What is Bioinformatics? A Proposed Definition and Overview of the ...

347What is Bioinformatics?Fig. 1 Plot showing the growth of scientific publications in bioinformatics between 1973 and 2000. The histogram bars(left vertical axis) counts the total number of scientific articles relating to bioinformatics, and the black line (right verticalaxis) gives the percentage of the annual total of articles relating to bioinformatics. The data are taken from PubMed.This needs more than just a simple textbasedsearch, and programs such as FASTA[8] and PSI-BLAST [9] must consider whatconstitutes a biologically significant match.Development of such resources dictatesexpertise in computational theory, as wellas a thorough understanding of biology.The third aim is to use these tools to analysethe data and interpret the results in abiologically meaningful manner. Traditionally,biological studies examined individualsystems in detail, and frequently comparedthem with a few that are related. Inbioinformatics, we can now conduct globalanalyses of all the available data with theaim of uncovering common principles thatapply across many systems and highlightnovel features.In this review, we provide a systematicdefinition of bioinformatics as shown inBox 1. We focus on the first and third aimsjust described, with particular reference tothe keywords: information, informatics,organisation, understanding, large-scaleand practical applications. Specifically, wediscuss the range of data that are currentlybeing examined, the databases into whichthey are organised, the types of analysesthat are being conducted using transcriptionregulatory systems as an example, andfinally some of the major practical applicationsof bioinformatics.2. “…the INFORMATIONassociated with theseMolecules…”Table 1 lists the types of data that areanalysed in bioinformatics and the range oftopics that we consider to fall within thefield. Here we take a broad view and includesubjects that may not normally belisted. We also give approximate valuesdescribing the sizes of data being discussed.We start with an overview of the sourcesof information. Most bioinformatics analysesfocus on three primary sources of data:DNA or protein sequences, macromolecularstructures and the results of functionalgenomics experiments. Raw DNA sequencesare strings of the four base-letterscomprising genes, each typically 1,000 baseslong. The GenBank [2] repository ofnucleic acid sequences currently holds atotal of 12.5 billion bases in 11.5 millionentries (all database figures as of April2001).At the next level are protein sequencescomprising strings of 20 amino acidletters.At present there are about 400,000known protein sequences [3], with a typicalbacterial protein containing approximately300 amino acids. Macromolecular structuraldata represents a more complex formof information. There are currently 15,000entries in the Protein Data Bank, PDB[6, 7], containing atomic structures of proteins,DNA and RNA solved by x-raycrystallography and NMR. A typical PDBfile for a medium-sized protein contains thexyz-coordinates of approximately 2,000atoms.Scientific euphoria has recently centredon whole genome sequencing. As with theraw DNA sequences, genomes consist ofstrings of base-letters, ranging from 1.6million bases in Haemophilus influenzae[10] to 3 billion in humans [11, 12]. TheEntrez database [13] currently has completesequences for nearly 300 archaeal,bacterial and eukaryotic organisms. Inaddition to producing the raw nucleotidesequence, a lot of work is involved inprocessing this data. An important aspectof complete genomes is the distinctionbetween coding regions and non-codingregions -‘junk’ repetitive sequences makingup the bulk of base sequences especially ineukaryotes. Within the coding regions,genes are annotated with their translatedprotein sequence, and often with theircellular function.Bioinformatics – a Definition 1(Molecular) bio – informatics: bioinformaticsis conceptualising biology interms of molecules (in the sense of Physicalchemistry) and applying “informaticstechniques” (derived from disciplinessuch as applied maths, computerscience and statistics) to understand andorganise the information associatedwith these molecules, on a large scale. Inshort, bioinformatics is a managementinformation system for molecular biologyand has many practical applications.1As submitted to the Oxford EnglishDictionary.Method Inform Med 4/2001

348Luscombe, Greenbaum, GersteinTable 1 Sources of data used in bioinformatics, the quantity of each type of data that is currently (April 2001) available,and bioinformatics subject areas that utilize this data.tities of data when experiments are conductedfor larger organisms and at moretime-points.Further genomic-scale data includebiochemical information on metabolicpathways, regulatory networks, proteinproteininteraction data from two-hybridexperiments, and systematic knockouts ofindividual genes to test the viability of anorganism.What is apparent from this list is thediversity in the size and complexity of differentdatasets. There are invariably moresequence-based data than others becauseof the relative ease with which they can beproduced.This is partly related to the greatercomplexity and information-content ofindividual structures or gene expressionexperiments compared to individual sequences.While more biological informationcan be derived from a single structurethan a protein sequence, the lack of depthin the latter is compensated by analysinglarger quantities of data.Method Inform Med 4/2001More recent sources of data have beenfrom functional genomics experiments, ofwhich the most common are gene expressionstudies.We can now determine expressionlevels of almost every gene in a givencell on a whole-genome level, howeverthere is currently no central repository forthis data and public availability is limited.These experiments measure the amount ofmRNA that is produced by the cell [14-18]under different environmental conditions,different stages of the cell cycle and differentcell types in multi-cellular organisms.Much of the effort has so far focused on theyeast [19-24] and human genomes [25, 26].One of the largest dataset for yeast hasmade approximately 20 time-point measurementsfor 6,000 genes [19]. However,there is potential for much greater quan-3. “… ORGANISE the Informationon a LARGE SCALE…”3.1 Redundancy and Multiplicityof DataA concept that underpins most researchmethods in bioinformatics is that much ofthe data can be grouped together based onbiologically meaningful similarities. Forexample, sequence segments are oftenrepeated at different positions of genomicDNA [27]. Genes can be clustered intothose with particular functions (eg enzymaticactions) or according to the metabolicpathway to which they belong [28],although here, single genes may actuallypossess several functions [29]. Goingfurther, distinct proteins frequently havecomparable sequences – organisms oftenhave multiple copies of a particular genethrough duplication and different specieshave equivalent or similar proteins thatwere inherited when they diverged fromeach other in evolution. At a structurallevel, we predict there to be a finite numberof different tertiary structures – estimatesrange between 1,000 and 10,000 folds[30, 31] – and proteins adopt equivalentstructures even when they differ greatly insequence [32]. As a result, although thenumber of structures in the PDB hasincreased exponentially, the rate of discoveryof novel folds has actually decreased.There are common terms to describe therelationship between pairs of proteins orthe genes from which they are derived:analogous proteins have related folds, butunrelated sequences, while homologousproteins are both sequentially and structurallysimilar. The two categories can sometimesbe difficult to distinguish especially ifthe relationship between the two proteinsis remote [33, 34]. Among homologues, it isuseful to distinguish between orthologues,proteins in different species that have evolv-

349What is Bioinformatics?ed from a common ancestral gene, andparalogues, proteins that are related bygene duplication within a genome [35].Normally, orthologues retain the samefunction while paralogues evolve distinct,but related functions [36].An important concept that arises fromthese observations is that of a finite “partslist” for different organisms [37-39]: aninventory of proteins contained within anorganism, arranged according to differentproperties such as gene sequence, proteinfold or function. Taking protein folds as anexample, we mentioned that with a fewexceptions, the tertiary structures of proteinsadopt one of a limited repertoireof folds. As the number of different foldfamilies is considerably smaller than thenumber of genes, categorising the proteinsby fold provides a substantial simplificationof the contents of a genome. Similar simplificationscan be provided by other attributessuch as protein function. As such, weexpect this notion of a finite parts list tobecome increasingly common in futuregenomic analyses.Clearly, an essential aspect of managingthis large volume of data lies in developingmethods for assessing similarities betweendifferent biomolecules and identifyingthose that are related. There are well-documentedclassifications for all of the maintypes of data we described earlier. Althoughdetailed descriptions of these classificationsystems are beyond the scope ofthe current review, they are of great importanceas they ease comparisons betweengenomes and their products. Links to themajor databases are available from oursupplementary website.3.2 Data IntegrationThe most profitable research in bioinformaticsoften results from integrating multiplesources of data [40]. For instance, the3D coordinates of a protein are more usefulif combined with data about the protein’sfunction, occurrence in different genomes,and interactions with other molecules. Inthis way, individual pieces of informationare put in context with respect to otherdata. Unfortunately, it is not alwaysstraightforward to access and crossreferencethese sources of information becauseof differences in nomenclature andfile formats.At a basic level, this problem is frequentlyaddressed by providing externallinks to other databases. For example inPDBsum, web-pages for individual structuresdirect the user towards correspondingentries in the PDB, NDB, CATH, SCOPand SWISS-PROT databases. At a moreadvanced level, there have been efforts tointegrate access across several data sources.One is the Sequence Retrieval System, SRS[41], which allows flat-file databases to beindexed to each other; this allows the userto retrieve, link and access entries fromnucleic acid, protein sequence, proteinmotif, protein structure and bibliographicdatabases. Another is the Entrez facility[42], which provides similar gateways toDNA and protein sequences, genomemapping data, 3D macromolecular structuresand the PubMed bibliographic database[43].A search for a particular gene in eitherdatabase will allow smooth transitions tothe genome it comes from, the proteinsequence it encodes, its structure, bibliographicreference and equivalent entries forall related genes. In our own group, we havedeveloped the SPINE [44] and PartsList[39] web resources; these databases integratemany types of experimental data andorganise them using the concept of thefinite “parts list” we described above.4. “…UNDERSTAND andOrganise the Information…”Having examined the data, we can discussthe types of analyses that are conducted.Asshown in Table 1, the broad subject areas inbioinformatics can be separated accordingto the type of information that is used. Forraw DNA sequences, investigations involveseparating coding and non-coding regions,and identification of introns, exons andpromoter regions for annotating genomicDNA [45, 46]. For protein sequences, analysesinclude developing algorithms forsequence comparisons [47], methods forproducing multiple sequence alignments[48], and searching for functional domainsfrom conserved sequence motifs in suchalignments. Investigations of structuraldata include prediction of secondary andtertiary protein structures, producingmethods for 3D structural alignments [49,50], examining protein geometries usingdistance and angular measurements, calculationsof surface and volume shapes andanalysis of protein interactions with othersubunits, DNA, RNA and smaller molecules.These studies have lead to molecularsimulation topics in which structural dataare used to calculate the energetics involvedin stabilising macromolecular structures,simulating movements within macromolecules,and computing the energiesinvolved in molecular docking. The increasingavailability of annotated genomicsequences has resulted in the introductionof computational genomics and proteomics– large-scale analyses of complete genomesand the proteins that they encode. Researchincludes characterisation of proteincontent and metabolic pathways betweendifferent genomes, identification of interactingproteins, assignment and prediction ofgene products, and large-scale analyses ofgene expression levels. Some of these researchtopics will be demonstrated in ourexample analysis of transcription regulatorysystems.Other subject areas we have included inTable 1 are: development of digital librariesfor automated bibliographical searches,knowledge bases of biological informationfrom the literature, DNA analysis methodsin forensics, prediction of nucleic acid structures,metabolic pathway simulations, andlinkage analysis – linking specific genes todifferent disease traits.In addition to finding relationships betweendifferent proteins, much of bioinformaticsinvolves the analysis of one typeof data to infer and understand the observationsfor another type of data. An exampleis the use of sequence and structuraldata to predict the secondary and tertiarystructures of new protein sequences [51].These methods, especially the former, areoften based on statistical rules derivedfrom structures, such as the propensity forcertain amino acid sequences to produceMethod Inform Med 4/2001

350Luscombe, Greenbaum, GersteinParadigm shifts during the past couple of decades have taken much of biology away from thelaboratory bench and have allowed the integration of other scientific disciplines, specificallycomputing. The result is an expansion of biological research in breadth and depth. The vertical axisdemonstrates how bioinformatics can aid rational drug design with minimal work in the wet lab.Starting with a single gene sequence, we can determine with strong certainty, the proteinsequence. From there, we can determine the structure using structure prediction techniques. Withgeometry calculations, we can further resolve the protein’s surface and through molecularsimulation determine the force fields surrounding the molecule. Finally docking algorithms canprovide predictions of the ligands that will bind on the protein surface, thus paving the way for thedesign of a drug specific to that molecule. The horizontal axis shows how the influx of biologicaldata and advances in computer technology have broadened the scope of biology. Initially with a pairof proteins, we can make comparisons between the between sequences and structures ofevolutionary related proteins. With more data, algorithms for multiple alignments of severalproteins become necessary. Using multiple sequences, we can also create phylogenetic trees totrace the evolutionary development of the proteins in question. Finally, with the deluge of data wecurrently face, we need to construct large databases to store, view and deconstruct theinformation. Alignments now become more complex, requiring sophisticated scoring schemes andthere is enough data to compile a genome census – a genomic equivalent of a population census –providing comprehensive statistical accounting of protein features in genomes.Fig. 2Organizing and understanding biological datadifferent secondary structural elements.Another example is the use of structuraldata to understand a protein’s function;here studies have investigated the relationshipdifferent protein folds and theirfunctions [52, 53] and analysed similaritiesbetween different binding sites in the absenceof homology [54]. Combined withsimilarity measurements, these studies provideus with an understanding of how muchbiological information can be accuratelyMethod Inform Med 4/2001transferred between homologous proteins[55].4.1 The Bioinformatics SpectrumFig. 2 summarises the main points weraised in our discussions of organisingand understanding biological data – thedevelopment of bioinformatics techniqueshas allowed an expansion of biologicalanalysis in two dimension, depth andbreadth. The first is represented by thevertical axis in the figure and outlines apossible approach to the rational drugdesign process. The aim is to take a singlegene and follow through an analysis thatmaximises our understanding of theprotein it encodes. Starting with a genesequence, we can determine the proteinsequence with strong certainty. From there,prediction algorithms can be used to calcu-

351What is Bioinformatics?late the structure adopted by the protein.Geometry calculations can define theshape of the protein’s surface and molecularsimulations can determine the forcefields surrounding the molecule. Finally,using docking algorithms, one couldidentify or design ligands that may bindthe protein, paving the way for designing adrug that specifically alters the protein’sfunction. In practise, the intermediate stepsare still difficult to achieve accurately, andthey are best combined with experimentalmethods to obtain some of the data, forexample characterising the structure of theprotein of interest.The aim of the second dimension, thebreadth in biological analysis, is to comparea gene or gene product with others. Initially,simple algorithms can be used tocompare the sequences and structures of apair of related proteins. With a larger numberof proteins, improved algorithms can beused to produce multiple alignments, andextract sequence patterns or structuraltemplates that define a family of proteins.Using this data, it is also possible to constructphylogenetic trees to trace the evolutionarypath of proteins. Finally, with evenmore data, the information must be storedin large-scale databases. Comparisonsbecome more complex, requiring multiplescoring schemes, and we are able to conductgenomic scale censuses that providecomprehensive statistical accounts ofprotein features, such as the abundance ofparticular structures or functions in differentgenomes. It also allows us to buildphylogenetic trees that trace the evolutionof whole organisms.5. “… applying INFORMATICSTECHNIQUES…”The distinct subject areas we mentionrequire different types of informatics techniques.Briefly, for data organisation, thefirst biological databases were simple flatfiles. However with the increasing amountof information, relational databasemethods with Web-page interfaces havebecome increasingly popular. In sequenceanalysis, techniques include string comparisonmethods such as text search and onedimensionalalignment algorithms. Motifand pattern identification for multiplesequences depend on machine learning,clustering and data-mining techniques. 3Dstructural analysis techniques include Euclideangeometry calculations combinedwith basic application of physical chemistry,graphical representations of surfacesand volumes, and structural comparisonand 3D matching methods. For molecularsimulations, Newtonian mechanics, quantummechanics, molecular mechanics andelectrostatic calculations are applied. Inmany of these areas, the computationalmethods must be combined with goodstatistical analyses in order to provide anobjective measure for the significance ofthe results.6. Transcription Regulation –a Case Study in BioinformaticsDNA-binding proteins have a central rolein all aspects of genetic activity within anorganism, participating in processes such astranscription, packaging, rearrangement,replication and repair. In this section, wefocus on the studies that have contributedto our understanding of transcriptionregulation in different organisms. Throughthis example, we demonstrate how bioinformaticshas been used to increase ourknowledge of biological systems and alsoillustrate the practical applications of thedifferent subject areas that were brieflyoutlined earlier. We start by consideringstructural analyses of how DNA-bindingproteins recognise particular base sequences.Later, we review several genomicstudies that have characterised the natureof transcription factors in different organisms,and the methods that have been usedto identify regulatory binding sites in theupstream regions. Finally, we provide anoverview of gene expression analyses thathave been recently conducted and suggestfuture uses of transcription regulatory analysesto rationalise the observations madein gene expression experiments. All theresults that we describe have been foundthrough computational studies.6.1 Structural StudiesAs of April 2001, there were 379 structuresof protein-DNA complexes in the PDB.Analyses of these structures have providedvaluable insight into the stereochemicalprinciples of binding, including how particularbase sequences are recognizedand how the DNA structure is quite oftenmodified on binding.A structural taxonomy of DNA-bindingproteins, similar to that presented in SCOPand CATH, was first proposed by Harrison[56] and periodically updated to accommodatenew structures as they are solved[57]. The classification consists of a two-tiersystem: the first level collects proteins intoeight groups that share gross structuralfeatures for DNA-binding, and the secondcomprises 54 families of proteins that arestructurally homologous to each other.Assembly of such a system simplifies thecomparison of different binding methods; ithighlights the diversity of protein-DNAcomplex geometries found in nature, butalso underlines the importance of interactionsbetween -helices and the DNAmajor groove, the main mode of binding inover half the protein families. While thenumber of structures represented in thePDB does not necessarily reflect the relativeimportance of the different proteins inthe cell, it is clear that helix-turn-helix,zinc-coordinating and leucine zipper motifsare used repeatedly.These provide compactframeworks to present the -helix on thesurfaces of structurally diverse proteins. Ata gross level, it is possible to highlight thedifferences between transcription factordomains that “just” bind DNA and thoseinvolved in catalysis [58]. Although thereare exceptions, the former typicallyapproach the DNA from a single face andslot into the grooves to interact with baseedges. The latter commonly envelope thesubstrate, using complex networks ofsecondary structures and loops.Focusing on proteins with -helices, thestructures show many variations, both inamino acid sequences and detailed geometry.They have clearly evolved independentlyin accordance with the requirementsof the context in which they are found.While achieving a close fit between theMethod Inform Med 4/2001

352Luscombe, Greenbaum, Gerstein-helix and major groove, there is enoughflexibility to allow both the protein andDNA to adopt distinct conformations.However, several studies that analysed thebinding geometries of -helices demonstratedthat most adopt fairly uniform conformationsregardless of protein family.They are commonly inserted in the majorgroove sideways, with their lengthwise axisroughly parallel to the slope outlined bythe DNA backbone. Most start with theN-terminus in the groove and extend out,completing two to three turns withincontacting distance of the nucleic acid [59,60].Given the similar binding orientations, itis surprising to find that the interactionsbetween each amino acid position alongthe -helices and nucleotides on the DNAvary considerably between different proteinfamilies. However, by classifying theamino acids according to the sizes of theirside chains, we are able to rationalise thedifferent interactions patterns. The rules ofinteractions are based on the simple premisethat for a given residue position on-helices in similar conformations, smallamino acids interact with nucleotides thatare close in distance and large amino acidswith those that are further [60, 61]. Equi-valentstudies for binding by other structuralmotifs, like -hairpins, have also been conducted[62]. When considering theseinteractions, it is important to rememberthat different regions of the protein surfacealso provide interfaces with the DNA.This brings us to look at the atomic levelinteractions between individual aminoacid-base pairs. Such analyses are based onthe premise that a significant proportion ofspecific DNA-binding could be rationalisedby a universal code of recognition betweenamino acids and bases, ie whether certainprotein residues preferably interact withparticular nucleotides regardless of thetype of protein-DNA complex [63]. Studieshave considered hydrogen bonds, van derWaals contacts and water-mediated bonds[64-66]. Results showed that about 2/3 of allinteractions are with the DNA backboneand that their main role is one ofsequence-independent stabilisation. Incontrast, interactions with bases displaysome strong preferences, including theMethod Inform Med 4/2001interactions of arginine or lysine withguanine, asparagine or glutamine withadenine and threonine with thymine. Suchpreferences were explained through examinationof the stereochemistry of the aminoacid side chains and base edges. Alsohighlighted were more complex types ofinteractions where single amino acidscontact more than one base-step simultaneously,thus recognising a short DNAsequence. These results suggested thatuniversal specificity, one that is observedacross all protein-DNA complexes, indeedexists. However, many interactions that arenormally considered to be non-specific,such as those with the DNA backbone, canalso provide specificity depending on thecontext in which they are made.Armed with an understanding ofprotein structure, DNA-binding motifs andside chain stereochemistry, a major applicationhas been the prediction of bindingeither by proteins known to contain a particularmotif, or those with structures solvedin the uncomplexed form. Most commonare predictions for -helix-major grooveinteractions – given the amino acid sequence,what DNA sequence would itrecognise [61, 67]. In a different approach,molecular simulation techniques have beenused to dock whole proteins and DNAs onthe basis of force-field calculations aroundthe two molecules [68, 69].The reason that both methods havebeen met with limited success is becauseeven for apparently simple cases like -helix-binding, there are many other factorsthat must be considered. Comparisonsbetween bound and unbound nucleic acidstructures show that DNA-bending is acommon feature of complexes formed withtranscription factors [58, 70].This and otherfactors such as electrostatic and cationmediatedinteractions assist indirectrecognition of the nucleotide sequence,although they are not well understood yet.Therefore, it is now clear that detailed rulesfor specific DNA-binding will be familyspecific, but with underlying trends such asthe arginine-guanine interactions.6.2 Genomic StudiesDue to the wealth of biochemical data thatare available, genomic studies in bioinformaticshave concentrated on modelorganisms, and the analysis of regulatorysystems has been no exception. Identificationof transcription factors in genomes invariablydepends on similarity search strategies,which assume a functional and evolutionaryrelationship between homologousproteins. In E. coli, studies have so farestimated a total of 300 to 500 transcriptionregulators [71] and PEDANT [72], a databaseof automatically assigned gene functions,shows that typically 2-3% of prokaryoticand 6-7% of eukaryotic genomescomprise DNA-binding proteins.As assignmentswere only complete for 40-60% ofgenomes as of August 2000, these figuresmost likely underestimate the actual number.Nonetheless, they already represent alarge quantity of proteins and it is clear thatthere are more transcription regulatorsin eukaryotes than other species. This isunsurprising, considering the organismshave developed a relatively sophisticatedtranscription mechanism.From the conclusions of the structuralstudies, the best strategy for characterisingDNA-binding of the putative transcriptionfactors in each genome is to group themby homology and to analyse the individualfamilies. Such classifications are providedin the secondary sequence databasesdescribed earlier and also those thatspecialise in regulatory proteins such asRegulonDB [73] and TRANSFAC [74].Of even greater use is the provision ofstructural assignments to the proteins;given a transcription factor, it is helpful toknow the structural motif that it uses forbinding, therefore providing us with abetter understanding of how it recognisesthe target sequence. Structural genomicsthrough bioinformatics assigns structuresto the protein products of genomes bydemonstrating similarity to proteins ofknown structure [75]. These studies haveshown that prokaryotic transcription factorsmost frequently contain helix-turnhelixmotifs [71, 76] and eukaryotic factorscontain homeodomain type helix-turnhelix,zinc finger or leucine zipper motifs.

353What is Bioinformatics?From the protein classifications in eachgenome, it is clear that different types ofregulatory proteins differ in abundance andfamilies significantly differ in size. A studyby Huynen and van Nimwegen [77] hasshown that members of a single family havesimilar functions, but as the requirementsof this function vary over time, so doesthe presence of each gene family in thegenome.Most recently, using a combination ofsequence and structural data, we examinedthe conservation of amino acid sequencesbetween related DNA-binding proteins,and the effect that mutations have onDNA sequence recognition. The structuralfamilies described above were expanded toinclude proteins that are related by sequencesimilarity, but whose structures remainunsolved. Again, members of the samefamily are homologous, and probably derivefrom a common ancestor.Amino acid conservations were calculatedfor the multiple sequence alignmentsof each family [78]. Generally, alignmentpositions that interact with the DNA arebetter conserved than the rest of the proteinsurface, although the detailed patternsof conservation are quite complex. Residuesthat contact the DNA backbone are highlyconserved in all protein families, providinga set of stabilising interactions that arecommon to all homologous proteins. Theconservation of alignment positions thatcontact bases, and recognise the DNAsequence, are more complex and could berationalised by defining a three-class modelfor DNA-binding. First, protein familiesthat bind non-specifically usually containseveral conserved base-contacting residues;without exception, interactions are made inthe minor groove where there is littlediscrimination between base types. Thecontacts are commonly used to stabilisedeformations in the nucleic acid structure,particularly in widening the DNA minorgroove. The second class comprise familieswhose members all target the samenucleotide sequence; here, base-contactingpositions are absolutely or highly conservedallowing related proteins to target thesame sequence.The third, and most interesting, classcomprises families in which binding is alsospecific but different members bind distinctbase sequences. Here protein residuesundergo frequent mutations, and familymembers can be divided into subfamiliesaccording to the amino acid sequencesat base-contacting positions; those in thesame subfamily are predicted to bind thesame DNA sequence and those of differentsubfamilies to bind distinct sequences. Onthe whole, the subfamilies correspondedwell with the proteins’ functions and membersof the same subfamilies were found toregulate similar transcription pathways.The combined analysis of sequence andstructural data described by this study providedan insight into how homologousDNA-binding scaffolds achieve differentspecificities by altering their amino acidsequences. In doing so, proteins evolveddistinct functions, therefore allowingstructurally related transcription factors toregulate expression of different genes.Therefore, the relative abundance of transcriptionregulatory families in a genomedepends, not only on the importance of aparticular protein function, but also in theadaptability of the DNA-binding motifs torecognise distinct nucleotide sequences.This, in turn, appears to be best accommodatedby simple binding motifs, such as thezinc fingers.Given the knowledge of the transcriptionregulators that are contained in eachorganism, and an understanding of howthey recognise DNA sequences, it is ofinterest to search for their potential bindingsites within genome sequences [79].For prokaryotes, most analyses have involvedcompiling data on experimentallyknown binding sites for particular proteinsand building a consensus sequence that incorporatesany variations in nucleotides.Additional sites are found by conductingword-matching searches over the entiregenome and scoring candidate sites bysimilarity [80-83]. Unsurprisingly, most ofthe predicted sites are found in non-codingregions of the DNA [80] and the results ofthe studies are often presented in databasessuch as RegulonDB [73]. The consensussearch approach is often complemented bycomparative genomic studies searchingupstream regions of orthologous genes inclosely related organisms. Through such anapproach, it was found that at least 27% ofknown E. coli DNA-regulatory motifs areconserved in one or more distantly relatedbacteria [84].The detection of regulatory sites ineukaryotes poses a more difficult problembecause consensus sequences tend to bemuch shorter, variable, and dispersed oververy large distances. However, initial studiesin S. cerevisiae provided an interestingobservation for the GATA protein in nitrogenmetabolism regulation. While the 5base-pair GATA consensus sequence isfound almost everywhere in the genome, asingle isolated binding site is insufficient toexert the regulatory function [85]. Thereforespecificity of GATA activity comesfrom the repetition of the consensus sequencewithin the upstream regions of controlledgenes in multiple copies. An initialstudy has used this observation to predictnew regulatory sites by searching for overrepresentedoligonucleotides in non-codingregions of yeast and worm genomes [86,87].Having detected the regulatory bindingsites, there is the problem of defining thegenes that are actually regulated, commonlytermed regulons. Generally, binding sitesare assumed to be located directly upstreamof the regulons; however there are differentproblems associated with this assumptiondepending on the organism. For prokaryotes,it is complicated by the presence ofoperons; it is difficult to locate the regulatedgene within an operon since it can lieseveral genes downstream of the regulatorysequence. It is often difficult to predict theorganisation of operons [88], especially todefine the gene that is found at the head,and there is often a lack of long-range conservationin gene order between relatedorganisms [89]. The problem in eukaryotesis even more severe; regulatory sites oftenact in both directions, binding sites areusually distant from regulons because oflarge intergenic regions, and transcriptionregulation is usually a result of combinedaction by multiple transcription factors in acombinatorial manner.Despite these problems, these studieshave succeeded in confirming the transcriptionregulatory pathways of well-characterizedsystems such as the heat shock responseMethod Inform Med 4/2001

354Luscombe, Greenbaum, Gersteinsystem [83]. In addition, it is feasible toexperimentally verify any predictions, mostnotably using gene expression data.6.3 Gene Expression StudiesMany expression studies have so farfocused on devising methods to clustergenes by similarities in expression profiles.This is in order to determine the proteinsthat are expressed together under differentcellular conditions. Briefly, the most commonmethods are hierarchical clustering,self-organising maps, and K-means clustering.Hierarchical methods originally derivedfrom algorithms to construct phylogenetictrees, and group genes in a “bottomup”fashion; genes with the most similarexpression profiles are clustered first, andthose with more diverse profiles areincluded iteratively [90-92]. In contrast, theself-organising map [93, 94] and K-meansmethods [95, 96] employ a “top-down”approach in which the user pre-defines thenumber of clusters for the dataset. Theclusters are initially assigned randomly, andthe genes are regrouped iteratively untilthey are optimally clustered.Given these methods, it is of interest torelate the expression data to other attributessuch as structure, function and subcellularlocalisation of each gene product.Mapping these properties provides aninsight into the characteristics of proteinsthat are expressed together, and also suggestsome interesting conclusions about theoverall biochemistry of the cell. In yeast,shorter proteins tend to be more highlyexpressed than longer proteins, probablybecause of the relative ease with which theyare produced [97]. Looking at the aminoacid content, highly expressed genes aregenerally enriched in alanine and glycine,and depleted in asparagine; these arethought to reflect the requirements ofamino acid usage in the organism, wheresynthesis of alanine and glycine are energeticallyless expensive than asparagine.Turningto protein structure, expression levels ofthe TIM barrel and NTP hydrolase foldsare highest, while those for the leucinezipper, zinc finger and transmembranehelix-containing folds are lowest. ThisMethod Inform Med 4/2001relates to the functions associated withthese folds; the former are commonly involvedin metabolic pathways and the latterin signalling or transport processes [98].This is also reflected in the relationshipwith subcellular localisations of proteins,where expression of cytoplasmic proteins ishigh, but nuclear and membrane proteinstend to be low [99, 100].More complex relationships have alsobeen assessed. Conventional wisdom is thatgene products that interact with each otherare more likely to have similar expressionprofiles than if they do not [101, 102]. However,a recent study showed that this relationshipis not so simple [103]. While expressionprofiles are similar for gene productsthat are permanently associated, forexample in the large ribosomal subunit,profiles differ significantly for productsthat are only associated transiently, includingthose belonging to the same metabolicpathway.As described below, one of the maindriving forces behind expression analysishas been to analyse cancerous cell lines[104]. In general, it has been shown that differentcell lines (eg epithelial and ovariancells) can be distinguished on the basis oftheir expression profiles, and that theseprofiles are maintained when cells aretransferred from an in vivo to an in vitroenvironment [105]. The basis for their physiologicaldifferences were apparent in theexpression of specific genes; for example,expression levels of gene products necessaryfor progression through the cell cycle,especially ribosomal genes, correlated wellwith variations in cell proliferation rate.Comparative analysis can be extended totumour cells, in which the underlyingcauses of cancer can be uncovered bypinpointing areas of biological variationscompared to normal cells. For example inbreast cancer, genes related to cell proliferationand the IFN-regulated signal transductionpathway were found to be upregulated[25, 106]. One of the difficulties incancer treatment has been to target specifictherapies to pathogenetically distinct tumourtypes, in order to maximise efficacyand minimise toxicity. Thus, improvementsin cancer classifications have been centralto advances in cancer treatment. Althoughthe distinction between different forms ofcancer – for example subclasses of acuteleukaemia – has been well established, it isstill not possible to establish a clinical diagnosison the basis of a single test. In arecent study, acute myeloid leukaemia andacute lymphoblastic leukaemia were successfullydistinguished based on the expressionprofiles of these cells [26]. As theapproach does not require prior biologicalknowledge of the diseases, it may provide ageneric strategy for classifying all types ofcancer.Clearly, an essential aspect of understandingexpression data lies in understandingthe basis of transcription regulation.However, analysis in this area is still limitedto preliminary analyses of expression levelsin yeast mutants lacking key components ofthe transcription initiation complex [19,107].7 “… many PRACTICALAPPLICATIONS…”Here, we describe some of the major usesof bioinformatics.7.1 Finding HomologuesAs described earlier, one of the drivingforces behind bioinformatics is the searchfor similarities between different biomolecules.Apartfrom enabling systematic organisationof data, identification of proteinhomologues has some direct practical uses.The most obvious is transferring informationbetween related proteins. For example,given a poorly characterised protein, it ispossible to search for homologues that arebetter understood and with caution, applysome of the knowledge of the latter to theformer. Specifically with structural data,theoretical models of proteins are usuallybased on experimentally solved structuresof close homologues [108]. Similar techniquesare used in fold recognition in whichtertiary structure predictions depend onfinding structures of remote homologuesand checking whether the prediction isenergetically viable [109]. Where biochemicalor structural data are lacking, studies

355What is Bioinformatics?Fig. 3 Above is a schematic outlining how scientists can use bioinformatics to aid rational drug discovery. MLH1 is a human gene encoding a mismatch repair protein (mmr) situated on theshort arm of chromosome 3. Through linkage analysis and its similarity to mmr genes in mice, the gene has been implicated in nonpolyposis colorectal cancer. Given the nucleotide sequence,the probable amino acid sequence of the encoded protein can be determined using translation software. Sequence search techniques can be used to find homologues in model organisms, andbased on sequence similarity, it is possible to model the structure of the human protein on experimentally characterised structures. Finally, docking algorithms could design molecules thatcould bind the model structure, leading the way for biochemical assays to test their biological activity on the actual protein.could be made in low-level organisms likeyeast and the results applied to homologuesin higher-level organisms such ashumans, where experiments are moredemanding.An equivalent approach is also employedin genomics. Homologue-finding is extensivelyused to confirm coding regions innewly sequenced genomes and functionaldata is frequently transferred to annotateindividual genes. On a larger scale, it alsosimplifies the problem of understandingcomplex genomes by analysing simpleorganisms first and then applying thesame principles to more complicated ones –this is one reason why early structuralgenomics projects focused on Mycoplasmagenitalium [75].Ironically, the same idea can be appliedin reverse. Potential drug targets are quicklydiscovered by checking whether homologuesof essential microbial proteins aremissing in humans. On a smaller scale,structural differences between similar proteinsmay be harnessed to design drugmolecules that specifically bind to onestructure but not another.7.2 Rational Drug DesignOne of the earliest medical applications ofbioinformatics has been in aiding rationalMethod Inform Med 4/2001

356Luscombe, Greenbaum, Gersteindrug design. Fig. 3 outlines the commonlycited approach, taking the MLH1 gene productas an example drug target. MLH1 is ahuman gene encoding a mismatch repairprotein (mmr) situated on the short arm ofchromosome 3 [110]. Through linkage analysisand its similarity to mmr genes in mice,the gene has been implicated in nonpolyposiscolorectal cancer [111]. Given thenucleotide sequence, the probable aminoacid sequence of the encoded protein canbe determined using translation software.Sequence search techniques can then beused to find homologues in model organisms,and based on sequence similarity, itis possible to model the structure of thehuman protein on experimentally characterizedstructures. Finally, docking algorithmscould design molecules that could bind themodel structure, leading the way for biochemicalassays to test their biologicalactivity on the actual protein.7.3 Large-scale CensusesAlthough databases can efficiently store allthe information related to genomes, structuresand expression datasets, it is useful tocondense all this information into understandabletrends and facts that users canreadily understand. Broad generalisationshelp identify interesting subject areas forfurther detailed analysis, and place new observationsin a proper context. This enablesone to see whether they are unusual in anyway.Through these large-scale censuses, onecan address a number of evolutionary, biochemicaland biophysical questions. Forexample, are specific protein folds associatedwith certain phylogenetic groups? Howcommon are different folds within particularorganisms? And to what degree arefolds shared between related organisms?Does this extent of sharing parallel measuresof relatedness derived from traditionalevolutionary trees? Initial studies showthat the frequency of folds differs greatlybetween organisms and that the sharing offolds between organisms does in fact followtraditional phylogenetic classifications[37, 112, 113]. We can also integrate data onprotein functions; given that the particularMethod Inform Med 4/2001protein folds are often related to specificbiochemical functions [52, 53], these findingshighlight the diversity of metabolicpathways in different organisms [36, 89].As we discussed earlier, one of the mostexciting new sources of genomic informationis the expression data. Combining expressioninformation with structural and functionalclassifications of proteins we can askwhether the high occurrence of a proteinfold in a genome is indicative of high expressionlevels [97]. Further genomic scaledata that we can consider in large-scale surveysinclude the subcellular localisations ofproteins and their interactions with eachother [114-116]. In conjunction with structuraldata, we can then begin to compile amap of all protein-protein interactions inan organism.8. ConclusionsWith the current deluge of data, computationalmethods have become indispensableto biological investigations. Originallydeveloped for the analysis of biological sequences,bioinformatics now encompassesa wide range of subject areas includingstructural biology, genomics and gene expressionstudies. In this review, we providedan introduction and overview of the currentstate of field. In particular, we discussedthe types of biological information anddatabases that are commonly used, examinedsome of the studies that are being conducted– with reference to transcriptionregulatory systems – and finally looked atseveral practical applications of the field.Two principal approaches underpin all studiesin bioinformatics. First is that of comparingand grouping the data according tobiologically meaningful similarities and second,that of analysing one type of data toinfer and understand the observations foranother type of data. These approaches arereflected in the main aims of the field,which are to understand and organise theinformation associated with biologicalmolecules on a large scale. As a result,bioinformatics has not only provided greaterdepth to biological investigations, butadded the dimension of breadth as well. Inthis way, we are able to examine individualsystems in detail and also compare themwith those that are related in order to uncovercommon principles that apply acrossmany systems and highlight unusual featuresthat are unique to some.AcknowledgmentsWe thank Patrick McGarvey for comments onthe manuscript.References1. Reichhardt T. It’s sink or swim as a tidal waveof data approaches. Nature 1999. 399 (6736):517-20.2. Benson DA, et al. GenBank. Nucleic AcidsRes 2000; 28 (1): 15-8.3. Bairoch A, Apweiler R. The SWISS-PROTprotein sequence database and its supplementTrEMBL in 2000. Nucleic Acids Res 2000; 28(1): 45-8.4. Fleischmann RD, et al. Whole-genome randomsequencing and assembly of Haemophilusinfluenzae Rd. Science 1995; 269(5223): 496-512.5. Drowning in data. The Economist 1999 (26June 1999).6. Bernstein FC, et al. The Protein Data Bank. Acomputer-based archival file for macromolecularstructures. Eur J Biochem 1977; 80 (2):319-24.7. Berman HM, et al. The Protein Data Bank.Nucleic Acids Res 2000; 28 (1): 235-42.8. Pearson WR, Lipman DJ. Improved tools forbiological sequence comparison. Proc NatlAcad Sci USA 1988; 85 (8): 2444-8.9. Altschul SF, et al. Gapped BLAST and PSI-BLAST: a new generation of protein databasesearch programs. Nucleic Acids Res 1997; 25(17): 3389-402.10. Fleischmann RD, et al. Whole-genome randomsequencing and assembly of Haemophilusinfluenzae Rd. Science 1995; 269(5223): 496-512.11. Lander ES, et al. Initial sequencing and analysisof the human genome. Nature 2001; 409:860-921.12. Venter JC, et al. The sequence of the humangenome. Science 2001; 291 (5507): 1304-51.13. Tatusova TA, Karsch-Mizrachi I, Ostell JA.Complete genomes in WWW Entrez: datarepresentation and analysis. Bioinformatics1999; 15 (7-8): 536-43.14. Eisen MB, Brown PO. DNA arrays for analysisof gene expression. Methods Enzymol,1999; 303: 179-205.15. Cheung VG, et al. Making and reading microarrays.Nat Genet 1999; 21 (1 Suppl): 15-9.16. Duggan DJ, et al. Expression profiling usingcDNA microarrays. Nat Genet 1999. 21(1 Suppl): 10-4.17. Lipshutz RJ, et al. High density syntheticoligonucleotide arrays. Nat Genet 1999; 21 (1):20-4.

357What is Bioinformatics?18. Velculescu VE, et al. Serial Analysis of GeneExpression. Detailed Protocol 1999.19. Holstege FC, et al. Dissecting the regulatorycircuitry of a eukaryotic genome. Cell 1998; 95(5): 717-28.20. Roth FP, Estep PW, Church GM. FindingDNA regulatory motifs within unaligned noncodingsequences clustered by whole-genomemRNA quantitation. Nat Biotech 1998; 16(10): 939-45.21. Jelinsky SA, Samson LD. Global response ofSaccharomyces cerevisiae to an alkylatingagent. Proc Natl Acad Sci USA 1999; 96 (4):1486-91.22. Cho RJ, et al. A genome-wide transcriptionalanalysis of the mitotic cell cycle. Mol Cell1998; 2 (1): 65-73.23. DeRisi JL, Iyer VR, Brown PO. Exploring themetabolic and genetic control of gene expressionon a genomic scale. Science 1997; 278(5338): 680-6.24. Winzeler EA, et al. Functional characterizationof the S. cerevisiae genome by genedeletion and parallel analysis. Science 1999;285 (5429): 901-6.25. Perou CM, et al. Molecular portraits of humanbreast tumours. Nature 2000; 406 (6797):747-52.26. Golub TR, et al. Molecular classification ofcancer: class discovery and class prediction bygene expression monitoring. Science 1999; 286(5439): 531-7.27. Pedersendagger AG, et al. A DNA structuralatlas for Escherichia coli. J Mol Biol 2000; 299(4): 907-30.28. Kanehisa M; Goto S. KEGG: kyoto encyclopediaof genes and genomes. Nucleic AcidsRes 2000; 28 (1): 27-30.29. Jeffery CJ. Moonlighting proteins. TIBS 1999;24 (1): 8-11.30. Chothia, C. Proteins. One thousand familiesfor the molecular biologist. Nature 1992; 357(6379): 543-4.31. Orengo CA, Jones DT, Thornton JM. Proteinsuperfamilies and domain superfolds. Nature1994; 372 (6507): 631-4.32. Lesk AM, Chothia C. How different aminoacid sequences determine similar proteinstructures: the structure and evolutionarydynamics of the globins. J Mol Biol 1980; 136(3): 225-70.33. Russell RB, et al. Recognition of analogousand homologous protein folds: analysis ofsequence and structure conservation. J MolBiol 1997; 269 (3): 423-39.34. Russell RB, et al. Recognition of analogousand homologous protein folds – assessment ofprediction success and associated alignmentaccuracy using empirical substitution matrices.Protein Eng 1998; 11 (1): 1-9.35. Fitch WM. Distinguishing homologous fromanalogous proteins. Syst Zool 1970; 19: 99-110.36. Tatusov RL, Koonin EV, Lipman DJ. A genomicperspective on protein families. Science1997; 278 (5338): 631-7.37. Gerstein M, Hegyi H. Comparing genomes interms of protein structure: surveys of a finiteparts list. FEMS Microbiol Rev 1998; 22 (4):277-304.38. Skolnick J, Fetrow JS. From genes to proteinstructure and function: novel applications ofcomputational approaches in the genomic era.Trends Biotech 2000; 18: 34-9.39. Qian J, et al. PartsList: a web-based system fordynamically ranking protein folds based ondisparate attributes, including whole-genomeexpression and interaction information.Nucleic Acids Res 2001; 29 (8): 1750-64.40. Gerstein M. Integrative database analysis instructural genomics. Nat Struct Biol 2000; 7Suppl: 960-3.41. Etzold T, Ulyanov A, Argos P. SRS: informationretrieval system for molecular biology databanks. Methods Enzymol 1996; 266: 114-28.42. Schuler GD, et al. Entrez: molecular biologydatabase and retrieval system. MethodsEnzymol 1996; 266: 141-62.43. Wade K. Searching Entrez PubMed anduncover on the internet. Aviat Space EnvironMed 2000; 71 (5): 559.44. Bertone P, et al. SPINE: An integratedtracking database and datamining approachfor high-throughput structural proteomics,enabling the determination of the propertiesof readily characterized proteins. NucleicAcids Res. In Press.45. Zhang MQ. Promoter analysis of co-regulatedgenes in the yeast genome. Comput Chem1999; 23 (3-4): 233-50.46. Boguski MS. Biosequence exegesis. Science1999; 286 (5439): 453-5.47. Miller C, Gurd J, Brass A. A RAPIDalgorithm for sequence database comparisons:application to the identification of vectorcontamination in the EMBL databases. Bioinformatics1999; 15 (2): 111-21.48. Gonnet GH, Korostensky C, Brenner S.Evaluation measures of multiple sequencealignments. J Comput Biol 2000; 7 (1-2):261-76.49. Orengo CA, Taylor WR. SSAP: sequentialstructure alignment program for protein structurecomparison. Methods Enzymol 1996; 266:617-35.50. Orengo CA. CORA – topological fingerprintsfor protein structural families. Protein Sci1999; 8 (4): 699-715.51. Russell RB, Sternberg MJ. Structure prediction.How good are we? Curr Biol 1995; 5 (5):488-90.52. Martin AC, et al. Protein folds and functions.Structure 1998; 6 (7): 875-84.53. Hegyi H, Gerstein M. The relationship betweenprotein structure and function: a comprehensivesurvey with application to theyeast genome. J Mol Biol 1999; 288 (1): 147-64.54. Russell RB, Sasieni PD, Sternberg MJE.Supersites within superfolds. Binding sitesimilarity in the absence of homology. J MolBiol 1998; 282 (4): 903-18.55. Wilson CA, Kreychman J, Gerstein M. Assessingannotation transfer for genomics:quantifying the relations between proteinsequence, structure and function throughtraditional and probabilistic scores. J Mol Biol2000; 297 (1): 233-49.56. Harrison SC. A structural taxonomy of DNAbindingdomains. Nature 1991; 353 (6346):715-9.57. Luscombe NM, et al.An overview of the structuresof protein-DNA complexes. GenomeBiology 2000; 1 (1): 1-37.58. Jones S, et al. Protein-DNA interactions: Astructural analysis. J Mol Biol 1999; 287 (5):877-96.59. Suzuki M, Gerstein M. Binding geometry ofalpha-helices that recognize DNA. Proteins1995; 23 (4): 525-35.60. Luscombe NM, Thornton JM. Protein-DNAinteractions: a 3D analysis of alpha-helixbindingin the major groove. Manuscript inpreparation.61. Suzuki M, et al. DNA recognition code oftranscription factors. Protein Eng 1995; 8 (4):319-28.62. Suzuki M. DNA recognition by a -sheet.Protein Eng 1995; 8 (1): 1-4.63. Seeman NC, Rosenberg JM, Rich A. Sequencespecific recognition of double helical nucleicacids by proteins. Proc Natl Acad Sci USA1976; 73: 804-8.64. Suzuki M. A framework for the DNA-proteinrecognition code of the probe helix intranscription factors: the chemical and stereochemicalrules. Structure 1994; 2 (4): 317-26.65. Mandel-Gutfreund Y, Schueler O, Margalit H.Comprehensive analysis of hydrogen bonds inregulatory protein-DNA complexes: in searchof common principles. J Mol Biol 1995; 253(2): 370-82.66. Luscombe NM, Laskowski RA, Thornton JM.Protein-DNA interactions: a 3D analysis ofamino acid-base interactions. Nucleic AcidsRes. In Press.67. Mandel-Gutfreund Y, Margalit H. Quantitativeparameters for amino acid-base interaction:inplications for prediction of protein-DNA binding sites. Nucleic Acids Res 1998;26: 2306-12.68. Sternberg MJ, Gabb HA, Jackson RM. Predictivedocking of protein-protein and protein-DNA complexes. Curr Opin Struct Biol 1998;8 (2): 250-6.69. Aloy P, et al. Modelling repressor proteinsdocking to DNA. Proteins 1998; 33 (4):535-49.70. Dickerson RE. DNA-binding: the prevalenceof kinkiness and the virtues of normality.Nucleic Acids Res 1998; 26 (8): 1906-26.71. Perez-Rueda E, Collado-Vides J. The repertoireof DNA-binding transcriptional regulatorsin Escherichia coli K-12. Nucleic AcidsRes 2000; 28 (8): 1838-47.72. Mewes HW, et al. MIPS: a database for genomesand protein sequences. Nucleic Acids Res2000; 28 (1): 37-40.73. Salgado H, et al. RegulonDB (version 3.0):transcriptional regulation and operon organizationin Escherichia coli K-12. NucleicAcids Res 2000; 28 (1): 65-7.Method Inform Med 4/2001

358Luscombe, Greenbaum, Gerstein74. Wingender E, et al. TRANSFAC: an integratedsystem for gene expression regulation. NucleicAcids Res 2000; 28 (1): 316-9.75. Teichmann SA, Chothia C, Gerstein M.Advances in structural genomics. Curr OpinStruct Biol 1999; 9 (3): 390-9.76. Aravind L, Koonin EV. DNA-binding proteinsand evolution of transcription regulationin the archaea. Nucleic Acids Res 1999; 27(23): 4658-70.77. Huynen MA, van Nimwegen E.The frequencydistribution of gene family sizes in completegenomes. Mol Biol Evol 1998; 15 (5): 583-9.78. Luscombe NM, Thornton JM. Protein-DNAinteractions: an analysis of amino acid conservationand the effect on binding specificity.Manuscript in preparation.79. Gelfand MS. Prediction of function in DNAsequence analysis. J Comp Biol 1995; 1:87-115.80. Robison K, McGuire AM, Church GM. Acomprehensive library of DNA-binding sitematrices for 55 proteins applied to thecomplete Escherichia coli K-12 genome. J MolBiol 1998; 284 (2): 241-54.81. Thieffry D, et al. Prediction of transcriptionalregulatory sites in the complete genomesequence of Escherichia coli K-12. Bioinformatics1998; 14 (5): 391-400.82. Mironov AA. et al. Computer analysis oftranscription regulatory patterns in completelysequenced bacterial genomes. Nucleic AcidsRes 1999; 27 (14): 2981-9.83. Gelfand MS, Koonin EV, Mironov AA.Prediction of transcription regulatory sites inArchaea by a comparative genomic approach.Nucleic Acids Res 2000; 28 (3): 695-705.84. McGuire AM, Hughes JD, Church GM.Conservation of DNA regulatory motifs anddiscovery of new motifs in microbial genomes.Genome Res 2000; 10 (6): 744-57.85. Bysani N, Daugherty JR, Cooper TG. Saturationmutagenesis of the UASNTR (GATAA)responsible for nitrogen catabolite repressionsensitivetranscriptional activation of theallantoin pathway genes in Saccharomycescerevisiae. J Bacteriol 1991; 173 (16): 4977-82.86. Clarke ND, Berg JM. Zinc fingers in Caenorhabditiselegans: finding families and probingpathways. Science 1998; 282 (5396): 2018-22.87. van Helden J, Andre B, Collado-Vides J.Extracting regulatory sites from the upstreamregion of yeast genes by computationalanalysis of oligonucleotide frequencies. J MolBiol 1998; 281 (5): 827-42.88. Salgado H, et al. Operons in Escherichia coli:genomic analyses and predictions. Proc NatlAcad Sci USA, 2000; 97 (12): 6652-7.89. Tatusov RL, et al. Metabolism and evolutionof Haemophilus influenzae deduced from awhole-genome comparison with Escherichiacoli. Curr Biol 1996; 6 (3): 279-91.90. Eisen MB, et al. Cluster analysis and displayof genome-wide expression patterns. ProcNatl Acad Sci USA 1998; 95 (25): 14863-8.91. Wen X, et al. Large-scale temporal gene expressionmapping of central nervous systemdevelopment. Proc Natl Acad Sci USA 1998;95 (1): 334-9.92. Alon U, et al. Broad patterns of gene expressionrevealed by clustering analysis oftumor and normal colon tissues probed byoligonucleotide arrays. Proc Natl Acad SciUSA 1999; 96 (12): 6745-50.93. Tamayo P, et al. Interpreting patterns of geneexpression with self-organizing maps: methodsand application to hematopoieticdifferentiation. Proc Natl Acad Sci USA 1999;96 (6): 2907-12.94. Toronen P, et al. Analysis of gene expressiondata using self-organizing maps. FEBS Lett1999; 451 (2): 142-6.95. Tavazoie S, et al. Systematic determination ofgenetic network architecture. Nat Genet 1999;22 (3): 281-5.96. Subrahmanyam YV, et al. RNA expressionpatterns change dramatically in human neutrophilsexposed to bacteria. Blood 2001; 97(8): 2457-68.97. Jansen R, Gerstein M. Analysis of the yeasttranscriptome with structural and functionalcategories: characterizing highly expressed proteins.Nucleic Acids Res 2000; 28 (6): 1481-8.98. Gerstein M, Jansen R. The current excitmentin bioinformatics, analysis of whole-genomeexpression data: how does it relate to proteinstructure and function. Curr Opin Struct Biol2000; 10: 574-84.99. Drawid A, Gerstein M. A Bayesian SystemIntegrating Expression Data with SequencePatterns for Localizing Proteins: ComprehensiveApplication to the Yeast Genome. JMol Biol 2000; 301:1059-75.100. Drawid A, Jansen R, Gerstein M. Genomwideanalysis relating expression level withprotein subcellular localisation. Trends Genet2000; 16: 426-30.101. Marcotte EM, et al. Detecting protein functionand protein-protein interactions fromgenome sequences. Science 1999; 285 (5428):751-3.102. Eisenberg D, et al. Protein function in thepost-genomic era. Nature 2000; 405 (6788):823-6.103. Jansen R, Greenbaum D, Gerstein M. Relatingwhole-genome expression data withprotein-protein interactions. Manuscript inpreparation.104. Marx J. DNA arrays reveal cancer in its manyforms. Science 2000; 289 (5485): 1670-2.105. Ross DT, et al. Systematic variation in geneexpression patterns in human cancer cell lines.Nat Genet 2000; 24 (3): 227-35.106. Perou CM, et al. Distinctive gene expressionpatterns in human mammary epithelial cellsand breast cancers. Proc Natl Acad Sci USA1999; 96 (16): 9212-7.107. Livesey FJ, et al. Microarray analysis of thetranscriptional network controlled by thephotoreceptor homeobox gene Crx. Curr Biol2000; 10 (6): 301-10.108. Sali A, Blundell TL. Comparative proteinmodelling by satisfaction of spatial restraints.Journal of Molecular Biology 1993; 234 (3):779-815.109. Jones DT, Taylor WR, Thornton JM. A newapproach to protein fold recognition. Nature1992; 358 (6381): 86-9.110. Kok K, Naylor SL, Buys CH. Deletions of theshort arm of chromosome 3 in solid tumorsand the search for suppressor genes.Advancesin Cancer Research 1997; 71: 27-92.111. Syngal S, et al. Sensitivity and specificity ofclinical criteria for hereditary non-polyposiscolorectal cancer associated mutations inMSH2 and MLH1. Journal Med Genet 2000;37 (9): 641-5.112. Lin J, Gerstein M. Whole-genome trees basedon the occurrence of folds and orthologs:implications for comparing genomes ondifferent levels. Genome Res 2000; 10 (6):808-18.113. Harrison PM, Echols N, Gerstein MB. Diggingfor dead genes: an analysis of the characteristicsof the pseudogene population in theCaenorhabditis elegans genome. NucleicAcids Res 2001; 29 (3): 818-30.114. Uetz P, et al. A comprehensive analysis ofprotein-protein interactions in Saccharomycescerevisiae. Nature 2000; 403 (6770): 623-7.115. Ross-Macdonald P, et al. Transposon mutagenesisfor the analysis of protein production,function, and localization. Methods Enzymol1999; 303: 512-32.116. Mewes HW, et al. MIPS: a database forgenomes and protein sequences. NucleicAcids Res 1999; 27 (1): 44-8.Correspondence to:Mark GersteinDepartment of Molecular Biophysics and BiochemistryYale University, 266 Whitney AvenuePO Box 208114, New Haven CT 06520-8114, USAE-Mail: mark.gerstein@yale.eduMethod Inform Med 4/2001