of versions (the first version 1 ,the second version2, or the third
version3), and methods involved, such as the regular MLM20 methods, CMLM6, ECMLM8, SUPER9,
P3D6, FarmCPU31, and BLINK13. Here is an example: “The GWAS was conducted using the BLINK
model17 implemented in GAPIT R Software package (version 3)3”.

2. What do I do if I get frustrated?

A: Try to go through this Q/A list and GAPIT Forum first before asking help from GAPIT team. If you
need to contact GAPIT team, email to Dr. Xiaolei Liu (email: [email protected]) on questions
related to FarmCPU, Dr. Meng Huang (email: [email protected]) for questions related to
BLINK, or Dr. Jiabo Wang (email: [email protected]) on rest questions. In all cases (Forum
or Emails), please state your names and your institutions.

3. Why GAPIT has different results from other software?

A: The most common reasons to have different results is that these software packages use different genetic
models (e.g. additive vs. additive + dominant), statistical models (e.g. GLM, MLM, CMLM, and
ECMLM), and processing of missing data. The GAPIT Bioinformatics paper demonstrated that GAPIT
and TASSEL gave identical results for inbred (additive only) without missing values for using MLM.

4. There are many methods implemented in GAPIT, which one should I use?
A: Literature demonstrated the order of statistical power: BLINK > FarmCPU> MLMM > SUPER >
ECMLM > CMLM > MLM > GLM.

5. How many PCs to include?

A: There no clear answer for this question. However, here are the two ways most of people do. 1) The
number of principal components (PCs) included in the GWAS models can be adjusted in GAPIT. To
help determine the number of PCs that adequately explain population structure, a screen plot is provided
in the GAPIT output (if at least one PC is selected for inclusion into the final model). Once the ideal
number of PCs is determined, GAPIT should be reran with this number PCs included in the GWAS
models; 2) Use BIC-based model selection (activated by writing Model.selection = TRUE in the
GAPIT() function) to determine the “optimal” number of PCs. The optimal is in quotations because no
evidence has been found for optimum statistical power.

6. Is it feasible I compare different models on my data?

A: Yes, you can compare different models implemented in GAPIT or other software packages through
simulation. All you need is a genotype file. The demo source code is amiable at the Workshop of
Assessment of Statistical Power in GWAS (http://zzlab.net/WorkshopISU).

7. How do I report an error?

A: In order to fix the problem, please copy and paste the error message from the R environment and attach
your R source code and the dataset that allow us to repeat the error.

8. What should I do with “Error in file (file, "rt") : cannot open the connection”?
A: In most cases this error is caused by incorrect file name or number of file specific is more than exist.

46
 GAPIT User Manual

9. What should I do with “Error in GAPIT (... : unused argument(s) ...”?

A: In most cases this error is caused by incorrect spelling of GAPIT key word such as upper or lower case,
or missing a comma.

10. How deal with “Error in solve.default(crossprod(X, X)) : system is computationally

singular”?
A: Check covariate variables and remove the ones that are linear dependent with others.

11. How to fix the error of using covariates from STRUCTURE as fixed effects?
A: This error is occurring because the covariates from STRUCTURE are linearly dependent. In particular,
for a given individual/taxa, these covariates sum to 1. To circumvent this error, remove one of the
STRUCTURE covariates from the corresponding input file.

12. Should I remove SNPs with MAF below 5%?

A: The answers are Yes/No. Rare SNPs with low Minor Allele Frequency (MAF) usually cause false
positives, especially for small samples and traits that do not have normal distribution. However, many
causal genetic variants are rare. A recommended practice is to not remove them, but interpret them with
caution.

13. My trait was measured in multiple environments, how do I use them simultaneously?
A: They can be averaged across environments, and use means by GAPIT. The genetic and environmental
interaction was implemented in a separated software package: GEMT (http://zzlab.net/GEMT).

14. Is it OK to analyze binary traits (case-control) with GAPIT?

A: Yes, there are many applications.

15. Does normality transformation help?

A: Yes, non-normality, rare variants and small samples jointly cause false positives. The transformation
helps in case of small samples and SNPs with low MAF.

16. Should I use PCs or Q matrix?

A: Keyan Zhao and et al (PLoS Genetics, 2007) compared the two methods and demonstrated that they had
similar statistical power.

47
 GAPIT User Manual

REFERENCES
1. Lipka, A. E. et al. GAPIT: genome association and prediction integrated tool. Bioinformatics 28, 2397–2399 (2012).
2. Tang, Y. et al. GAPIT Version 2: An Enhanced Integrated Tool for Genomic Association and Prediction. Plant
Genome 9, 1–9 (2016).
3. Wang, J. & Zhang, Z. GAPIT Version 3: Boosting Power and Accuracy for Genomic Association and Prediction.
Genomics. Proteomics Bioinformatics 19, 1–12 (2021).
4. Price, A. L. et al. Principal components analysis corrects for stratification in genome-wide association studies. Nat
Genet 38, 904–909 (2006).
5. Yu, J. M. et al. A unified mixed-model method for association mapping that accounts for multiple levels of relatedness.
Nat. Genet. 38, 203–208 (2006).
6. Zhang, Z. et al. Mixed linear model approach adapted for genome-wide association studies. Nat. Genet. 42, 355–360
(2010).
7. Zhang, Z., Todhunter, R. J., Buckler, E. S. & Van Vleck, L. D. Technical note: Use of marker-based relationships with
multiple-trait derivative-free restricted maximal likelihood. J. Anim. Sci. 85, 881–885 (2007).
8. Li, M. et al. Enrichment of statistical power for genome-wide association studies. BMC Biol. 12, 73 (2014).
9. Wang, Q., Tian, F., Pan, Y., Buckler, E. S. E. S. & Zhang, Z. A SUPER Powerful Method for Genome Wide
Association Study. PLoS One 9, e107684 (2014).
10. Segura, V. et al. An efficient multi-locus mixed-model approach for genome-wide association studies in structured
populations. Nature Genetics 44, 825–830 (2012).
11. Liu, X., Huang, M., Fan, B., Buckler, E. S. & Zhang, Z. Iterative Usage of Fixed and Random Effect Models for
Powerful and Efficient Genome-Wide Association Studies. PLoS Genet. 12, e1005767 (2016).
12. Wang, J. et al. Expanding the BLUP alphabet for genomic prediction adaptable to the genetic architectures of complex
traits. Heredity (Edinb). 121, 648–662 (2018).
13. Huang, M., Liu, X., Zhou, Y., Summers, R. M. & Zhang, Z. BLINK: A package for the next level of genome-wide
association studies with both individuals and markers in the millions. Gigascience giy154 (2019).
doi:10.1093/gigascience/giy154
14. Atwell, S. et al. Genome-wide association study of 107 phenotypes in Arabidopsis thaliana inbred lines. Nature 465,
627–631 (2010).
15. Zhang, Z. et al. Mixed linear model approach adapted for genome-wide association studies. Nat Genet 42, 355–360
(2010).
16. Purcell, S. et al. PLINK: A Tool Set for Whole-Genome Association and Population-Based Linkage Analyses. Am J
Hum Genet 81, 559-575. (2007).
17. Huang, M. et al. BLINK: A package for the next level of genome-wide association studies with both individuals and
markers in the millions. Gigascience giy154 (2018). doi:10.1093/gigascience/giy154
18. Pritchard, J. K., Stephens, M. & Donnelly, P. Inference of population structure using multilocus genotype data.
Genetics 155, 945–959 (2000).
19. Zhao, K. et al. An Arabidopsis example of association mapping in structured samples. PLoS Genet. 3, 0071–0082
(2007).
20. Yu, J. M. et al. A unified mixed-model method for association mapping that accounts for multiple levels of relatedness.
Nat. Genet. 38, 203–208 (2006).
21. Zhang, Z. et al. Mixed linear model approach adapted for genome-wide association studies. Nat. Genet. 42, 355–60
(2010).
22. Yu, J. et al. A unified mixed-model method for association mapping that accounts for multiple levels of relatedness.
Nat. Genet. 38, 203–208 (2006).
23. Purcell, S. et al. PLINK: a tool set for whole-genome association and population-based linkage analyses. Am J Hum
Genet 81, 559–575 (2007).
24. Kang, H. M. et al. Variance component model to account for sample structure in genome-wide association studies. Nat.
Genet. 42, 348–354 (2010).
48
 GAPIT User Manual

25. Benjamini, Y. & Hochberg, Y. Controlling the false discovery rate: a practical and powerful approach to mutliple
testing. J.R. Statis. Soc. B. 57, 289–300 (1995).
26. Guo, G. et al. Canine hip dysplasia is predictable by genotyping. Osteoarthr. Cartil. 19, 420–429 (2011).
27. Flint-Garcia, S. A. et al. Maize association population: a high-resolution platform for quantitative trait locus dissection.
Plant J 44, 1054–1064 (2005).
28. Loiselle, B. A., Sork, V. L., Nason, J. & Graham, C. Spatial Genetic-Structure of a Tropical Understory Shrub,
Psychotria Officinalis (Rubiaceae). Am. J. Bot. 82, 1420–1425 (1995).
29. FALUSH, D., STEPHENS, M. & PRITCHARD, J. K. Inference of population structure using multilocus genotype
data: dominant markers and null alleles. Mol. Ecol. Notes 7, 574–578 (2007).
30. VanRaden, P. M. Efficient methods to compute genomic predictions. J Dairy Sci 91, 4414–4423 (2008).
31. Liu, X., Huang, M., Fan, B., Buckler, E. S. E. S. & Zhang, Z. Iterative Usage of Fixed and Random Effect Models for
Powerful and Efficient Genome-Wide Association Studies. PLoS Genet. 12, e1005767 (2016).

49