A System for Knowledge Management in Bioinformatics

Sudeshna Adak, Vishal S Batra, Deo N Bhardwaj, P V Kamesam,

Pankaj Kankar, Manish P Kurhekar, Biplav Srivastava

IBM India Research Laboratory
Block 1, IIT Campus, Hauz Khas
New Delhi 110016, India
{asudeshn,

bvishal, dbhardwa, pkamesam, kpankaj, kmanish, sbiplav}@in.ibm.com

ABSTRACT

Many exploration tools

semantically working together
across disciplines and across
heterogeneous sources

The emerging biochip technology has made it possible to simultaneously study expression (activity level) of thousands
of genes or proteins in a single experiment in the laboratory. However, in order to extract relevant biological knowledge from the biochip experimental data, it is critical not
only to analyze the experimental data, but also to crossreference and correlate these large volumes of data with information available in external biological databases accessible online. We address this problem in a comprehensive
system for knowledge management in bioinformatics called
e2e. To the biologist or biological applications, e2e exposes
a common semantic view of inter-relationship among biological concepts in the form of an XML representation called
eXpressML, while internally, it can use any data integration
solution to retrieve data and return results corresponding
to the semantic view. We have implemented an e2e prototype that enables a biologist to analyze her gene expression
data in GEML or from a public site like Stanford, and discover knowledge through operations like querying on relevant annotated data represented in eXpressML using pathways data from KEGG, publication data from Medline and
protein data from SWISS-PROT.

Client

Biochips Knowledge Management Applications

Gene Clustering
Disease Classification
Expression Pattern Visualization

Sequence Alignment
Medical Literature Summarization
Biochemical Pathway Recognition

Biochips
Query
Interface

Biologist Cares
Only Here

Semantic Biological Model (Includes Alias list, Synonyms, etc)

Biochips Enterprise Information System (Middleware choices)
Data Model:
Relational
XML
Implementation: DB2/SQL
XML or XML over Relational
Query Interface: SQL
Quilt/ Xquery
Common issues: Metadata and Wrapper management, Security

Wrappers

Sources

External

biochip

Other model(e.g. tuples)

Specialized database
Specialized language

pathway

literature

sequence

Figure 1: e2e Biochips Information System Framework Architecture.

1. INTRODUCTION
DNA based microarray technologies[14] have been used
extensively in generating the expression levels of all or most
of the genes of several organisms under a variety of experimental conditions. Specialized repositories and data warehousing projects are being built (e.g., NCBIs Gene Expression Omnibus1 (GEO), Stanford Microarray Database2 (SMD))
to store the vast quantities of data that are being generated
by the biochips. A biologist starts with analysis of the gene
expression data for insightful patterns among some clusters
of genes. Once a gene cluster is obtained, the main interest of a biologist lies in finding out the underlying biological mechanisms and functions causing these genes to be coexpressed and assign biological significance to this cluster.
The biological relations among these genes span multidisciplinary islands of biology. Downstream annotation involves
combining expression data with other sources of information
to improve the range and quality of conclusions that can be
drawn. However, related biomedical data[3] are numerous
and our focus is to develop an infrastructural framework
for building knowledge discovery tools for microarrays that
can leverage related but continuously updated diverse online
data.

Categories and Subject Descriptors

H.0 [Information Systems]: General

General Terms
Management, Performance

Keywords
Knowledge Management, Bioinformatics, Biochips
Lead contact. Author names appear in alphabetic order.
# Formerly with IBM.

Permission to make digital or hard copies of all or part of this work for
personal or classroom use is granted without fee provided that copies are
not made or distributed for profit or commercial advantage and that copies
bear this notice and the full citation on the first page. To copy otherwise, to
republish, to post on servers or to redistribute to lists, requires prior specific
permission and/or a fee.
CIKM02, November 49, 2002, McLean, Virginia, USA.
Copyright 2002 ACM 1-58113-492-4/02/0011 ...$5.00.

1
2

638

http://www.ncbi.nlm.nih.gov/geo/
http://genome-www4.Stanford.EDU/MicroArray/SMD/

To the stated end, we describe a comprehensive system

for knowledge management in bioinformatics called e2e (see
Figure 1) in which data generated by the biochip experiments can be cross-referenced and validated with additional
insights from related concepts. To the biologist or biological applications, e2e exposes a common semantic view of
inter-relationship among biological concepts in the form of
an XML representation called eXpressML. Internally, e2e
can use any data integration solution (like DiscoveryLink[9] ,
Kleisli[5] or natively XML-based) to retrieve data and return
results corresponding to the semantic view. We have implemented an e2e prototype that demonstrates our framework
by allowing a biologist to analyze her gene expression data
in GEML or from a public site like Stanford, and discover
knowledge through operations like querying on annotated
data in eXpressML, pathway scoring, text summarization,
etc. More details on e2e can be found in the extended paper[2].
Here is the layout of the paper: we start with a background of gene expression data and discuss current approaches
for data integration for bioinformatics. Next, we introduce
the e2e framework that maintains the users biological perspective. We discuss the different components of e2e - the
data integration middleware, eXpressML, a unified representation in XML for the complete annotation data necessary
to gain insight into gene expression patterns and the knowledge management (KM) applications. We conclude with our
contributions and future work.

2.

A Biological Environment
for Knowledge Discourse
and Relevant Tool Flows

Microarray Data

Clustered Genes

Novel
KM
(Knowledge Management)

Pathways
Analysis

Protein Structure
Analysis

Biomedical
Literature
Summarization

Apps
Sequence Data
Analysis

Semantic
Domain
Model

Chemical Compound
Analysis

Visualization

Q
U
E
R
Y

Data Integration Middleware

Figure 2: A Schemata of Flow between KM Analyses in e2e.

BACKGROUND

The developments in the area of microarrays and data

integration are shaping genomics today.

2.1 Gene Expression and Biochips

When a gene is expressed, the coded information contained in its DNA is first transcribed into messenger-RNA
and then translated into the proteins present and operating
in the cell. Changes in gene expression are associated with
almost all biological phenomena, including aging, onset and
progression of diseases, adaptive responses to the environment, and biochemical effects of drugs.
In order to realize the full potential of biochips, the main
challenges faced by the life sciences industry today are: (a)
Improvements in the core microarray technology to improve
the accuracy of gene expression measurement (b) Development of the full spectrum of specialized analytics and
(bio)informatics tools required for making (biological) knowledge discoveries from biochip data.

2.2 Integration of Heterogeneous Data

There are several stand-alone analysis tools today (e.g.
GeneSightT M from Biodiscovery, biotechnology solutions from
Spotfire, etc.) that detect gene expression patterns. However, since new genomic data is continuously produced and
made available online, a stand-alone tool, however sophisticated, will fail to provide the scalable, heterogeneous information integration infrastructure.
A variety of approaches have been developed for integrated access to heterogeneous data sources in genomics. In
the link-driven federation approach, the user can switch between sources using system-provided links in a hypermedia
environment. The user has to still interact with individual
sources; only the interaction is easier through convenient

639

links and not invoking the sources directly. SRS[7] is an example of this approach. Though the link driven approach
is very convenient for non-expert users and provides limited keyword search capability on the content of a source,
it does not scale well and has no across-source capabilities.
Another approach is that of view integration in which a virtual global schema is created in a common data model using
the descriptions of the individual sources so that the user
can declaratively pose queries on the common data model
that may span the content of multiple sources. The system
seamlessly and automatically figures out how data from the
different sources has to be retrieved [12]. A variation of view
integration is the warehousing approach where instantiation
of the global schema is created, i.e., all data of interest in
remote sources is locally replicated and maintained for predictable performance. Examples are IBMs DiscoveryLink[9]
and Kleisli[5] which provide powerful querying capabilities,
but fail to provide the in-depth analysis that are provided
by the point solutions.
A biologist wants to access only relevant data that she can
easily correlate in the pursuit of understanding the biochip
assay. Hence, what is needed is semantic integration in
which the user sees domain concepts like proteins and pathways while the infrastructural artifacts like source names
(SWISS-PROT, KEGG, etc) and attribute fields (protein id,
etc) are handled transparently by the user. A system related
to our definition of semantic integration is TAMBIS[8] where
a common ontology of about 1900 terms is constructed to
describe the concepts and relationships in molecular biology. Users interact with TAMBIS in the ontological realm
while the system internally maps them to source schemas using Kleisli[5] as its data integration middleware. However,
TAMBIS is not targeted towards microarrays and does not
provide the full spectrum of query/analytical capabilities
(breadth) that is needed in making (biological) knowledge
discoveries from biochip data.
E2E FRAMEWORK
We now discuss e2e in which semantic relationship among
biological concepts is represented in the XML representation
of eXpressML[1] and analytical KM tools can work from this

3.

abstraction. The biologist only needs to know about the

biological domain while the system will hide the peculiarities
of the sources involved to answer a domain query. As seen
in Figure 1, e2e envisages a two stage approach.
The underlying infrastructure for e2e is a view integration middleware (called Enterprise Information System to
emphasize the fact that it should be able to handle large
data sizes) which can retrieve either microarray experimental data or external information from publicly available biological data sources. In choosing a middleware, one has to
consider the issues of uniform data model, the query language to support and availability of source wrappers. For
example, with a relational data model and SQL query language, DiscoveryLink[9] is a middleware solution on commercial DB2T M database while Kleisli[5] uses a complex
value model of data and Collection Programming Language
(CPL), but either could be used within e2e. In the present
prototype, the data management is in XML and in-memory,
but will be migrated to a relational database in the next
version.
The semantic biological model provides the user with a
common biological context to view and manipulate related
data and issue XML queries in Quilt[13] through a query
interface. Finally, e2e envisages an application layer where
knowledge management tools are available for detecting gene
expression patterns and downstream annotation of these
patterns. For example, some tools that can be used are:
pathway visualization tools[11] for annotating gene clusters
with pathway information, text summarization tools[10] for
annotating gene clusters with biological function, and sequence alignment tools[4] for annotating gene clusters with
motifs/domains. Figure 2 shows a schematic flow between
KM analyses tools that a biologist may take in pursuit of
discovery. Note that the input for any KM tool is a group of
genes and (optionally) eXpressML while the output is some
insight about the group.

or by running KM tools on data from heterogeneous data

sources. Note that gene expression data itself is not part
of eXpressML. A related effort is MAGE-ML4 which represents useful annotations that describe the experimental
conditions and environments (array type, number of spots,
sample source, etc). However, MAGE-ML does not support annotation derived from heterogeneous external sources
while eXpressML extends to this as well.
Now, both GEML and eXpressML are available from e2e
and can be queried with an expressive XML query language.
The Biochips Query Interface (refer to Figure 1) select supports queries in Quilt XML query language [13] (specifically,
Kweelt5 implementation of Quilt). The query interface has
templates for a number of pre-canned queries and the user
can also pose any Quilt query which is valid.
Example: A query like list all regulatory pathways and
enzymes associated with genes that are similar in expression
to gene HXK1 (Hexokinase-1) can be formulated against
the eXpressML but it not possible with existing representations like GEML or MAGE-ML. This is because the query
involves determining the genes in the same cluster (group)
as gene HXK1 and finding the pathways and enzymes associated with the resulting gene list, which are representable
in eXpressML.

3.2 KM Layer
The KM layer consists of two types of applications: (a)
Tools for detecting gene expression patterns by supporting
clustering, classification, and visualization of biochip experimental data, and (b) Downstream annotation tools combining expression data with other sources of information to
improve the range and quality of conclusions that can be
drawn. Below, we describe some of the implemented frontend tools in detail but note that new tools can be built that
have as input a group of genes and optionally, subset of data
represented in eXpressML.

3.1 Integration in e2e

4. KM APPLICATIONS

e2e works on two types of data - the gene expression data

from microarray experiments and annotations of gene expression as well as relevant distributed data necessary to
gain insight into gene expression patterns. The annotations
are semantically arranged in the XML representation of eXpressML[1].
For gene expression, we adopted Rosetta Inpharmatics
Gene Expression Markup Language, GEMLT M 3 , which has
been accepted relatively widely by the industry as a uniform syntax for storing and exchanging gene expression data
from multiple biochip experiments. For annotations, we
developed the eXpressML representation keeping following
into consideration: (a) The semi-structured nature of XML
makes it the appropriate language for unified view of annotations as it guarantees flexibility and scalability in the data
model for future extensions. (b) The common view should
allow querying, modeling, and browsing of complex annotations. (c) The unified model should arrange the annotation
information in a compact hierarchy but reflect the relationship among the biological data items and facilitate complex
queries.
Related data includes DNA, protein, 3D structure, genetic
maps of disease, biochemical pathways, enzyme, keywords,
medical citation information, and is obtained either directly
3

4.1 Microarray Analysis

The first tier of Microarray data analysis typically involves
clustering or classification of the microarray data. In clustering (or cluster analysis), genes with similar expression
patterns are grouped together. Then, it is the gene cluster
rather than the individual genes that get associated with biological functions (e.g., DNA repair, galactose metabolism).
For example, hierarchical clustering[6] has been used to determine the functions of gene clusters in regulating cell-cycle
in yeast.
e2e provides a platform for integrating algorithms made
available through third-party vendors or academic researchers
seamlessly as long as they provide following basic information: (a) Any initialization parameters and the format of input gene expression data (tabular or XML). (b) The format
of output result. (c) If the algorithm supports visualization,
a handle of the input and output panels.
We have implemented hierarchical clustering and K-means
clustering in the e2e prototype.

4.2 Text Summarization

4
5

http://www.geml.org

640

http://www.mged.org/Workgroups/MAGE/mage.html
http://db.cis.upenn.edu/Kweelt/

matics. Biologists who have used the e2e prototype value

the ability it provides to cross-relate concepts and analytics from different areas. However, they want to run it with
larger expression data (1000s of genes), something for which
the current e2e prototype is slow due to the in-memory storage of XML.
We are looking at extending e2e along various directions:
(a) Address middleware issues of effective query decomposition and scalability in the presence of domain knowledge
of biology[15] and large data (through available database
technologies). (b) Extend the range of annotations and the
types of related data. (c) Improve query interface to allow
the biologist to issue natural language queries. (d) Improve
retrieval of unstructured data along with issues like change
detection and caching of results.

The biomedical literature databases are rich source of information from various disciplines of biomedical sciences.
Text mining of these databases can be used to augment, confirm, or discover biologically significant information for gene
clusters spanning different biological domains. The main
challenges in handling biomedical citations are: (1) Querying on even a small cluster of genes retrieves tens of thousands of documents. (2) Use of multiple names and conventions in referring to genes makes it difficult to cross-reference
documents with gene names. (3) Non-uniform nomenclature and language usage for same biological concepts make
it difficult for text mining of the citations retrieved. (4)
Highly complex and parallel interrelations among biological
processes across multiple biological domains.
We have developed a specialized text-mining system called
MedMeSH summarizer [10] that provides a summary of the
citations pertaining to a group of genes in a given cluster. The MedMeSH summarizer system uses PubMed as
the literature database and provides an automated document extraction and summarization solution PubMed, the
most widely used biomedical literature database has more
than 11 million citations (since 1960) and about 30,000 new
citations are added each month. The user is required to provide only a list of genes (gene cluster) as input. The output
is a summary of the documents, which shows the most important MeSH terms which describe the whole cluster and
produces summaries across all biological domains.

6.
REFERENCES
[1] Adak, S., Srivastava, B., Kankar, P., and Kurhekar, M. 2002.
A Common Data Representation for Organizing and
Managing Annotations of Biochip Expression Data. IBM
Research Report RI02017. Available at
http://domino.watson.ibm.com/library/CyberDig.nsf/Home
[2] Adak, S., Batra, V., Bhardwaj, D., Kamesam, P., Kankar,
P., Kurhekar, and Srivastava, B. 2002. Bioinformatics for
Microarrays. IBM Research Report RI02016. Available at
http://domino.watson.ibm.com/library/CyberDig.nsf/Home
[3] Baxevanis, A. 2001. The Molecular Biology Database
Collection: an updated compilation of biological database
resources. Numcleic Acids Research, Vol. 29, No. 1.
[4] Brazma, A., Jonassen, I., Vilo, J., and Ukkonen, E. (1998).
Predicting gene regulatory elements in silico on a genomic
scale. Genome Research, 8:1202-1215.
[5] Buneman, P., Davidson, S. Hart, K., Overton, C., and
Wong, L. (1995). A Data Transformation System for
Biological Data Sources. Proc. VLDB, pp 158169.
[6] Eisen, M., Spellman, P., Brown, P., and Botstein, D. (1998).
Cluster analysis and display of genome-wide expression
patterns. Proc. Natl Acad Sci USA, 95:14863-14868, 1998.
[7] Etzold, T., and Argos, P. (1993). SRS: An Indexing and
Retrieval Tool for Flat File Data Libraries. Computer
Application of Biosciences, 9:49-57.
[8] Goble, C., Stevens, R., Ng, G., Bechhofer, S., Paton, N.,
Baker, P., Peim, M., and Brass, A. (2001). Transparent
Access to Multiple Bioinformatics Information Sources. IBM
Systems Journal, Vol. 40, No.2, pp 532-551.
[9] Haas, L., Schwarz, P., Kodali, P., Kotlar, E., Rice, J., and
Swope, W. (2001). DiscoveryLink: A system for integrated
access to life sciences data sources. IBM Systems Journal,
Volume 40, Number 2, 2001.
[10] Kankar, P., Adak, S., Sarkar, A., Murari, K. and Sharma,
G. (2002). MedMeSH Summarizer: Text Mining for Gene
Clusters. In Proc. of the SIAM Conf. in Data Mining.
[11] Kurhekar, M., Adak, S., Jhunjhunwala, S., and
Raghupathy, K. (2002). Genome-wide pathway analysis and
visualization using gene expression data. In Proc. of the
Pacific Symposium of Biocomputing.
[12] Levy, A. 1998. Combining Artificial Intelligence and
Databases for Data Integration. At http://citeseer.nj.nec.com
[13] Robie, D., Chamberlin, D. and Florescu, D. (2001). Quilt:
an XML Query Language.
http://www.almaden.ibm.com/cs/people/chamberlin/quilt euro.html
[14] Shalon, D., Smith, S. and Brown, P. (1996). A DNA
microarray system for analyzing complex DNA samples using
two-color fluorescent probe hybridization. Genome Research,
6:639-645.
[15] Srivastava, B. 2002. Using Planning for Query
Decomposition in Bioinformatics Sixth Intl. Conf. on AI
Planning & Scheduling (AIPS-02) Workshop on Is There
Life Beyond Operator Sequencing? Exploring Real World
Planning.

4.3 Pathways Scoring

At the cellular level, organisms function through intricate
networks of chemical reactions (metabolic pathways) and
interacting molecules (regulatory pathways). Annotation of
microarray data with pathway information can help in understanding the functions and roles of the proteins involved
in various cellular processes. The pathway scoring systems
serve as an important tool for interpreting the large amount
of data from microarrays, in assessing the behavior of pathways at different cell stages or the effect of stimuli on cellular
processes.
We have implemented algorithms [11] which use gene expression data and putative metabolic and regulatory pathways database of KEGG. The outputs are: pathway scores
which quantify activity, coregulation, and cascade effects in pathways as measured by the gene expression levels
from the microarray experimental data, and pathway animated visuals which show the effects on individual pathways.

4.4 Protein Sequence Analysis

By comparing the complete genome of one organism to another, it is clear that certain genes have been conserved since
evolutionary divergence from a common ancestor. Genes
can be found in the different organisms, with identical functions and/or protein motifs. The way to do this is by sequence analysis. The sequence analyser has a host of sequence similarity tools including BLAST and FASTA and
uses the SWISS-Prot database.

5.

CONCLUSION AND FUTURE WORK

In this paper, we presented a comprehensive bioinformatics KM framework called e2e which provides a uniform window to biochip data and related annotations. We demonstrated an e2e prototype that gives an early glimpse of the
wide potential of an integrated KM solution for bioinfor-

641