See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/11552390

Challenges for Biomedical Informatics and Pharmacogenomics

Article  in  Annual Review of Pharmacology · February 2002

DOI: 10.1146/annurev.pharmtox.42.082401.140850 · Source: PubMed

CITATIONS READS
109 403

2 authors, including:

Russ B Altman
Stanford University
853 PUBLICATIONS   43,323 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Protein structure annotation View project

Education in HBI View project

All content following this page was uploaded by Russ B Altman on 30 May 2014.

The user has requested enhancement of the downloaded file.

 Annu. Rev. Pharmacol. Toxicol. 2002. 42:113–33
Copyright

c 2002 by Annual Reviews. All rights reserved

CHALLENGES FOR BIOMEDICAL INFORMATICS

AND PHARMACOGENOMICS

Russ B. Altman and Teri E. Klein

Stanford Medical Informatics, Stanford, California 94305-5479;
e-mail: [email protected], [email protected]

Key Words bioinformatics, pharmacogenetics, computation, databases

■ Abstract Pharmacogenomics requires the integration and analysis of genomic,
molecular, cellular, and clinical data, and it thus offers a remarkable set of challenges to
biomedical informatics. These include infrastructural challenges such as the creation
of data models and databases for storing these data, the integration of these data with
external databases, the extraction of information from natural language text, and the
protection of databases with sensitive information. There are also scientific challenges
in creating tools to support gene expression analysis, three-dimensional structural anal-
ysis, and comparative genomic analysis. In this review, we summarize the current uses
of informatics within pharmacogenomics and show how the technical challenges that
remain for biomedical informatics are typical of those that will be confronted in the
postgenomic era.

WHAT IS BIOMEDICAL INFORMATICS?

Biomedical informatics is the study of information flow within biology and medi-
cine. The use of computational techniques in biomedical research dates back to the
first general purpose computers but interest in the techniques has exploded in the
last decade (1). The increased interest stems from the availability of experimental
techniques that create data that simply cannot be manually analyzed and require
computational intervention. Many areas of biology and medicine are being revo-
lutionized by the introduction of new experimental techniques, accompanied by
informatics methodologies that fundamentally change the way that investigators
do their work.
The two flows of information that are studied by informatics are the flow of
information from the DNA code to biological function and the flow of information
in the design and analysis of experiments. In the first flow, we are interested in the
transfer of information within biology, while in the other, we are interested in the
transfer of information about biology. Thus, the first information flow deals with

0362-1642/02/0210-0113$14.00 113
114 ALTMAN
KLEIN

the central dogma of biology: DNA is transcribed into RNA, RNA is translated into
protein, and protein molecules have functions that carry out biological processes.
Interacting proteins produce signaling and metabolic pathways that coalesce to
form networks at the cellular level, and cells interact at an organismal level to
produce physiology. Informatics approaches to studying different aspects of this
flow, therefore, include methods for gene finding (2–5), 3D structure prediction
(6, 7), modeling of genetic networks (8–14), and statistical population biology
(15, 16).
In the second flow, we are interested in the ways in which biological and medi-
cal information is gathered. This flow begins with a scientific hypothesis, followed
by a plan to collect data, execution of an experiment, analysis of the results, and
subsequent refinement of the hypothesis. Informatics applications within this flow
are usually created to support investigators in the practice of science. Informat-
ics approaches to studying this flow, therefore, include methods for organizing
and searching databases of literature, sequence, and function, as well as meth-
ods for helping to create and evaluate scientific models (17). If both of these
information flows are included in a definition of biomedical informatics, then vir-
tually all biomedical informatics research can be placed in one or both of these
areas.
Biomedical informatics has gained prominence recently because biologists can
now collect more data. The success of the genome sequencing projects has cat-
alyzed a new way of thinking in biology, whereby data are collected on a large scale
and without a particular hypothesis in mind. The data are then placed in a database,
and scientists with hypotheses can extract information from the database in order
to evaluate the merits of the hypotheses. This leads to a fundamental change in
how some investigators do their work: Instead of first moving to the laboratory,
they first move to the database, and only after assessment of the available data
are experiments planned. There has been much debate about the merits of such an
approach, but there is no doubt that the emergence of these large-scale and high-
throughput methods for data collection makes such an approach feasible (18). The
data explosion is not limited to DNA sequencing, and we are seeing increased ca-
pacity to assess the levels of mRNA expression (19, 20), to detect protein-protein
interactions (21), to locate gene products within the cell (22), to detect and iden-
tify compounds using mass spectroscopy (23), and even to understand the detailed
atomic three-dimensional structure of macromolecules and their small molecule
ligands (24).
As long as clever experimentalists continue to create these high-throughput
experimental methods, informatics professionals will have a surfeit of data and
data analytic challenges. Success in informatics usually means the acceleration in
understanding the processes of interest, and increased access to the information
required to generate and test scientific hypotheses. One of the areas that has re-
cently attracted the attention of biomedical informaticians is pharmacology, and
particularly pharmacogenetics and pharmacogenomics.
 INFORMATICS FOR PHARMACOGENOMICS 115

WHAT ARE PHARMACOGENETICS AND

PHARMACOGENOMICS?
Pharmacogenetics is the study of how variation in genes affects the response to
drugs. It has existed as a field for more than four decades, and forms the basic
intellectual framework for understanding phenomena such as the idiosyncratic re-
sponses to anesthesia, to opiates, and to anticancer agents (25). Pharmacogenetics
has tended to study single genes with a focused, hypothesis-directed set of experi-
ments. Most pharmacogenetic studies begin with the recognition of high variability
in response to a medication and then a search for the genetic basis of the variation.
Thus, for example, the blood levels of a metabolite may be measured and noted
to vary widely, investigations of the pathway of metabolism may suggest that en-
zymes in this pathway are behaving differently, and the genetic analysis of these
enzymes might detect variations in the protein sequence (or regulatory sequence)
that explain different catalytic rates or binding constants.
Pharmacogenomics emerged recently, and scientists do not entirely agree on its
relationship to pharmacogenetics (26–29). “Genomics” is generally used to indi-
cate the study of the entire complement of genes within an organism, and the -omics
suffix has been used generally to indicate the comprehensive analysis of the capa-
bilities of an organism. Thus, proteomics studies the full set of proteins and how
they interact within a cell. Based on this understanding, pharmacogenomics can
be construed as the study of the entire complement of pharmacologically relevant
genes, how they manifest their variations, how these variations interact to produce
phenotypes, and how these phenotypes affect drug response. A key element of
pharmacogenomics is, not surprisingly, the large-scale and high-throughput col-
lection of data, including DNA sequence variations, mRNA expression analysis,
enzyme kinetic assays, and cellular localization experiments. The move toward
these types of experiments of course creates a magnet for biomedical informatics
investigators, who see an opportunity to apply their methodologies to an exciting
area with promise to revolutionize medical care.
The development of pharmacogenomics is a natural sequela to the success of the
initial human genome sequencing project. The promise of that project was that an
understanding of all human genes would create the opportunity for new diagnostic,
prognostic, and therapeutic technologies. The variation in response to medications
across patients can be large, and the occurrence of side effects and adverse events
limits the success of many therapeutic strategies. A systematic understanding of
the gene systems that modulate response to medications may therefore change
the way medications are prescribed. With the success of pharmacogenomics, it
may become possible routinely to check the genetic background of a patient in
order to ensure that the prescribed medications are effective and free from adverse
side effects (30–33). Although it is not entirely clear how many of the 35,000
genes assigned in the rough draft of the human genome are relevant to drug re-
sponse (or even how to define “relevance”), a systematic analysis of pharmacology
116 ALTMAN
KLEIN

textbooks indicates a core set of 500 to 1000 genes, as shown in the PharmGKB
Web site.1

WHAT ARE THE APPROACHES TO

PHARMACOGENOMICS?
There are generally two approaches to pharmacogenomic research, which are
summarized as the “genotype-to-phenotype” and the “phenotype-to-genotype”
approaches. In the genotype-to-phenotype approach, the investigators start with a
set of genes that are known (or strongly suspected) to be important in modulating
the response to drugs, and then they search for variation in their sequences (that
is, their genotype). Given an understanding of genetic variation, they can search
for the phenotypic consequences. Examples of approaches amenable to genotype-
to-phenotype analysis might include gene families known to be important for
pharmacokinetics (the study of how medications are absorbed, distributed, and
cleared from the body), such as phase I metabolism enzymes (the mixed function
oxygenases of the cytochrome p450 system) (34), phase II metabolism enzymes
(the conjugation system) (35), and membrane transporter molecules (36). Other
systems amenable to the genotype-to-phenotype approach are those that are in-
volved in pharmacodynamics (the study of how medications have their therapeutic
effect) and those whose mechanisms are well understood at the receptor and path-
way level. Examples might include the well described pathways of inflammation
in asthma (37, 38), the purine/pyrimidine biosynthetic pathways that are targeted
by some anticancer agents (39), or the enzyme cascade that controls blood clotting
(40). The steps of a genotype-to-phenotype approach can be summarized in this
simplified way:
1. Identify the genes that belong to the system that is involved in modulating
drug response.
2. Catalog the variation in the DNA sequences across the population.
3. Search for phenotypes associated with the sequence variation.
4. Confirm clinical relevance of the genotype-phenotype associations.
Each of these steps is nontrivial and complicated. The first involves searching DNA
databases and perhaps using comparative genomic techniques to identify target
genes (41, 42). The second includes high-throughput experimental methods for
detecting DNA variations, including single nucleotide polymorphisms (SNPs), the
most common type of DNA variation (43–45), and also for associating individual
SNPs into haplotypes (46). The third step involves the collection of molecular,
cellular, or clinical data (reviewed below), and the final step requires clinical trials
to prove the associations of interest and to demonstrate clinical relevance.

1
http://www.pharmgkb.org/
 INFORMATICS FOR PHARMACOGENOMICS 117

Phenotype-to-Genotype Approaches
Phenotype-to-genotype approaches toward pharmacogenomic discovery are dif-
ferent. Instead of identifying a family of genes in which to characterize genetic
variation, investigators search for a phenotypic measure that shows significant
variation. This measure can be a clinical measure (such as the rate of clearance
of a drug or the peak level of the drug for a given dose), a cellular measure (the
rate of cellular uptake of a drug or the profile of gene expression), or a molecular
measure (the enzymatic turnover rate of an enzyme or a substrate binding con-
stant). In any case, it is the phenotypic variation that first draws attention and then
follows a search for the genes that are responsible for this variation. The steps of
a phenotype-to-genotype approach, therefore, can thus be summarized:
1. Identify a phenotype that shows significant variation.
2. Search for genes that may explain this variation.
3. Characterize genetic variations and check for association with the phenotype.
4. Confirm proposed genetic basis for the variation and its clinical relevance.
The challenges in the first step are to identify phenotypes that are both clinically
relevant and also measurable. The second step is the most difficult and requires
the investigator to use any means available to identify genes that could be involved
with the phenotypes. It may involve using animal models and comparative ge-
nomics, DNA microarray analysis to measure changes in expression in response to
drugs, database (literature and sequence) searches for associations between genes
and related phenotypes, or analytic chemistry methods to identify gene products
contributing to variation (47). The third step is similar to the second step of the
genotype-to-phenotype process. A major challenge in this step is the large amount
of variability in human genes that is not functionally significant, so investigators
must focus efforts on variations that can be shown to have functional consequence.
The final step is focused particularly on this problem of ensuring that the discovered
genetic component really explains the phenotypic variation of interest.
Both approaches to pharmacogenomics have strengths and weaknesses. Investi-
gators must assess the current knowledge base for a given drug class of interest
in order to determine whether there is enough genetic information to justify a
genotype-to-phenotype approach, or whether there are more striking phenotypic
data suggesting a phenotype-to-genotype approach.

CHALLENGES FOR BIOMEDICAL INFORMATICS

IN PHARMACOGENOMICS
The challenges for biomedical informatics within the study of pharmacogenomics
all follow directly from the preceding discussion. Pharmacogenomics is relatively
new, so the current excitement derives, in part, from the great range of opportunity
118 ALTMAN
KLEIN

for contributions that now exists. One of the key themes in pharmacogenomics is
that the relevant informatics expertise includes information from molecular biol-
ogy (sequences, structures, pathways) as well as from clinical medicine (medica-
tions, diseases, side effects), and of course from pharmacology (pharmacokinetics
and pharmacodynamics). Thus it represents a new wave of informatics problems
where both basic biological and clinical information must be combined and ana-
lyzed. Whereas previously bioinformatics focused solely on issues of relevance to
molecular biology (sequence and structure analysis), applications are now mov-
ing closer to the parts of clinical informatics that focus on the organization of
clinical information, particularly for research purposes. The main challenges for
biomedical informatics within pharmacogenomics fall into nine areas:
1. Representing the diversity of pharmacogenomic data
2. Developing standards for data exchange
3. Integrating data from multiple data resources
4. Mining literature for knowledge
5. Using expression data to understand regulation
6. Understanding the structural basis for variability
7. Using comparative genomics
8. Managing laboratory information
9. Protecting sensitive patient information

Representing the Diversity of Pharmacogenomic Data

One of the principal challenges for pharmacogenomics is the creation of data struc-
tures that store relevant information in a form that is easy for computer programs to
manipulate. There is a difference between data formats that are useful for human
readers (journals, tables, figures) and those that are useful for computers (data
structures in computer programs that label all data for easy retrieval and analysis).
The classes of data that must be represented are diverse, as are the connections
between the data that must be maintained.

GENOMIC DATA The representation of DNA sequence information for pharma-

cogenomics is similar to that required for many other applications. The main
requirement is that the gene structure for a protein product be understood and
labeled so that observed DNA sequence variations can be interpreted as be-
longing to coding or noncoding regions, and their likely significance can be
evaluated. The key concepts that must be modeled include genomic sequence,
unprocessed mRNA transcript, processed transcript, and protein sequence. Within
each of these models are details such as the 3
- and 5
-untranslated regions of
genes, genetic regulators (enhancers, silencers), exons that are coding or non-
coding (or partially coding), and alternative splicing strategies. There have been
a number of proposed standards for tracking genomic data, including the data
 INFORMATICS FOR PHARMACOGENOMICS 119

structures behind Genbank (48), Human Genome Database (49), BIOML2, and
others. The human genome browsers offered by UC Santa Cruz3, Ensembl4,
National Center for Biotechnology Information (NCBI)5, and Celera6 offer a basic
look at gene structure, but these are still evolving because the genome is in draft.
In addition to the representation of basic gene structure, it is critical to also under-
stand the locations and types of genetic variation. The dbSNP resource at NCBI
provides an excellent source of reported SNPs (50), including those submitted by
The SNP Consortium (51), an industrial group that is performing large-scale SNP
detection and submitting many of these to the public domain.
Genome data are also made more useful by their connections to databases of
biological function, including the Online Mendelian Inheritance in Man (OMIM)
database of inherited human disorders (52, 53), and a number of specialty databases
that provide valuable in-depth information about individual gene families, such
as the Cell Signaling Network Database (54), the transcription factor database
TRANSFAC (55), and the protein kinase database (56).

MOLECULAR AND CELLULAR DATA The characterization of phenotype is important

for both the genotype-to-phenotype methods as well as the phenotype-to-genotype
methods. Phenotype is difficult to precisely define, but it can be thought of as
functional features of gene products, ranging in detail from the molecular to the
individual and population levels. Unfortunately, phenotype data are not as “digi-
tal” as sequence data, so they are much more difficult to represent. Nevertheless,
the success of pharmacogenomics depends on the establishment of standards for
describing these data.
In pharmacogenomics, a few types of molecular and cellular data are clearly
critical to represent. These include enzyme kinetic data (such as the binding and
catalytic constants of enzymes, and their associated kinetic parameters), three-
dimensional structural data (when available) for enzymes and their substrates/
ligands, and protein localization data (often images) that show where different gene
products are found within the cell. Standard representations of these data types are
not generally available, but the creation of databases to store such information will
require that they be developed. Fortunately, there is fairly good agreement on the
basic vocabulary of enzyme kinetics and the definition of these parameters in basic
pharmacology texts (and associated programs for computing the parameters), as
well as the representation of three-dimensional structural data in the Cambridge
Crystallographic Database7 and the Protein Data Bank (PDB)8.

2
http://www.bioml.com/BIOML/
3
http://genome.ucsc.edu/
4
http://www.ensembl.org/genome/central/
5
http://www.ncbi.nlm.nih.gov/genome/guide/central.html
6
http://.public.celera.com/index.cfm
7
http://www.ccdc.cam.ac.uk/
8
http://www.rcsb.org/pdb/
120 ALTMAN
KLEIN

Microarray expression data are clearly going to become an important phe-

notypic data source for pharmacogenomics (57–59), and standards are in active
development at this time, although they have not settled yet. There is a proposal
for a MicroArray Markup Language (MAML) that would specify a minimal set of
information for exchange. MAML is part of a larger effort to develop standards for
microarray data and databases in the Microarray Gene Expression Database9 effort.
The challenges here involve developing standards for representing the experimen-
tal conditions, the quality control parameters, the list of genes being assayed, and
the actual expression measurements (and background measurements) recorded.

CLINICAL DATA Pharmacogenomics requires a connection to clinical medicine in

order to establish the relevance and importance of the systems that are studied. As
such, clinical medicine can be considered simply another phenotype, as molecular
or cellular data. However, the techniques used to collect and describe clinical data
are sufficiently different from basic biological data that the approaches should be
distinguished. At the most basic level, pharmacogenomics information resources
need to store clinical information pertaining to the most commonly measured
phenotypes: one, pharmacokinetic profiles of drug levels in response to dosing and
two, measures of pharmacodynamic efficacy based on the target effects. In addition,
the incidence of side effects in response to medications must be represented.
Although there is a large literature on the representation of clinical data, much
of this is for the purposes of supporting the delivery of clinical care and not
for clinical research. Clinical research requires precision in ways different from
clinical care, so the portability of most clinical data standards is not clear. The
standards that exist for coding diagnoses (International Classification of Diseases
standard10), pathology (Systematized Nomenclature of Medicine11), procedures
(Current Procedural Terminology12), and others offer a good starting point for
pharmacogenomics research but in general do not provide the precision required
for high-quality data storage.
As is the case for enzyme kinetics, there is a fair amount of uniformity within the
pharmacology community on how to represent pharmacokinetic profiles. Programs
such as ADAPT II13, NONMEM14, SAAM 30, and CONSAM15 exist that all
allow first- and second-order kinetics and associated parameters to be computed.
A standard set of parameters, including Ki, Km, and Vmax, have fairly consistent
definitions and thus provide a good initial opportunity for modeling of the data.
One issue that arises in modeling phenotypic data is the relationship between
raw data and the more processed parameters and intermediate representations.

9
http://www.mged.org/
10
http://www.cdc.gov/nchs/about/otheract/icd9/abticd10.htm
11
http://www.snomed.org/
12
http://www.ama-assn.org/ama/pub/category/3113.html
13
http://www.usc.edu/dept/biomed/BMSR/Software/adptmenu.html
14
http://c255.ucsf.edu/nonmem0.html
15
http://www-saam.nci.nih.gov/index.html
 INFORMATICS FOR PHARMACOGENOMICS 121

Although the raw data are used to compute these intermediate representations (for
example, the raw time points of blood levels are used to compute pharmacokinetic
parameters), it can be difficult to determine the appropriate level of data to make
routinely available in databases. The raw data can be cumbersome, and the com-
puted parameters may be of real interest to most. However, there are times when
the raw data must be retrieved in order to check conclusions or alternative inter-
pretations. Thus, a major challenge for pharmacogenomic information resources
is to provide easy access both to the computed/derived parameters as well as to
the basic information upon which they are based.

Developing Communication Standards in Pharmacogenomics

One way in which informatics technologies can help accelerate progress in a field
is the development of standards for representing and exchanging data. It is clear
that shared understanding of the basic data elements within pharmacogenomics is a
critical building block with which to build an information infrastructure. Methods
for communicating these data are therefore equally important. The two main ar-
eas that require progress are the definition of shared syntax (how information is
structured in a data file) and semantics (how the information should be interpreted
by others). Two contributing technologies within informatics that address these
problems are technologies for defining shared vocabularies and technologies for
exchanging them.
A standard vocabulary is a controlled set of terms that can be used instead
of free text to communicate information. For example, whereas an abstract in
Medline is free text, the list of Medical Subject Heading keywords from the
MEDLINE database16 represents a controlled vocabulary. The advantage of a
controlled vocabulary is that computer programs can be written to expect cer-
tain phrases and can be instructed how to process data based on the occurrence
of these phrases. Whereas free text has much more power for expressiveness,
it is difficult for computer programs to understand because of the inherent
ambiguities in human-to-human communication. Some of these difficulties are
addressed by natural language processing techniques (reviewed below). However,
a principal means to address these problems is to develop and adopt controlled
vocabularies.
Some vocabularies have been developed to facilitate the delivery of clinical care,
and these may be helpful in pharmacogenomics. Many contributing vocabularies
have been related to one another as part of the Unified Medical Language System
project at the National Library of Medicine (60). To support these endeavors fully,
however, these vocabularies need to be supplemented by establishing the following
standards.
1. Human gene names and links to other organisms. The Human Genome
Nomenclature Committee has created a reference set of symbols for human

16
http://www.ncbi.nlm.nih.gov/PubMed/
122 ALTMAN
KLEIN

genes, and this should stabilize over time and provide a useful set of indices
for the human genome browsers. Included in this activity is the identifica-
tion of the function of new genes of pharmacogenomic interest, particularly
transporters [which are classified in a taxonomy by Saeir17 (61)] and the
cytochrome p450 system (classified based on isoform similarity) (62, 63).
2. Drug and compound names. There are efforts proposed to build a con-
trolled list of drug categories, their structural and biological features, and
the associated specific compounds. The Unified Medical Language System
(UMLS) contains the 1997 Food and Drug Administration Standard Product
Nomenclature.18
3. Side effects. Standards are required for coding drug side effects, at a clinical
level and perhaps at a lower biological level. A vocabulary that is used for
clinical trials is a good initial start. The UMLS contains the World Health
Organization Adverse Drug Reaction Terminology and the coding symbols
for a thesaurus of adverse reaction terms (COSTART) from the Food and
Drug Administration.19

Data Exchange Standards

eXensible Markup Language (XML) has emerged as a common standard for the
exchange of data (64)20. XML is a syntax for specifying how text data can be la-
beled so that computer programs can load the data items into their memory struc-
tures. Without a standard such as XML, competing file formats would abound,
and programs would work only with a subset of these formats. It then becomes
very difficult to exchange data and run programs. Similarly, databases often are
constructed independently and organize their data quite differently (including de-
cisions as simple as whether a drug dose should be specified as a single string
“20 mg oral” or as separate fields). XML provides a partial solution to this prob-
lem by providing a standard syntax for specifying the elements within the file
and how they are presented. It becomes relatively easy, then, to read files in one
format and then to translate them into a newer format. Even better, however, is for
the community to adopt a single format for uniform representation of data. The
PharmGKB database21 is attempting to define some XML standards for data that
will allow basic pharmacogenetic data to be formatted in a standard manner.
A more difficult issue is the semantics of the data representation. Syntax only
enforces the order of the elements and their basic types (integers, real numbers,
strings, and the like), but specifications of semantics require more constraints on the

17
http://www-biology.ucsd.edu/∼msaier/transport/titlepage.html
18
http://www.fda.gov/cder/ndc/database/default.htm and http://www.fda.gov/cder/ob/
default.htm
19
http://www.fda.gov/cder/aers/index.htm
20
http://www.w3.org/XML/
21
http://pharmgkb.stanford.edu/xml-schemas.html
 INFORMATICS FOR PHARMACOGENOMICS 123

logical relationships between data items (for example, a “nonsynonymous SNP”

must be at a genome sequence position that is a SNP, must be within a coding
region, and must change the amino acid for which it is coded). The specification of
semantics is an active area of research, but the Resource Description Framework
(written in XML itself22) is an attempt to enable the standardization of semantics.
In addition, knowledge base management systems support the logic and constraint
checking that is required for computationally enforcing semantics (65, 66).

Integrating Data From Diverse and Heterogeneous Databases

Pharmacogenomics research is marked by the diversity of databases that must
be used in order to answer important questions. For example, in order to find
all three-dimensional protein structures with SNPs that change the amino acid
in the coding region of proteins that are involved in diabetes, we must combine
gene sequence data [such as are found in GENBANK (48)], SNP data [dbSNP
(50)], three-dimensional structural databases [Protein Data Bank (67)], databases
of genetic diseases and their gene defects [OMIM (52)], and the medical literature.
Other queries might require databases of drugs or drug-drug interactions, which
are not publicly available at this time.
The problem of integrating data is a difficult one within computer science.
One approach is to create a single large data model and to dump the contents
of all the contributing databases into the new “mega” database. This approach,
called consolidation, suffers because as the contributing databases evolve, the
consolidated database becomes out of date. In addition, it is very difficult to build a
large data model. Another class of approaches to database integration is federation,
which can take three forms. In the first, databases are linked together loosely with
hyperlinks on the web, offering little help to automatic programs but useful for
human users. In the second, programs are written to extract certain types of data
from each database and combine them to answer queries that require data from
more than one database (68). In the third, programs are written to extract the data
on a regular basis from the contributing databases and dump them into a common
database that is then updated regularly [but functions as a consolidated database
between updates (69)].

Mining the Published Literature for Pharmacogenomic Data

The Medline/PubMED resource contains references to more than ten million
biomedical publications, and many of these offer online abstracts. In addition,
many journals are now making their articles available online in full text. Although
this mode of publication is effective for human users, it is difficult for comput-
ers to extract information from natural language text. At the same time, there are
decades of biological knowledge stored in written natural language text. In order

22
http://www.w3.org/RDF/
124 ALTMAN
KLEIN

to avoid the loss of this information and to assist in the automatic population of
databases, informatics researchers are building systems for extracting information
from text. The general problem of understanding the full details of a natural lan-
guage text has been studied for more than four decades and remains unsolved (70);
however, the more tractable goal of reliably identifying relationships within text
is within reach. For example, texts can be analyzed to extract protein names (71),
and protein-protein (11, 72, 73) or protein-drug interactions (74), based on the oc-
currence of protein names and verbs such as “inhibits,” “activates,” “represses,”
“enhances,” etc.).
Within pharmacogenomics there are good opportunities for natural language
processing (NLP) techniques to assist in the organization of data. First, there is no
definitive list of drug-gene interactions, and the literature (both published medical
literature and the U.S. patent application literature23) is filled with associations
that are pharmacokinetic (e.g., “X is metabolized by CYP2D6”) and pharmaco-
dynamic (e.g., “Y is active at the beta-adrenergic receptor”). In general, NLP
techniques work best in well-defined domains that use standardized vocabulary. A
second area of opportunity within pharmacogenomics is the extraction of cellular
localization information (“X is localized to the Golgi”) from text (75). A third
area that holds much promise is the processing of mRNA expression data with
microarrays (8, 20, 57, 59, 76–79). Because of the great volume of information
generated by these experiments, it is often useful to cluster the genes based on
expression pattern, and it is very challenging subsequently to summarize the key
features of each cluster. NLP-based techniques may be able to combine the pub-
lished information about genes with information about how they cluster to create
automatic cluster labels that provide biological insight. A fourth area for NLP ap-
plications is in the identification of genes of pharmacogenomic interest from text
and in the classification of these genes as primarily of pharmacokinetic or phar-
macodynamic importance. The language within pharmacokinetic papers is quite
idiosyncratic (including discussions of “area under curve” and “bioavailability”),
so it may be relatively straightforward to classify abstracts that are discussing this
topic and then to extract key data elements from them.

Using Expression Data to Assess the

Phenotypes of Drug Response
The emergence of DNA microarrays to measure mRNA expression has created
excitement in many areas of biomedical research, including pharmacogenomics
(8, 20, 57, 59, 76–79). These microarrays use hybridization of amplified RNA from
samples of interest to DNA of known sequence (that have been affixed to small
spots that are arranged into a square array) in order to measure the level of gene
expression in the samples. The use of microarrays for pharmacogenomics has only
begun, but it has the potential of bringing great gains because they can be used to

23
http://www.uspto.gov/
 INFORMATICS FOR PHARMACOGENOMICS 125

address the most difficult steps of both genotype-to-phenotype and phenotype-to-

genotype approaches. In genotype-to-phenotype investigations, microarrays can
assist in the third step to find phenotypes at the cellular and molecular level that
are associated with variations in genotype (and in the context of administering
certain drugs). In phenotype-to-genotype studies, microarray measurements can
be used in the second step to find genes whose expression alters in the context of
an important new phenotype.
The most common informatics analyses of microarray data are currently clus-
tering of genes and classification of genes based on shared expression patterns
(80). These groupings can be used as evidence for “guilt-by-association” assign-
ments of function, whereby the function of a gene is assumed to be similar to the
function of genes with which it is grouped. The most exciting work in microarrays,
however, involves combining information from microarrays with other data sets.
One important study measured the expression levels of genes within sixty cancer
cell lines and compared the sensitivity of each of these cell lines to over 70,000
different potential anticancer drugs (58). The key expression features that identi-
fied potential sensitivity to the anticancer drugs were defined, and cell lines were
clustered based on common potential sensitivities. Microarray analysis has also
been used to study the pharmacogenomics of cystic fibrosis (81), schizophrenia
(82), and others. We expect that in the future microarray analysis of cells before
and after drug exposure will provide an important set of pharmacogenomic data
for determining the full set of changes that occur at a cellular level, as well as for
determining the cells’ kinetics (59).

Understanding the Structural Consequences

of Genetic Variations
Pharmacology has always had a strong structural component because the three-
dimensional structure of drugs can be critical for understanding mechanisms of ac-
tion and for building pharmacophores for drug design (83). The increasing number
of structures in the 3D structural database also allows us to model the interactions
between proteins and their ligands in order to gain a high-resolution understanding
of drug action. The PDB now has over 15,000 individual structures (67), and the
emergence of high-throughput structure determination efforts promises to main-
tain a rapid rate of data acquisition (84). The availability of more 3D structures
makes it increasingly likely that a homologous protein of known structure will be
available for most proteins of pharmacogenomic interest. The types of structural
analyses that are becoming important include the following methods
1. Methods for docking small molecules into binding pockets of proteins in
order to predict affinity. There have been a number of successful reports in
this area, including algorithms based on energetics and based on statistical
analysis of pockets (85–87).
2. Methods for homology modeling in order to build models of protein varia-
tions. A variety of programs have been developed and tested and offer good
126 ALTMAN
KLEIN

options for building 3D models of proteins that are globular and share 30%
or more sequence identity with a known structure (88). In these applica-
tions, the location and possible functional significance of nonsynonymous
SNPs in the coding portion of proteins can be evaluated (89). A recent paper
estimated that 30% of all nonsynonymous SNPs may be associated with
significant changes in function (90) based on an analysis of a large set of
mutations in DNA binding proteins. There are also some early indications
that even synonymous SNPs may change RNA stability and affect the level
of activity for some proteins.
3. Methods for predicting protein-protein interactions. It is clear that many
proteins have multiple partners with which they interact as activators, in-
hibitors, or otherwise as modifiers. There has been progress in molecular
docking algorithms (91–96) that allows investigators to combine geometric
and energetic properties for the purpose of understanding how two protein
surfaces may interact.

Comparing Genomes to Develop Pharmacogenomic Models

Comparative genomics is the study of multiple genomes from different organ-
isms in order to define the shared characteristics between organisms as well as the
distinguishing characteristics of each organism (41, 42). For pharmacogenomics,
comparative genomics can identify analogs to human drug response phenotypes in
organisms that are more readily manipulated experimentally. In general, the coding
regions of rat, mouse, and pig are more strongly conserved relative to human than
the noncoding regions. Recent work shows that noncoding regions can be con-
served across species and can be important for regulating gene expression (97).
Thus, the availability of complete genomes for related species is likely to assist
pharmacogenomics researchers in identifying areas outside coding regions that
may be worth studying for polymorphisms. Because the background rate of poly-
morphisms in humans is so high, it is critical to have these clues from comparative
genomics to guide the selection of regions to emphasize in the search for critical
variations. These analyses depend on accurate alignments of large segments of
genomic DNA, and special purpose algorithms have been developed (97, 98).
Comparative genomics techniques can also be used to understand metabolic
and genetic regulatory pathways. Metabolic databases such as EcoCYC (99) and
the Kyoto Encyclopedia of Genes and Genomes (KEGG) (100) have been con-
structed for a number of complete bacterial genomes and are beginning to emerge
for eucaryotes as well. These resources will allow a computational analysis of
pathways in determining genes that should be studied and perhaps in determining
their role in metabolism or drug action (101). Genetic regulatory pathways are
summarized in the Cell Signaling Network Database (54) and the signal transduc-
tion knowledge environment (STKE24), and can also be studied with databases
of transcription factors and regulatory sequences (55). The use of comparative

24
http://stke.sciencemag.org/
 INFORMATICS FOR PHARMACOGENOMICS 127

microarray expression analysis is also being evaluated as a strategy for filtering

important signals from microarray expression experiments (102, 103).

Managing Laboratory Information Data

Although it is not peculiar to pharmacogenomics, the development of reliable lab-
oratory information management systems (LIMS) is as critical to this field as any.
Tracking the large number of samples that must be tracked of patients, tissues,
cell lines, and individual genes and gene products is a nontrivial bookkeeping
challenge. Excellent systems have been developed in the context of disease re-
search networks for web-based tracking and linking of samples to core data sets
(104). Pharmacogenomics offers a particularly challenging application area for
LIMS because of the diverse array of data and samples that are relevant and be-
cause genomic data are being linked to clinical data for the purposes of finding
genotype-phenotype associations. In addition, the emergence of publicly available
tissue samples for sampling genomic diversity (105) creates a tracking problem
for samples worldwide.

Protecting the Confidentiality and Privacy

of Clinical Phenotype Data
The study of pharmacogenomics and the desire to disseminate data pertaining
to pharmacogenomics raise a number of important issues in ethics and patient
confidentiality. The need to integrate molecular data with clinical data implies
that clinical phenotype information may be distributed on the internet. It is cru-
cial, however, that the privacy and security of patient identity be maintained
while disseminating these data. Simple methods of “de-identification” (in which
basic identifying information such as name, address, and other demographic infor-
mation is removed) for patient protection is not adequate. As more information is
provided about a patient, even though it is not directly identifying, it can often be
combined with other data sources (such as hospital discharge records and voter and
driver registration information) to reconstruct, either exactly or probabilistically,
the identity of patients (106). There are precedents for the publication of patient-
related data in journals as well as in databases (such as Genbank). Nonetheless,
it is critical to ensure that the availability of clinical phenotype data sets does not
lead to the loss of study-subject confidentiality or privacy.
There are generally two approaches to protecting patient privacy. The first,
called mediation, inserts a computer program in between a user and a database
and monitors the queries that are asked and the answers that are provided by
the database. The mediator has rules about the kinds of queries that can be writ-
ten and the kinds of responses that can be supplied from the database. It stops
queries that are inappropriate (based on the rules, which often include infor-
mation about the privileges of the user), and it filters results that are inappro-
priate. These mediators have been shown to provide reasonable semi-automated
protections (107). A second method, called scrubbing, is based on the principle of
removing information from a data set so that the details of the data cannot be used
128 ALTMAN
KLEIN

to re-identify a patient (108, 109). Thus, for example, a list of pharmacokinetic

parameters can be either rounded off to reduce precision or can be provided as
ranks instead of absolute values. Each of these maneuvers would serve to increase
the pool of subjects from whom these values could come and thus protect the pri-
vacy of individual subjects. Crucial to scrubbing is the idea of “bin size,” which is
the minimum number of people (in some defined population) who match a certain
query. Thus, we may say that in a study with 500 patients, we will never answer a
query with results containing less than 10 subjects, so that individual subject data
cannot be teased out. The bin size has been used by other organizations, such as
the social security administration in the United States, and the U.S. census bureau
(106). The obvious disadvantage to scrubbing is the loss of precision, which can
make certain statistical analyses much more expensive. Although not within the
scope of this review, there are associated implications about how study subjects
should exercise informed consent when participating in pharmacogenomic stud-
ies. The other critical issue that is outside the technical scope of this paper, but
deserves mention, is the problem of having pharmacogenomic information used
to discriminate (either in terms of the research agenda, insurance, or employment)
against groups within the population based on statistical associations.

PHARMACOGENOMICS: A NEW CHALLENGE

FOR BIOMEDICAL INFORMATICS
The focus of biomedical informatics has, in the past, been fragmented into in-
formatics to meet the challenges of genomics (sequence analysis, structure anal-
ysis, biochemical and regulatory pathway analysis) and informatics to meet the
challenges of organizing clinical data (medical records, information extraction,
database integration). In the post-genome period, there will be an increasing num-
ber of applications that require the combination of basic bioinformatics with clini-
cal informatics. Pharmacogenomics is an excellent example of such an application
area. Many different branches of biomedical informatics clearly will play a critical
role in gathering, organizing, and analyzing pharmacogenomic data. The National
Institutes of Health has recently formed a Pharmacogenetics Research Network
and Database25 program whereby several research groups are cooperating to gather
pharmacogenomic data (genomic, molecular, cellular and clinical) and deposit it
in a common database for public use. The database, PharmGKB26, is intended to
gather data not just from these groups, but from all groups worldwide wishing
to disseminate their data of pharmacogenomic relevance. The initial focus is on
building representations and XML standards for the submission of basic genomic
variation data, enzyme kinetic data, and clinical pharmacokinetic data, but work
is being done in all the areas reviewed here. As the basic infrastructure is created,

25
http://www.nigms.nih.gov/pharmacogenetics/
26
http://www.pharmgkb.org/
 INFORMATICS FOR PHARMACOGENOMICS 129

there will be opportunities for the community to create and distribute informatics
methodologies that address each of the challenges outlined here. There is no doubt
that other, perhaps unexpected, informatics challenges will arise in the course of
creating these resources.

Visit the Annual Reviews home page at www.AnnualReviews.org

LITERATURE CITED
1. Tarter ME. 1979. Biocomputational human genes for high-throughput analysis
methodology: an adjunct to theory and of gene expression. Nat. Genet. 28:21–28
applications. Biometrics 35:9–24 12. Legrain P, Wojcik J, Gauthier J. 2001.
2. Besemer J, Lomsadze A, Borodovsky M. Protein-protein interaction maps: a lead
2001. GeneMarkS: a self-training method towards cellular functions. Trends Genet.
for prediction of gene starts in micro- 17:346–52
bial genomes. Implications for finding se- 13. Hasty J, McMillen D, Isaacs F, Collins
quence motifs in regulatory regions. Nu- JJ. 2001. Computational studies of gene
cleic Acids Res. 29:2607–18 regulatory networks: in numero molecu-
3. Burge CB, Karlin S. 1998. Finding the lar biology. Nat. Rev. Genet. 2:268–79
genes in genomic DNA. Curr. Opin. 14. Salazar-Ciudad I, Newman SA, Sole RV.
Struct. Biol. 8:346–54 2001. Phenotypic and dynamical tran-
4. Lewis S, Ashburner M, Reese MG. sitions in model genetic networks. I.
2000. Annotating eukaryote genomes. Emergence of patterns and genotype-
Curr. Opin. Struct. Biol. 10:349–54 phenotype relationships. Evol. Dev. 3:84–
5. Mural RJ. 1999. Current status of compu- 94
tational gene finding: a perspective. Meth- 15. Terwilliger JD, Goring HH, Besemer J,
ods Enzymol. 303:77–83 Lomsadze A, Borodovsky M. 2000. Gene
6. Landgraf R, Xenarios I, Eisenberg D. mapping in the 20th and 21st centuries:
2001. Three-dimensional cluster analy- statistical methods, data analysis, and ex-
sis identifies interfaces and functional perimental design. Hum. Biol. 72:63–
residue clusters in proteins. J. Mol. Biol. 132
307:1487–502 16. Jansen RC, Nap J. 2001. Genetical ge-
7. Al-Lazikani B, Jung J, Xiang Z, Honig B. nomics: the added value from segregation.
2001. Protein structure prediction. Curr. Trends Genet. 17:388–91
Opin. Chem. Biol. 5:51–56 17. Chen RO, Altman RB. 1999. Automated
8. Altman RB, Raychaudhuri S. 2001. diagnosis of data-model conflicts using
Whole-genome expression analysis: chal- metadata. J. Am. Med. Inform. Assoc. 6:
lenges beyond clustering. Curr. Opin. 374–92
Struct. Biol. 11:340–47 18. Broder S, Venter JC. 2000. Sequencing
9. Vohradsky J. 2001. Neural model of ge- the entire genomes of free-living organ-
netic network. J. Biol. Chem. 6:6 isms: the foundation of pharmacology in
10. Reis BY, Butte AS, Kohane IS. 2001. the new millennium. Annu. Rev. Pharma-
Extracting knowledge from dynamics in col. Toxicol. 40:97–132
gene expression. J. Biomed. Inform. 34: 19. Bartlett J. 2001. Technology evaluation:
15–27 SAGE, Genzyme molecular oncology.
11. Jenssen TK, Laegreid A, Komorowski Curr. Opin. Mol. Ther. 3:85–96
J, Hovig E. 2001. A literature network of 20. Hu Y. 2001. An integrated approach
130 ALTMAN
KLEIN

for genome-wide gene expression analy- nology for clinical diagnostics. Mol. Di-
sis. Comput. Methods Programs Biomed. agn. 5:341–48
65:163–74 33. Hess P, Cooper D. 1999. Impact of phar-
21. Tucker CL, Gera JF, Uetz P. 2001. To- macogenomics on the clinical laboratory.
wards an understanding of complex pro- Mol. Diagn. 4:289–98
tein networks. Trends Cell Biol. 11:102–6 34. Meisel C, Roots I, Cascorbi I, Brinkmann
22. Cubitt AB, Heim R, Adams SR, Boyd U, Brockmoller J, et al. 2000. How to
AE, Gross LA, Tsien RY. 1995. Under- manage individualized drug therapy: ap-
standing, improving and using green flu- plication of pharmacogenetic knowledge
orescent proteins. Trends BioChem. Sci. of drug metabolism and transport. Clin.
20:448–55 Chem. Lab. Med. 38:869–76
23. Papac DI, Shahrokh Z. 2001. Mass spec- 35. Yan L, Otterness DM, Weinshilboum
trometry innovations in drug discovery RM. 1999. Human nicotinamide N-
and development. Pharm. Res. 18:131–45 methyltransferase pharmacogenetics:
24. Teichmann SA, Murzin AG, Chothia C. gene sequence analysis and promoter
2001. Determination of protein function, characterization. Pharmacogenetics 9:
evolution and interactions by structural 307–16
genomics. Curr. Opin. Struct. Biol. 11: 36. Glatt CE, DeYoung JA, Delgado S, Ser-
354–63 vice SK, Giacomini KM, et al. 2001.
25. Weber W. 1997. Pharmacogenetics. Ox- Screening a large reference sample to
ford, UK: Oxford Univ. Press identify very low frequency sequence
26. Evans WE, Relling M. 1999. Pharma- variants: comparisons between two genes.
cogenomics: translating functional ge- Nat. Genet. 27:435–38
nomics into rational therapeutics. Science 37. Kuehl P, Zhang J, Lin Y, Lamba J, As-
286:487–91 sem M, et al. 2001. Sequence diversity in
27. Rusnak JM, Kisabeth RM, Herbert DP, CYP3A promoters and characterization of
McNeil DM. 2001. Pharmacogenomics: a the genetic basis of polymorphic CYP3A5
clinician’s primer on emerging technolo- expression. Nat. Genet. 27:383–91
gies for improved patient care. Mayo Clin. 38. Israel E, Drazen JM, Liggett SB,
Proc. 76:299–309 Boushey HA, Cherniack RM, et al. 2001.
28. Meyer UA. 1991. Genotype or phenotype: Effect of polymorphism of the beta(2)-
the definition of a pharmacogenetic poly- adrenergic receptor on response to regular
morphism. Pharmacogenetics 1:66–67 use of albuterol in asthma. Int. Arch. Al-
29. McLeod HL, Evans WE. 2001. Pharma- lergy Immunol. 124:183–86
cogenomics: unlocking the human 39. Ewesuedo RB, Iyer L, Das S, Koenig A,
genome for better drug therapy. Annu. Mani S, et al. 2001. Phase I clinical and
Rev. Pharmacol. Toxicol. 41:101–21 pharmacogenetic study of weekly TAS-
30. Murphy MP. 2000. Current pharmacoge- 103 in patients with advanced cancer. J.
nomic approaches to clinical drug devel- Clin. Oncol. 19:2084–90
opment. Pharmacogenomics 1:115–23 40. Furuya H, Fernandez-Salguero P, Gre-
31. Murphy MP, Beaman ME, Clark LS, gory W, Taber H, Steward A, et al.
Cayouette M, Benson L, et al. 2000. 1995. Genetic polymorphism of CYP2C9
Prospective CYP2D6 genotyping as an and its effect on warfarin maintenance
exclusion criterion for enrollment of a dose requirement in patients undergoing
phase III clinical trial. Pharmacogenetics anticoagulation therapy. Pharmacogenet-
10:583–90 ics 5:389–92
32. Leushner J, Chiu NH. 2000. Automated 41. O’Brien SJ, Menotti-Raymond M, Mur-
mass spectrometry: a revolutionary tech- phy WJ, Nash WG, Wienberg J, et al.
 INFORMATICS FOR PHARMACOGENOMICS 131

1999. The promise of comparative geno- mental disorders. Clin. Genet. 57:253–
mics in mammals. Science 286:458–62, 66
479–81 54. Takai-Igarashi T, Nadaoka Y, Kaminuma
42. Clark MS. 1999. Comparative genomics: T. 1998. A database for cell signaling net-
the key to understanding the Human works. J. Comput. Biol. 5:747–54
Genome Project. Bioessays 21:121–30 55. Wingender E, Chen X, Hehl R, Karas
43. Bray MS, Boerwinkle E, Doris PA. 2001. H, Liebich I, et al. 2000. TRANSFAC:
High-throughput multiplex SNP genotyp- an integrated system for gene expression
ing with MALDI-TOF mass spectrome- regulation. Nucleic Acids Res. 28:316–19
try: practice, problems and promise. Hum. 56. Smith CM, Shindyalov IN, Veretnik S,
Mutat. 17:296–304 Gribskov M, Taylor SS, et al. 1997. The
44. Brookes AJ. 1999. The essence of SNPs. protein kinase resource. Trends BioChem.
Gene 234:177–86 Sci. 22:444–46
45. Schork NJ, Fallin D, Lanchbury JS. 57. Jain KK. 2000. Applications of biochip
2000. Single nucleotide polymorphisms and microarray systems in pharmacoge-
and the future of genetic epidemiology. nomics. Pharmacogenomics 1:289–307
Clin. Genet. 58:250–64 58. Scherf U, Ross DT, Waltham M, Smith
46. Judson R, Stephens JC. 2001. Notes from LH, Lee JK, et al. 2000. A gene expres-
the SNP vs. haplotype front. Pharmacoge- sion database for the molecular pharma-
nomics 2:7–10 cology of cancer. Nat. Genet. 24:236–44
47. Mann M, Hendrickson RC, Pandey A. 59. Slonim DK. 2001. Transcriptional profil-
2001. Analysis of proteins and pro- ing in cancer: the path to clinical pharma-
teomes by mass spectrometry. Annu. Rev. cogenomics. Pharmacogenomics 2:123–
BioChem. 70:437–73 36
48. Benson DA, Karsch-Mizrachi I, Lip- 60. Lindberg DA, Humphreys BL, McCray
man DJ, Ostell J, Rapp BA, Wheeler AT. 1993. The Unified Medical Language
DL. 2000. GenBank. Nucleic Acids Res. System. Methods Inf. Med. 32:281–91
28:15–8 61. Saier MH, Jr. 2000. A functional-
49. Cuticchia AJ. 2000. Future vision of the phylogenetic classification system for
GDB human genome database. Hum. Mu- transmembrane solute transporters. Mi-
tat. 15:62–67 croBiol. Mol. Biol. Rev. 64:354–411
50. Sherry ST, Ward MH, Kholodov M, 62. Nelson DR, Kamataki T, Waxman DJ,
Baker J, Phan L, et al. 2001. dbSNP: the Guengerich FP, Estabrook RW, et al.
NCBI database of genetic variation. Nu- 1993. The P450 superfamily: update on
cleic Acids Res. 29:308–11 new sequences, gene mapping, accession.
51. Sachidanandam R, Weissman D, DNA Cell Biol. 12:1–51
Schmidt SC, Kakol JM, Stein LD, et al. 63. Nelson DR, Koymans L, Kamataki T,
2001. A map of human genome sequence Stegeman JJ, Feyereisen R, et al. 1996.
variation containing 1.42 million sin- P450 superfamily: update on new se-
gle nucleotide polymorphisms. Nature quences, gene mapping, accession num-
409:928–33 bers. Pharmacogenetics 6:1–42
52. Hamosh A, Scott AF, Amberger J, 64. Harold E, Means W. 2001. XML in a Nut-
Valle D, McKusick VA. 2000. Online shell: A Desktop Quick Reference. Cam-
Mendelian Inheritance in Man (OMIM). bridge, MA: O’Reilly
Hum. Mutat. 15:57–61 65. Abernethy N, Wu J, Hewett M, Altman
53. Boyadijiev SA, Jabs EW. 2000. Online R. 1999. SOPHIA: a flexible, web-based
Mendelian Inheritance in Man (OMIM) knowledge server. IEEE Intell. Sys. Appl.
as a knowledgebase for human develop- 14:79–85
132 ALTMAN
KLEIN

66. Musen MA. 1998. Domain ontologies in 78. Kurella M, Hsiao LL, Yoshida T, Randall
software engineering: use of Protege with JD, Chow G, et al. 2001. DNA microarray
the EON architecture. Methods Inf. Med. analysis of complex biologic processes. J.
37:540–50 Am. Soc. Nephrol. 12:1072–78
67. Berman HM, Westbrook J, Feng Z, 79. Ideker T, Thorsson V, Ranish JA, Christ-
Gilliland G, Bhat TN, et al. 2000. The mas R, Buhler J, et al. 2001. Integrated
Protein Data Bank. Nucleic Acids Res. genomic and proteomic analyses of a sys-
28:235–42 tematically perturbed metabolic network.
68. Chung SY, Wong L, Blaschke C, An- Science 292:929–34
drade MA, Ouzounis C, et al. 1999. 80. Raychaudhuri S, Sutphin PD, Chang JT,
Kleisli: a new tool for data integration in Altman RB. 2001. Basic microarray anal-
biology. Trends Biotechnol. 17:351–55 ysis: grouping and feature reduction.
69. McEntyre J. 1998. Linking up with En- Trends Biotechnol. 19:189–93
trez. Trends Genet. 14:39–40 81. Srivastava M, Eidelman O, Pollard HB.
70. Allen J. 1995. Natural Language Under- 1999. Pharmacogenomics of the cystic fi-
standing. Redwood City, CA: Benjamin/ brosis transmembrane conductance regu-
Cummings lator (CFTR) and the cystic fibrosis drug
71. Yoshida M, Fukuda K, Takagi T. 2000. CPX using genome microarray analysis.
PNAD-CSS: a workbench for construct- Mol. Med. 5:753–67
ing a protein name abbreviation dictio- 82. Kawanishi Y, Tachikawa H, Suzuki T.
nary. Bioinformatics 16:169–75 2000. Pharmacogenomics and schizo-
72. Jenssen TK, Vinterbo S. 2000. A set- phrenia. Eur. J. Pharmacol. 410:227–41
covering approach to specific search for 83. Ekins S, de Groot MJ, Jones JP. 2001.
literature about human genes. Proc. AMIA Pharmacophore and three-dimensional
Symp. pp. 384–88. Philadelphia, PA: Han- quantitative structure activity relationship
ley & Belfus methods for modeling cytochrome p450
73. Marcotte EM, Xenarios I, Eisenberg active sites. Drug Metab. Dispos. 29:936–
D. 2001. Mining literature for protein- 44
protein interactions. Bioinformatics 17: 84. Blundell TL, Mizuguchi K. 2000. Struc-
359–63 tural genomics: an overview. Prog. Bio-
74. Rindflesch TC, Tanabe L, Weinstein JN, phys. Mol. Biol. 73:289–95
Hunter L. 2000. EDGAR: extraction of 85. Ewing TJ, Makino S, Skillman AG,
drugs, genes and relations from the bio- Kuntz ID. 2001. DOCK 4.0: search strate-
medical literature. Pac. Symp. Biocomput. gies for automated molecular docking of
pp. 517–28 flexible molecule databases. J. Comput.
75. Drawid A, Jansen R, Gerstein M. 2000. Aided Mol. Des. 15:411–28
Genome-wide analysis relating expres- 86. Pang YP, Kozikowski AP. 1994. Predic-
sion level with protein subcellular local- tion of the binding site of 1-benzyl-4-
ization. Trends Genet. 16:426–30 [(5,6-dimethoxy-1-indanon-2-yl)methyl]
76. Marcotte EM. 2000. Computational ge- piperidine in acetylcholinesterase by
netics: finding protein function by nonho- docking studies with the SYSDOC pro-
mology methods. Curr. Opin. Struct. Biol. gram. J. Comput. Aided Mol. Des. 8:
10:359–65 683–93
77. Masys DR, Welsh JB, Lynn Fink J, Grib- 87. Sun Y, Ewing TJ, Skillman AG, Kuntz
skov M, Klacansky I, Corbeil J. 2001. ID. 1998. CombiDOCK: structure-based
Use of keyword hierarchies to interpret combinatorial docking and library design.
gene expression patterns. Bioinformatics J. Comput. Aided Mol. Des. 12:597–604
17:319–26 88. Sanchez R, Sali A. 2000. Comparative
 INFORMATICS FOR PHARMACOGENOMICS 133

protein structure modeling. Introduction 98. Delcher AL, Kasif S, Fleischmann RD,
and practical examples with modeller. Peterson J, White O, Salzberg SL. 1999.
Methods Mol. Biol. 143:97–129 Alignment of whole genomes. Nucleic
89. Sunyaev S, Lathe W, Bork P III. 2001. Acids Res. 27:2369–76
Integration of genome data and protein 99. Karp PD, Riley M, Saier M, Paulsen IT,
structures: prediction of protein folds, Paley SM, Pellegrini-Toole A. 2000. The
protein interactions and “molecular phe- EcoCyc and MetaCyc databases. Nucleic
notypes” of single nucleotide polymor- Acids Res. 28:56–59
phisms. Curr. Opin. Struct. Biol. 11:125– 100. Wixon J, Kell D. 2000. The Kyoto ency-
30 clopedia of genes and genomes—KEGG.
90. Chasman D, Adams RM. 2001. Predic- Yeast 17:48–55
ting the functional consequences of non- 101. Werner T. 2001. Cluster analysis and pro-
synonymous single nucleotide polymor- moter modelling as bioinformatics tools
phisms: structure-based assessment of for the identification of target genes
amino acid variation. J. Mol. Biol. 307: from expression array data. Pharmacoge-
683–706 nomics 2:25–36
91. Claussen H, Buning C, Rarey M, 102. Zien A, Küffner R, Zimmer R, Lengauer
Lengauer T. 2001. FlexE: efficient molec- T. 2000. Analysis of gene expression data
ular docking considering protein structure with pathway scores. ISMB 2000:407–17
variations. J. Mol. Biol. 308:377–95 103. Nolan PM. 2000. Generation of mouse
92. Goldman BB, Wipke WT. 2000. QSD mutants as a tool for functional genomics.
quadratic shape descriptors. 2. Molecular Pharmacogenomics 1:243–55
docking using quadratic shape descriptors 104. Nadkarni PM, Marenco L, Chen R,
(QSDock). Proteins 38:79–94 Skoufos E, Shepherd G, Miller P. 1999.
93. Moont G, Gabb HA, Sternberg MJ. 1999. Organization of heterogeneous scientific
Use of pair potentials across protein inter- data using the EAV/CR representation. J.
faces in screening predicted docked com- Am. Med. Inform. Assoc. 6:478–93
plexes. Proteins 35:364–73 105. Collins FS, Brooks LD, Chakravarti A.
94. Morris GM, Goodsell DS, Huey R, Olson 1998. A DNA polymorphism discovery
AJ. 1996. Distributed automated docking resource for research on human genetic
of flexible ligands to proteins: parallel ap- variation. Genome Res. 8:1229–31
plications of AutoDock 2.4. J. Comput. 106. Sweeney L. 1998. Privacy and medical-
Aided Mol. Des. 10:293–304 records research. N. Engl. J. Med. 338:
95. Ritchie DW, Kemp GJ. 2000. Protein 1077; discussion-8
docking using spherical polar Fourier cor- 107. Wiederhold G, Bilello M, Sarathy V,
relations. Proteins 39:178–94 Qian X. 1996. A security mediator for
96. Sternberg MJ, Aloy P, Gabb HA, Jack- health care information. Proc. AMIA
son RM, Moont G, et al. 1998. A com- Annu. Fall Symp. 120–24
putational system for modelling flexible 108. Sweeney L, Nolan PM. 1997. Guaran-
protein-protein and protein-DNA dock- teeing anonymity when sharing medical
ing. Proc. Int. Conf. Intell. Syst. Mol. Biol. data, the Datafly System. Proc. AMIA
6:183–92 Annu. Fall Symp. 1:51–55. Philadelphia,
97. Dubchak I, Brudno M, Loots GG, PA: Hanley & Belfus
Pachter L, Mayor C, et al. 2000. Active 109. Malin B, Sweeney L. 2000. Determin-
conservation of noncoding sequences re- ing the identifiability of DNA database
vealed by three-way species comparisons. entries. Proc. AMIA Symp. pp. 537–41.
Genome Res. 10:1304–6 Philadelphia, PA: Hanley & Belfus

View publication stats