INTEGRATING COMPUTATIONAL PROTEIN FUNCTION PREDICTION INTO DRUG DISCOVERY INITIATIVES

Abstract

Pharmaceutical researchers must evaluate vast numbers of protein sequences and formulate innovative strategies for identifying valid targets and discovering leads against them as a way of accelerating drug discovery. The ever increasing number and diversity of novel protein sequences identified by genomic sequencing projects and the success of worldwide structural genomics initiatives have spurred great interest and impetus in the development of methods for accurate, computationally empowered protein function prediction and active site identification. Previously, in the absence of direct experimental evidence, homology-based protein function annotation remained the gold-standard for in silico analysis and prediction of protein function. However, with the continued exponential expansion of sequence databases, this approach is not always applicable, as fewer query protein sequences demonstrate significant homology to protein gene products of known function. As a result, several non-homology based methods for protein function prediction that are based on sequence features, structure, evolution, biochemical and genetic knowledge have emerged. Herein, we review current bioinformatic programs and approaches for protein function prediction/annotation and discuss their integration into drug discovery initiatives. The development of such methods to annotate protein functional sites and their application to large protein functional families is crucial to successfully utilizing the vast amounts of genomic sequence information available to drug discovery and development processes.

INTRODUCTION

The molecular details of protein function are of fundamental importance in designing specific and selective inhibitors or ligands to modulate protein activity as part of the process of developing small-molecule drug candidates. By identifying and characterizing protein structure, function, and active site information early in the discovery process, pharmaceutical researchers can study multiple targets using integrated biological, structural, and chemical methods simultaneously to more rapidly, cost effectively discover selective lead compounds.

Since the emergence of initiatives in the early 1990’s, genome sequencing projects have been exceedingly successful in producing an impressive array of fully sequenced genomes and vast amounts of protein sequence information. The availability of these genome sequences and their associated curated annotations has generated a wealth of information for new avenues of investigation, including drug discovery efforts. However, one of the fundamental challenges for the post-genomic era is to develop methods to incorporate the exponentially growing protein sequence information for thousands of functionally-uncharacterized proteins into large-scale drug discovery strategies. In the human genome and in the genomes of pathogenic agents there will be thousands of potential, unexplored drug targets, without the development of robust methods for the computational prediction of protein function or functional site identification.

Similarly, global protein structure initiatives underway today are providing new high-resolution protein structures [Burley and Bonanno 2003; Weigelt et al. 2008] that are representative of both current and novel pharmaceutical targets and endowing a strong foundation for drug discovery. Representative structures provide templates for comparative modeling and knowledge-based potentials for ab initio structure folding methods so that new structural models can be generated for protein sequences with high sequence similarity (>30%) to expand the known structural space relative to experimentally derived templates [Grant 2009]. The accuracy of these computational models can be detailed enough to provide valuable information in lead development for the structure and chemistry of binding sites identified in the protein structure. However, while structural genomics initiatives continue to produce new protein structures, the current focus is on characterizing the largest number of different folds to have the best possible sampling of structure space. As a consequence, a large number of structures and potential comparative models belong to proteins of unknown function, annotated merely as ‘hypothetical proteins’. This fact has greatly increased the interest in computational methods for functional inference. Herein, we review the research approaches and recently developed tools in the field of computational protein function prediction and discuss the ways these can be integrated into the process of drug discovery.

FUNCTIONAL ANNOTATION OF PROTEINS

Biological function can be highly contextual with different degrees of functional specificity and can be described at many levels ranging from biochemical, process, pathway, organ, or organism levels. To make protein function annotation available universally and for throughput computational processing, there is an essential need to describe the function of any gene product in any organism with a controlled and well-defined vocabulary. Several schemes for classifying protein function have been developed, most notably the Enzyme Commission (EC) Classification, which described enzymatic reactions using four-levels of indentified hierarchy. More recently, to address the need to describe complex protein functions beyond biochemical ones, the open-source Gene Ontology (GO) schema [2009a; Ashburner et al. 2000] has become the standard approach for a controlled vocabulary and a machine-readable ontology for functional annotation. GO comprises a framework of controlled vocabularies describing three aspects of gene product function: molecular function, biological process, and cellular location. This scheme represents the expanded view of protein function, whereby a protein is defined as an element in a network of its interactions. The GO Annotation (GOA) project aims to annotate all of the complete and incomplete proteomes that exist in the SWISS-PROT Protein Knowledgebase sequence database and its supplement, TrEMBL, using defined GO terms [Camon et al. 2003; Camon et al. 2004], as well as evidence codes reflecting how the annotation was obtained or determined. Through such standardization, protein function annotations may be computationally processed and a means for programs to output protein function predictions exists.

PROTEIN FUNCTION PREDICTION METHODS

Protein function prediction methods mainly fall into sequence- and structure-based approaches. Herein, we outline the best described bioinformatic technologies for sequence- and structure-based protein function prediction. A schematic overview of these protein function prediction methods is shown in Figure 1. Table 1 lists a number of important databases and collections (sequence, structural, and ontology) that are extremely useful for approaches to protein function annotation.

Figure 1.

A schematic overview of protein function prediction methods. Various approaches to protein function prediction (grey boxes) are described in the text along with various methodologies employed in these approaches (white boxes). Both protein sequences and structures can provide information for family classification and functional inference.

Table 1.

Useful databases for protein function annotation.

Database	URL
CATH	http://www.biochem.ucl.ac.uk/bsm/cath/
COGs	http://www.ncbi.nlm.nih.gov/COG/
Catalytic Site Atlas	http://www.ebi.ac.uk/thornton-srv/databases/CSA/
DBAli	http://www.salilab.org/DBAli/
EntrezStructure/MMDB	http://www.ncbi.nlm.nih.gov/sites/entrez?db=Structure
GenBank	http://www.ncbi.nlm.nih.gov/Genbank/GenbankSearch.html
Genecensus	http://bioinfo.mbb.yale.edu/genome/
Gene Ontology (GO)	http://www.genontology.org
GOOD	http://goods.ibms.sinica.edu.tw/goods/
InParanoid7	http://inparanoid.sbc.su.se/cgi-bin/index.cgi
MACiE	http://www.ebi.ac.uk/thornton-srv/databases/MACiE/
ModBase	http://modbase.compbio.ucsf.edu/modbase-cgi/index.cgi
NCBI	http://www.ncbi.nlm.nih.gov/
OrthoDB	http://cegg.unige.ch/orthodb3
Pfam	http://pfam.sanger.ac.uk/
PDB	http://www.rcsb.org/pdb/home/home.do
PSI-StructuralGenomics Knowledgebase	http://kb.psi-structuralgenomics.org/KB/
SCOP	http://scop.mrc-lmb.cam.ac.uk/scop/
TIGR	http://www.tigr.org/tdb/mdb/mdbcomplete.html
TIGRFAMS	http://www.tigr.org/TIGRFAMs/index.shtml
TargetDB	http://targetdb.pdb.org
TreeFam	http://www.treefam.org/
UniProt	http://www.uniprot.org/
UniProtKB/Swiss-Prot	http://www.expasy.org/sprot/
UniProtKB/TrEMBL	http://www.expasy.org/sprot/

Similarity-based approaches

Sequence homology-based methods

The most widely used approach for function prediction is homology transfer. For a given unannotated protein, this approach is based on searches for an annotated homolog (highly sequence-similar) and uses the experimentally verified function of the latter to infer the function of the former. The rationale for this approach is the assumption that two sequences with a high degree of similarity most likely evolved from a common ancestor and thus must have similar function. An important distinction in this context is between orthologous and parologous sequences, however. In general, function tends to be more conserved in orthologs than in paralogs [Theissen 2002], however, there are examples of more functionally divergent orthologs than corresponding paralogs [Punta and Ofran 2008]. Several data bases have been created to identify orthologous genes, for example COGs [Tatusov et al. 1997] and InParanoid [Remm et al. 2001], and also to catalog groups of orthologous genes in a hierarchical manner, such as in OrthoDB [Kriventseva et al. 2008].

Homology-based approaches require clustering of proteins into evolutionary families using sequence similarity-detection and/or alignment-search tools such as BLAST [Altschul et al. 1990] or tools based on multiple sequence alignments such as PSI-BLAST [Altschul et al. 1997], MAFFT [Katoh et al. 2002], and ProbCons [Do et al. 2005]. Several available resources provide pre-compiled sequence-based family assignments for proteins on a genomic scale, for example, PIRSF [Nikolskaya et al. 2006] in which a set of rules is applied to define primary and curated clusters (protein superfamilies) divided into common domain architectures. Table 2 lists a number of important resources, implementations, and tools for sequence searching and sequence alignment. Full-length similarity methods have been largely superseded by more sophisticated sequence pattern-based methods.

Table 2.

Useful sequence search and alignment programs and servers.

Program/Server	URL
Align	http://bioinfo.mbb.yale.edu/Align/
CLUSTALW2	http://www.ebi.ac.uk/Tools/clustalw2/index.html
COMPASS	http://prodata.swmed.edu/compass/compass.php
DALIGN-TX	http://dialign-tx.gobics.de/
FFAS03	http://ffas.ljcrf.edu
HMMER3	http://hmmer.janelia.org/
MAFFT	http://align.bmr.kyushu-u.ac.jp/mafft/software/
MultAlign	http://mendel.ethz.ch:8080/Server/MultAlign.html
MUSCLE	http://www.drive5.com/muscle/
PFP	http://kiharalab.org/web/pfp.php
PSI-BLAST	http://blast.ncbi.nlm.nih.gov/Blast.cgi
PROBCONS	http://probcons.stanford.edu/
RE-MuSiC	http://140.113.239.131/RE-MUSIC/
SAM-T08	http://compbio.soe.ucsc.edu/HMM-apps/HMM-applications.html
T-coffee	http://www.tcoffee.org/
Espresso/3D-coffee	http://www.tcoffee.org/Projects_home_page/expresso_home_page.html

Sequence motif and pattern-based methods

Two protein sequences that would not match in sequence searches may still have common sequence signatures that could reveal their functional relatedness. Finding one of these well-characterized motifs in a newly discovered sequence could offer some insights into its function. Several resources classify proteins on the basis of locally conserved sequence patterns (active site motifs), which often reflect the function(s) of the whole protein. The advantage of these profile methods are that they provide greater sensitivity compared to simple sequence-sequence comparisons because the profiles inherently contain both well-conserved residue and variable residue information for the motif/pattern across protein families. The most common type of profile is the hidden Markov model (HMM) and several methods exist for creating them from sequences of whole protein families. Profiling tools based on multiple sequence alignments are PSI-BLAST, HMMER [Bateman et al. 1999], and SAM [2009b], for example. The recent PFP [Hawkins et al. 2009] tool uses three rounds of PSI-BLAST and tolerant sequence similarity thresholds to include the annotations of remote homologues or homologous domains.

The PROSITE [Hulo et al. 2008] resource comprises manually selected biologically important motifs and has three types of signatures: patterns, rules, and profiles. While the two local signatures, patterns and rules, extend over just a few residues, profiles extend to the level of entire domains. PROSITE scans a query sequence against short, position-specific residue profiles that are characteristic of distinct protein families. Another widely used database is Pfam [Bateman et al. 2004; Sammut et al. 2008] comprising motifs that span over entire domains and focuses on the functional aspect of the domain definition. Pfam currently contains more than 10,000 family profiles and covers roughly 75% of UniProt [Wu et al. 2006] sequences, reflecting about half their amino acids. SMART [Letunic et al. 2009] utilizes the same approach and consists of a considerably smaller but completely manually curated set of families, and other well-known functional motif databases include BLOCKS [Henikoff et al. 2000], PRINTS [Attwood et al. 2003] and ELM [Puntervoll et al. 2003]. Other resources, for example CDD [Marchler-Bauer et al. 2009], use externally defined profiles to provide rapid assignments to sequence queries using a BLAST-like engine. The PANTHER [Mi et al. 2005] database distinguishes functional divergence within homologous protein families by defining groups of protein sequences into functional subfamilies. TIGRFAMs [Haft et al. 2003] uses models of full-length proteins and shorter regions at the levels of superfamilies, subfamilies and functional conservation in families of ‘equivalogs’ – sets of homologous proteins conserved with respect to function since their last common ancestor. Table 3 lists these and a number of additional important sequence similarity searching and sequence-based function assignment programs or servers.

Table 3.

Sequence similarity search and sequence-based function assignment methods.

Database	URL
BLOCKS	http://blocks.fhcrc.org/
CDD/CD-Search	http://www.ncbi.nlm.nih.gov/Structure/cdd/wrpsb.cgi
ELM	http://elm.eu.org/
EVEREST	http://www.everest.cs.huji.ac.il/
HHsearch/FHMMmer/HHpred	http://toolkit.lmb.uni-muenchen.de/sections/search
InterPro	http://www.ebi.ac.uk/interpro/
PANTHER	http://www.pantherdb.org/
Pfam	http://pfam.sanger.ac.uk/
PIRSF	http://pir.georgetown.edu/pirwww/dbinfo/pirsf.shtml
PRINTS	http://www.bioinf.manchester.ac.uk/dbbrowser/PRINTS/index.php
ProDom	http://prodom.prabi.fr/prodom/current/html/home.php
PROSITE	http://www.expasy.org/prosite/
SMART	http://smart.embl-heidelberg.de/
SUPERFAMILY	http://supfam.cs.bris.ac.uk/SUPERFAMILY/
TIGR	http://www.tigr.org/tdb/mdb/mdbcomplete.html
TIGRFAMS	http://www.tigr.org/TIGRFAMs/index.shtml

Genomic context and phylogenomic-based methods

Genomic context-based prediction, also termed phylogenomic profiling, comprises protocols for predicting protein function based on the observation that proteins with similar inter-genomic profiles are thought to have evolved in tandem and share a common function [Eisen and Fraser 2003; Gomase and Tagore 2009; Sjolander 2004]. The active sites of proteins can be predicted using phylogenetic analysis and assessment of tree-determinant residues [del Sol et al. 2003]. The evolutionary trace (ET) method is well described and uses trees to rank residues by evolutionary importance and map these onto the structure to identify clusters and functional sites [Yao et al. 2003]. Phylomat [Graham et al. 2004] is a motif analysis tool for phylogenomics that scans predicted proteome sets for proteins containing highly conserved amino acid motifs or domains for analysis of the evolutionary history of these motifs/domains. RIO [Zmasek and Eddy 2002], SIFTER [Engelhardt et al. 2005; Zmasek and Eddy 2002], and OrthoStrapper [Storm and Sonnhammer 2002] algorithms search for protein orthologs by inferring gene duplications on a gene tree by comparing it to a species tree, thereby distinguishing orthologous from paralogous events.

Phylogenomics refers to the application of phylogenetic information to genomic studies and considers the evolutionary history of homologs in the prediction of function to increase the accuracy of annotation transfer. Annotation transfer is performed from the closest ortholog, rather than from the most similar sequence. Additionally, in the prokaryote genomes, the loci of functionally related proteins tend to be chromosomally co-localized. Several phylogenomic-based protein function prediction methods thus combine co-evolution and chromosomal proximity observations into function prediction algorithms, such as in Phydbac2 [Enault et al. 2004].

Expression-based prediction methods

Following the rationale of co-location, genes involved in similar cellular functions tend to be co-transcribed. Thus, from the analysis of gene expression arrays, unknown genes co-expressed with genes of known function may be functionally annotated through co-transcriptional associations. Unlike sequence motif-based approaches which center on molecular function, expression-based predictions can be useful for the annotation of the cellular aspect of protein function [Sleator and Walsh]. An extension of this concept is that most cellular processes are carried out by groups of physically interacting proteins and therefore, interacting proteins may have similar cellular functions. If so, protein-protein interaction (PPI) data may represent great potential for facilitating protein function annotation. Several PPI databases are available including the STRING [Jensen et al. 2009], DIP [Lehne and Schlitt 2009], GRID [Breitkreutz et al. 2003], MINT [Zanzoni et al. 2002], OPHID [Brown and Jurisica 2005], HAPPI [Chen et al. 2009b], HPRD [Keshava Prasad et al. 2009] and PIPs [McDowall et al. 2009] databases, as well as servers for mining and predicting protein-protein interactions, such as PPISearch [Chen et al. 2009a], PRISM [Keskin et al. 2008], and PPI Finder [He et al. 2009].

As sequence databases continue to expand, the homology-based transfer approach losses utility. As a direct consequence of exponential sequence expansion of unannotated novel sequence data from large-scale genomic sequencing projects, the number of clustered similar proteins for which no single annotated reference sequence exists is rapidly growing. Indeed, it has been estimated that <35% of all proteins can be annotated automatically with homology-based transfer, while >30% of all proteins cannot [Rost et al. 2003].

Structure-based approaches

As homologous proteins evolve, their three-dimensional structures often remain more conserved than their sequences. Proteins sharing similar function often have similar overall protein folds as a result of descent from a common ancestral protein. Additionally, similarities in protein structure can be more reliable than sequence similarities for grouping together distant homologs [Brenner et al. 1996; Rost 1997]. Many protein sequences that exhibit little or no sequence similarity [Gherardini and Helmer-Citterich 2008; Watson et al. 2005] still retain significant structural similarity due to evolutionary constraints, and thus structure is a powerful potential indicator of function [Bartlett et al. 2003; Todd et al. 2001]. Additionally, as function is critically related to structure, the structure of a protein directly suggests the mechanistic determinants of its function [Watson et al. 2005].

Methods for predicting protein function from three-dimensional structure can be classified according to the level of protein structure and specificity at which they perform their analysis, ranging from analysis of the overall protein fold to the identification of highly specific three-dimensional clusters (motifs and patterns) of functional residues. In the cases of the latter, existing methods can be classified generally into two groups: those that use comparative approaches to look for the presence of structural motifs associated with known biochemical function, and those methods consisting of analyzing the physicochemical characteristics of a protein surface to identify patches that have features (e.g. shape, electrostatic properties, etc.) characteristic of functional sites [Gherardini and Helmer-Citterich 2008]. Overall, structural comparison methods can identify even distant evolutionary relationships between proteins and make the identification of independently evolved sites possible. In general, it is suggested to use more than one method since different methods may find different valid matches [Kolodny et al. 2005].

Comparative structure-based approaches

Similar to sequence comparison methods, structural comparison methods can be classified as either global or local searches. Global comparison algorithms are mainly used in protein structure classification and to identify evolutionary links between distant homologues. They can also be used for function prediction, however one should caution that the relationship between fold and function is extremely complex and numerous examples are known of folds supporting a great variety of functions [42]. Proteins sharing similar functions often have similar global structure folds. Finding a fold match serves as the first approach in structure-based functional prediction. Computational tools that can scan the Protein Data Bank (PDB) [Berman et al. 2002; Berman et al. 2000] for global structural similarity or structure classification databases SCOP [Andreeva et al. 2004; Hubbard et al. 1999] and CATH [Greene et al. 2007; Orengo et al. 1997] given a query sequence using structural alignment methods include, CE [Guda et al. 2004; Shindyalov and Bourne 2001], DALI/FSSP [Holm et al. 2006; Holm and Sander 1993], FATCAT [Ye and Godzik 2004], PAST [Taubig et al. 2006], and FAST [Zhu and Weng 2005], Matras [Kawabata 2003], GRATH [Harrison et al. 2003], and FragBag [Budowski-Tal et al.], along with others. Table 4 lists the various programs, servers, and databases, described here and elsewhere for use in structural comparisons and structure-based protein function prediction. The recent Annolite [Marti-Renom et al. 2007] program was developed specifically as a structure-alignment based tool for protein function prediction given a query structure using annotation transfer from similar structures. An important assessment of several structure alignment servers has been performed and concluded that multiplicity in efforts for structure alignments, using multiple methods and algorithms, generates more accurate results than any single approach [Novotny et al. 2004].

Table 4.

Useful structural comparison and structure-based function assignment methods.

Method	Program/Server	URL/Webserver
Fold similarity	Annolite	http://salilab.org/DBAli/?page=tools&action=f_annolitechain
	CATHEDRAL	http://www.cathdb.info/cgi-bin/CathedralServer.pl
	CE	http://cl.sdsc.edu/
	DALI/DaliLite	http://ekhidna.biocenter.helsinki.fi/dali_server/start
	FATCAT	http://fatcat.burnham.org/
	GRATH	http://www.biochem.ucl.ac.uk/cgi-bin/cath/Grath.pl
	MATRAS	http://biunit.aist-nara.ac.jp/matras/
	MAMMOTH	http://ub.cbm.uam.es/mammoth/
	SCALI	http://www.bioinfo.rpi.edu/bystrc/SCALI/
	SSAP	http://www.cathdb.info/cgi-bin/cath/SsapServer.pl
	SSM	http://www.ebi.ac.uk/msd-srv/ssm/ssmstart.html
	STRUCTAL	http://molmovdb.mbb.yale.edu/align/
	TOPS+	http://balabio.dcs.gla.ac.uk/mallika/WebTOPS/
	VAST	http://www.ncbi.nlm.nih.gov/Structure/VAST/vastsearch.html
Active sites	LigBase	http://modbase.compbio.ucsf.edu/ligbase/
	MarkUs	http://luna.bioc.columbia.edu/honiglab/mark-us/cgi-bin/submit.pl
	MOTIF Search	http://motif.genome.jp/
	PatchFinder	http://patchfinder.tau.ac.il/
	PDBSiteScan	http://wwwmgs.bionet.nsc.ru/mgs/gnw/pdbsitescan/
	PINTS	http://www.russell.embl.de/pints/
	PROCAT	http://www.biochem.ucl.ac.uk/bsm/PROCAT/getPDBFILE.html
	Query3D	http://pdbfun.uniroma2.it/
	RIGOR/SPASM	http://xray.bmc.uu.se/usf/spasm.html
	SARIG	http://bioinfo2.weizmann.ac.il/~pietro/SARIG/V3/index.html
	SuMo	http://sumo-pbil.ibcp.fr/cgi-bin/sumo-welcome
	THEMATICS	http://pfweb.chem.neu.edu/thematics/submit.html
Pockets/clefts	CASTp	http://sts.bioengr.uic.edu/castp/index.php
	CAVER	http://loschmidt.chemi.muni.cz/caver/
	ef-Site	http://ef-site.hgc.jp/eF-site/index.jsp
	FEATURE	http://feature.stanford.edu/webfeature/
	FINSITE	http://cssb.biology.gatech.edu/skolnick/files/FINDSITE/index.html
	LIGSITEcsc	http://projects.biotec.tu-dresden.de/pocket/
	PocketPicker	http://gecco.org.chemie.uni-frankfurt.de/pocketpicker/index.html
	SURF’sUP!	http://asia.genesilico.pl/surfs_up/
	SURFNET	http://www.biochem.ucl.ac.uk/~roman/surfnet/surfnet.html
Protein-protein Interaction	BIND	http://bond.unleashedinformatics.com/
	DIP	http://dip.doe-mbi.ucla.edu/dip/Main.cgi
	HAPPI	http://discern.uits.iu.edu:8340/HAPPI/index.html
	IntAct	http://www.ebi.ac.uk/intact/main.xhtml
	MINT	http://mint.bio.uniroma2.it/mint/Welcome.do
	MIPS	http://mips.helmholtz-muenchen.de/proj/ppi/
	PIPs	http://www.compbio.dundee.ac.uk/www-pips/
	PPISearch	http://gemdock.life.nctu.edu.tw/ppisearch/index.php
	PRISM	http://prism.ccbb.ku.edu.tr/prism/
	ProMate	http://bioinfo.weizmann.ac.il/promate/promate.html
	STRING	http://string-db.org/
MetaServers	Gene3D	http://gene3d.biochem.ucl.ac.uk/Gene3D/
	JAFA	http://jafa.burnham.org
	Interpro/InterproScan	http://www.ebi.ac.uk/Tools/InterProScan/
	ProFunc	http://www.ebi.ac.uk/thornton-srv/databases/ProFunc/
	ProKnow	http://proknow.mbi.ucla.edu/
	ProtFun	http://www.protfun.com/
	PSiFR	http://psifr.cssb.biology.gatech.edu/
	SiteEngine	http://bioinfo3d.cs.tau.ac.il/SiteEngine/

Arguably, the function of a protein depends more on the identity and location of a few residues comprising the active site than on the overall fold. In order to directly analyze and compare the residues effectively involved in protein function, local structural comparison methods have been developed. Local structural comparison refers to detecting similar three-dimensional arrangements or motifs (patterns) of a small set of residues, in the context of different global protein folds. As such, in local structure comparisons, one can either compare two entire protein structures looking for local similarities, or one can use a pre-defined structural template, which represents the spatial arrangement of a local functional residue motif, to screen a structure.

Residue template local search methods

Local structural comparison refers to identifying a similar three-dimensional arrangement of a small set of residues or spatial sub-regions within the protein structure, possibly in the context of completely different protein structures (folds). In applying such algorithms, one can either compare two entire protein structures in search for local similarities or use a pre-defined structural template to screen a structure. A template represents the spatial arrangement of the residues involved in some biochemical function and can be regarded as a three-dimensional extension of the linear sequence motif idea. Often the specific arrangement/conformation of the residues is crucial to the performance of the function and remains strongly conserved. The various methods available for local structure comparison differ essentially in two aspects: the way the protein structure is represented and the computational strategy that is used to search for similarities. Structure arrangements or patterns range from three-dimensional shapes dissociated from the amino acids to a string of characters representing amino acids and their physical environment. The level of detail in the representation ranges from very approximate, to elaborate schemes that take into account the presence of different chemical groups along the amino acid side chains.

Several databases and search algorithms have been developed including Catalytic Site Atlas (CSA) [Porter et al. 2004], pdbFun [Ausiello et al. 2005], PDBSiteScan [Ivanisenko et al. 2004; Ivanisenko et al. 2005], SuMo [Jambon et al. 2005], pvSOAR [Binkowski et al. 2004], SARIG [Amitai et al. 2004], FEATURE [Wei and Altman 1998; Wei and Altman 2003], JESS [Barker and Thornton 2003], RIGOR [Kleywegt 1999] and PatchFinder [Nimrod et al. 2005]. The well-known PINTS [Stark and Russell 2003] allows comparison of a protein structure against a database of patterns or a PDB format pattern against a pattern database. As other examples, Phunctioner [Pazos and Sternberg 2004] extracts conserved residues and uses GO annotation as a core element of its assignment and validation, while SuMo [Jambon et al. 2005] uses a stereo-chemical group representation for residues arranged in triangles and graph theory to superimpose them for data base searching. FEATURE [Wei and Altman 1998] defines subdomains in the protein structure as a series of concentric spheres or as a three-dimensional cubic lattice, while seqFEATURE [Wei and Altman 2003] enables the creation of structure patterns from known sequence patterns. To overcome the difficulty that structural templates must be derived manually, methods for the automatic discovery of structural motifs characterizing a protein family have been developed [Bandyopadhyay et al. 2006; Polacco and Babbitt 2006; Wangikar et al. 2003].

Approaches based on physiochemical characteristics and structure calculations

In general, these methods are based on the observation that the functional patches of a protein have unique physicochemical features which set them apart from the protein surface and exhibit function. The aim of these methods is usually to predict either the location of a ligand-binding or active site. Numerous algorithms employ the notion that functional sites are usually located in clefts on the protein surface [Laskowski et al. 1996]. This basic idea is used either directly to predict the location of functional sites, or as a first step to identify candidate residues before further scoring procedures are applied. Methods for identifying cavities, clefts, pockets and surfaces in a protein include PASS [Brady and Stouten 2000], CASTp [Dundas et al. 2006], LIGSITEcsc [Huang and Schroeder 2006], VICE [Tripathi and Kellogg], SURFNET/SURFNET-ConSurf [Glaser et al. 2006; Laskowski 1995], BSAlign [Aung and Tong 2008], CAVER [Petrek et al. 2006], pvSPAR [Binkowski et al. 2004] and PocketPicker [Weisel et al. 2007], among others. Besides being located in clefts, active site residues are reported as being close to the centroid of the structure, having a destabilizing effect on the structure, interacting with a high number of residues of the same protein, having perturbed pKa values and inducing clusters in the electrostatic potential around the protein. All of these observations have been used to develop methods aimed at the inference of active site location from structure. Electrostatic calculations have also been used to predict DNA binding sites, combined with the analysis of the curvature of the molecular surface and the detection of specific structural motifs.

Combining multiple methods and multiple data sources

Since protein function is a multifaceted concept, its comprehensive prediction and characterization requires data from many sources. Recognizing the power and thoroughness in predictive multiplicity, several recent methods or applications, also referred to as ‘metasevers’ in some cases, have been developed to integrate information pertaining to function such as structure, sequence information, physiochemical features, and protein interaction data and also to provide a consensus view that can better identify the most likely functional predictions. This approach has been used by ProtFun [Jensen et al. 2003] which combines numerous different sequence-based methods to generate GO term predictions. InterPro [Hunter et al. 2009] integrates together predictive models or ‘signatures’ representing protein domains, families and functional sites from multiple, diverse databases and predicts the occurrence of functional domains, repeats and important sites. Another resource, ProFunc [Laskowski et al. 2005] uses varied sequence and structure-based methods, combined with the identification of active and binding sites and integrates them with interaction data and knowledge of genomic sequences to yield a comprehensive prediction summary of the most likely GO term-represented functions. ProKnow [Pal and Eisenberg 2005] considers structural features that are associated with specific functions in addition to sequence motifs, fold similarity, templates, and interaction data. Joined Assembly of Function Annotations [Friedberg et al. 2006], or JAFA, is a metaserver that surveys several function-prediction servers with a query sequence, returning a summary of predicted GO terms.

APPLICATION OF PROTEIN FUNCTION IN DRUG DISCOVERY INITIATIVES

One of the most important challenges for computational biology and drug discovery efforts has been to predict the function of previously uncharacterized proteins for which there is no known experimental three-dimensional structure [Betz et al. 2002; Chanda and Caldwell 2003; Ofran et al. 2005]. New protein sequences with the potential to be involved in important genetic and parasitic disease are being discovered at rapid pace and spurring the need to predict their structures in order to understand their function and investigate their potential as therapeutic targets. Despite significant, emerging advances in our understanding of the relationship between structure and function in this era of structural genomics, the identification of new drug targets and the successful development of potent and specific therapeutic drugs is still a slow and resource expensive process. However, with the remarkable research efforts by many into the development of accurate methods for protein function prediction, homology molecular modeling, virtual screening of chemical libraries by docking experiments, and new avenues for drug design and optimization, certain future success in drug discovery initiatives that implement these methods will lead to promising drug candidates more rapidly.

Having a comprehensive, functionally annotated sequence database cataloging of all the members of a given gene family in the human genome allows one to integrate a very different perspective into the drug-discovery process. The availability of the complete sets of related genes for a given target and a representative subset of protein structures allows one to build three-dimensional models for the entire protein family and to map the interactions of a given substrate/inhibitor and specific residues in the target even when detailed structural data is not available. Specificity predictions at the target level can also be applied to target selection. Knowing the sequences and structures of the target and those proteins that are physiochemically and/or structurally related to the target in the inhibitor-binding site can be essential in the evaluation of the target. Having this information in hand for all targets within a biological pathway would allow one to use the specificity prediction to guide target selection to the most appropriate and specific target in the pathway.

Optimization of a lead compound to a clinical development candidate involves iterative cycles to improve targeting and specificity. There is a significant need to understand fully how small molecules interact with other targets, outside of the target of therapeutic value. Knowing the detailed receptor-ligand interactions within a binding pocket and having a catalog of all the motifs, patterns, and functional residue subsets found in all sequences allows the prediction of the relatedness of various targets for broad arrays of inhibitors. With such model sets, a comprehensive evaluation of inhibitor modifications at certain positions can reveal interactions with increased specificity for the target over others. Furthermore, having the complete set of residues that provide key inhibitor interactions across a catalog of functionally related sequences enables the prediction of inhibitor specificity significantly earlier in the discovery process. Genomic data can then be used to focus inhibitor design towards areas of the novel target molecule that might facilitate engineering of inhibitor/ligand specificity.

We have witnessed significant changes to the pharmaceutical drug discovery process over the past few years. Protein function prediction can have an impact on target selection and on various stages of lead compound design [Betz et al. 2002] [Chanda and Caldwell 2003]. One change that is impacting strongly on the development process in its early stages is the shift from focusing on optimizing the chemistry of a small number of targets to focusing on the validation process of a few thousand druggable targets [Ofran et al. 2005]. With only a small fraction of proteins used as drug targets, the goal is to explore the space of druggable proteins and reveal the relationship between them and the chemistry space, and to use computational tools to address this challenge.

It is becoming apparent that an important goal in pharmacological targeting will be to identify and select disease modifying nodes or functional hubs that control or link several desired targets within a biochemical network. Functional knowledge about each potential target will be crucial to fully evaluating complex biochemical networks, their integrated protein functions, and their respective role in disease. With the poor coverage of all potential targets by direct experimental structures and experimental functional annotations, functional information from computational or in silico function prediction methods might be the most readily available means to provide critical functional information to inform such network-based targeting approaches. Protein function prediction will have an impact on adding novel gene products to the considered target space and on narrowing down the number of potential targets, as well, if applied in the early stages of discovery.

SUMMARY

To understand the function of whole proteomes in nature is one of the grand goals of molecular biology. The implications of this knowledge will have a tremendous impact on understanding the biochemical details of molecular processes, the molecular basis of diseases and the non-trivial relationships between structure and function. As the available protein structural and experimental data grow and computational methods to exploit this information improve, as is the strong case for emerging methods for protein function prediction, the derived knowledge will have positive consequences for future successful drug discovery initiatives. Here we have reviewed current approaches and programs for protein function prediction and discussed their integration into drug discovery initiatives. The development of such methods to annotate protein functional sites and their application to large protein functional families will be crucial to successfully utilizing the vast amounts of genomic sequence information available to drug discovery and development processes.