
==== Front
Nucleic Acids ResNucleic Acids ResnarNucleic Acids ResearchNucleic Acids Research0305-10481362-4962Oxford University Press 1713519210.1093/nar/gkl932ArticlesMiST: a microbial signal transduction database Ulrich Luke E. 13*Zhulin Igor B. 1231Joint Institute for Computational Sciences, The University of Tennessee–Oak Ridge National LaboratoryOak Ridge, TN 37831-6173, USA2Graduate School of Genome Science and Technology, The University of Tennessee–Oak Ridge National LaboratoryOak Ridge, TN 37831-6173, USA3Center for Bioinformatics and Computational Biology, School of Biology, Georgia Institute of TechnologyAtlanta, GA 30332-0230, USA*To whom correspondence should be addressed. Tel: +1 865 974 7687; Fax: +865 576 4368; Email: ulrichle@ornl.gov1 2007 28 11 2006 28 11 2006 35 Database issue D386 D390 15 8 2006 12 10 2006 13 10 2006 © 2006 The Author(s)2006This is an Open Access article distributed under the terms of the Creative Commons Attribution Non-Commercial License (http://creativecommons.org/licenses/by-nc/2.0/uk/) which permits unrestricted non-commercial use, distribution, and reproduction in any medium, provided the original work is properly cited.Signal transduction pathways control most cellular activities in living cells ranging from regulation of gene expression to fine-tuning enzymatic activity and controlling motile behavior in response to extracellular and intracellular signals. Because of their extreme sequence variability and extensive domain shuffling, signal transduction proteins are difficult to identify, and their current annotation in most leading databases is often incomplete or erroneous. To overcome this problem, we have developed the microbial signal transduction (MiST) database (), a comprehensive library of the signal transduction proteins from completely sequenced bacterial and archaeal genomes. By searching for domain profiles that implicate a particular protein as participating in signal transduction, we have systematically identified 69 270 two- and one-component proteins in 365 bacterial and archaeal genomes. We have designed a user-friendly website to access and browse the predicted signal transduction proteins within various organisms. Further capabilities include gene/protein sequence retrieval, visualized domain architectures, interactive chromosomal views for exploring gene neighborhood, advanced querying options and cross-species comparison. Newly available, complete genomes are loaded into the database each month. MiST is the only comprehensive and up-to-date electronic catalog of the signaling machinery in microbial genomes.
==== Body
INTRODUCTION
Microbial signal transduction (MiST) links environmental stimuli to specific adaptive cellular responses. Consequently, signal transduction pathways are vital to an organism's survival and function. In microorganisms, they control the majority of cellular functions including chemotaxis, respiration, development, osmoregulation, transport, metabolism, virulence, host-recognition and antibiotic resistance. Such signal transduction pathways behave as information processing circuits that link input (stimulus—nutrients, temperature, redox, etc.) and output (regulatory response—gene expression, enzyme activity, flagellar motor switch, etc.) events (1,2).

The detection of a signal (input) and coupling this with an adaptive cellular response (output) is common to all signal transduction systems; however, microorganisms employ diverse mechanisms for linking these events. These range from single-domain transducers to several interacting proteins and multi-protein complexes. The most widely recognized signaling systems are the so-called two-component systems that utilize protein phosphorylation as the fundamental signaling mechanism (2,3). The prototypical two-component system consists of two proteins: a membrane-bound, sensor histidine kinase and a cytoplasmic, response regulator. The sensor kinase detects environmental signals via one or more, amino-terminal sensory domains. Subsequently, the sensor kinase undergoes a conformational shift that results in an ATP-dependent autophosphorylation of a conserved histidine residue within its carboxy-terminal, transmitter domain. The cognate response regulator then catalyzes the transfer of this phosphoryl group to a conserved aspartate residue within its amino-terminal, receiver domain. Phosphorylation of the receiver domain activates the response regulator's output domain(s) and effects a particular adaptive response—typically regulation of gene expression at the transcriptional level (1).

Despite the importance of two-component systems, most signal transduction events in prokaryotes are carried out by one-component systems that consist of a single protein molecule containing both input and output domains but lacking the phosphotransfer domains typical of two-component systems (4). Many one-component systems have been extensively studied including the LacI lactose operon repressor (5) and the catabolite activator, CAP, of Escherichia coli (6); the arginine catabolism regulator, RocR, of Bacillus subtilis (7); and the quorum-sensing regulator, TraR, of Agrobacterium tumefaciens (8). These proteins bind ligands via their input domain and regulate gene expression with their output domain. Most one-component systems typically consist of at least one input and one output domain, yet in some cases a single domain is capable of both detecting a ligand and producing a regulatory response. For example, both the metalloregulators, CzrA of Staphylococus aureus (9) and CmtR of Mycobacterium tuberculosis (10), consist of a single domain, which directly binds DNA in response to metal ligands. One-component systems are more widely distributed among bacteria and archaea, and display a greater diversity of domains than two-component systems (4).

Signal transduction proteins are highly modular with diverse and mosaic domain architectures. The domains comprising these proteins may be categorized into four major types according to their function: input, transmitter, receiver and output (1). Relatively few input domains specific to signal transduction have been characterized and include the following: PAS (11,12), GAF (13), Cache (14), CHASE (15,16), CHASE2 through CHASE6 (17), NIT (18) and 4HB_MCP (19). These domains exhibit extreme sequence variability due to the broad range of signals they detect and are the least conserved of all the signaling modules. In contrast, transmitter and receiver domains display remarkable sequence conservation, which reflects the conserved phosphorylation reaction between these domain types that links sensor kinases and response regulators (20). The most commonly found output domains are DNA-binding helix–turn–helix (HTH) domains because the predominant adaptive response of two- and one-component systems is regulation of gene expression (4). Several novel output domains have been recently described in response regulators, which implicate these systems in other types of control, such as the regulation of enzyme activity. These include adenylate and diguanylate cyclases, c-di-GMP-phosphodiesterase, phosphohydrolase and other related domains (21–23). Output domains are more conserved than input domains yet still considerably divergent due to their moderate number of regulatory roles (e.g. gene-regulation, protein–protein interactions, etc.).

One of the first efforts to catalog signal transduction proteins was the SENTRA database (24). It contains information on classical two-component systems and a few other signaling systems that interact via phosphorylation or methylation reactions. Currently, SENTRA includes information on two-component signal transduction proteins for 43 genomes. Several other databases attempt to document the signal transduction machinery within various genomes; however, these projects are usually a part of a larger initiative and therefore limited in scope and/or accuracy. For example, the KEGG project (25) focuses on deriving higher-order information (e.g. pathways) from genomic data and has successfully mapped metabolic pathways across multiple genomes, yet the curators have recently begun mapping some regulatory pathways including signal transduction. KEGG transfers pathway information among genomes by first defining reference pathways based on scientific literature and then computationally extending these networks across genomic data via orthologous relationships established from sequence similarity searches and the positional correlation of genes. This approach is constrained to experimentally defined pathways and does not provide a complete record of the signal transduction repertoire. In addition, KEGG defines orthologous relationships using bi-directional best hits with respect to the entire protein sequence—this poorly ascertains related signal transduction proteins, given their highly modular nature and extreme sequence variability. Model organism databases such as the E.coli EcoCyc (26) and B.subtilis DBTBS (27) often contain information about signal transduction for their associated organisms, since these are primarily restricted to literature-based curation efforts and therefore limited in coverage (albeit with high accuracy). Finally, as one-component systems have only recently been recognized as a major part of signal transduction (4), most database resources when describing signal transduction contain only information on two-component systems.

We have developed a novel resource, the MiST database that contains a comprehensive compilation of the signal transduction proteins within bacterial genomes. As these proteins are modular in nature, our approach focuses on determining signal transduction proteins from a protein's domain composition, which we derive using the HMMER software (28). From the Pfam (29) and SMART (30) domain libraries, we have designated a set of profile hidden Markov models (HMM) that represent signaling domains (e.g. HATPase_c transmitter domain or DNA-binding output domains) and implicate a role in signal transduction. We classify a protein as belonging to signal transduction if it contains one or more of these specific signaling domains. In this manner, we systematically produce the repertoire of signal transduction proteins for microbial genomes (or any other collection of protein sequences). To our knowledge, signal transduction is the only current resource on signal transduction in prokaryotes that provides a thorough catalog of both two- and one-component systems within bacterial genomes.

HIGH-THROUGHPUT IDENTIFICATION OF SIGNAL TRANSDUCTION PROTEINS
The biological function of a signal transduction protein is determined by specific signaling domains that detect an environmental cue (input), mediate protein–protein communication (transmitter, receiver), or initiate a cellular response (output). Thus, domains that perform these roles serve as markers and facilitate the identification of signal transduction proteins given two conditions: (i) an adequate mechanism for detecting domains in an amino acid sequence and (ii) a comprehensive set of domains known or predicted to participate in signal transduction. Protein domains may be robustly represented by profile HMMs (28), which statistically model the primary structure of homologous domain sequences and enable their rapid identification from a protein sequence with tools such as the HMMER software package. Furthermore, domain profiles are more sensitive than pairwise sequence comparisons and typically contain manually-curated score thresholds from which the significance and membership of domain matches may be automatically evaluated. To meet the second requirement, we selected 133 signaling domains (Supplementary Table S1) from the Pfam (29) and SMART databases (30) based on the following: (i) known domain function (information from Pfam, SMART, InterPro (31), COG (32) resources and analysis of literature on signal transduction), and (ii) predicted domain function based on the association with other signaling domains. This approach has been recently described in detail (4).

MiST incorporates a straightforward and systematic methodology for identifying signal transduction proteins within completely sequenced bacterial genomes (Figure 1). Perl scripts download all available genomes from NCBI in the XML (eXtensible Markup Language) format and load this information into the database. Each XML file contains the full RefSeq annotation (33) for a genome, which includes its associated nucleotide data, genes and translated proteins. Secondly, we predict the domain architecture for each protein including signal peptides, transmembrane regions, coiled-coils and low-complexity segments. Finally, we identify the set of putative signal transduction proteins by scanning each protein for the presence of transmitter, receiver or output signaling domains (see above). Input domains are often found in pathways other than signal transduction (e.g. metabolic pathways) and therefore proteins identified solely from an input domain are not classified as belonging to signal transduction. During this process, we also filter out various domain combinations that indicate a role other than signal transduction. For example, due to structural similarities, DNA topoisomerase IV often contains a predicted HATPase_c transmitter domain (Pfam accession no. PF02518, SMART accession no. SM00387) in addition to other domains suggestive of topoisomerase activity (DNA_gyraseB and DNA_gyraseB_C), yet this protein is not involved in signal transduction. This pipeline may be executed on a regular schedule such that new genomes are seamlessly integrated into the database and signal transduction blueprints for these organisms are automatically generated. MiST is updated on a monthly basis to maintain a current record of the signal transduction repertoire within newly sequenced genomes.

Figure 1 Overview of the high-throughput process for identifying signal transduction proteins. First, all complete, bacterial genomes are downloaded from NCBI and loaded into the MiST database. Second, the complete domain architecture of each protein is predicted. Finally, a protein is classified as belonging to signal transduction if it contains at least one transmitter, receiver, or output signaling domain.

The MiST database is implemented using the PostgreSQL () version 8.1.3 relational database management system on the Gentoo distribution () of the Linux/GNU operating system. All protein domain architectures are derived on our 34-node Linux cluster using HMMER version 2.3.2 (28) and the Pfam version 19.0 (29) and SMART version 5.0 (30) domain libraries. Every year, the domain architectures and signal transduction predictions will be updated based on the latest releases of the Pfam and SMART databases. Phobius version 1.01 (34) is used to predict signal peptides and transmembrane regions. Coiled-coils and regions of low-complexity are predicted with the COILS version 2.2 (35) and SEG (36) programs, respectively. Custom Perl scripts handle program execution, formatting and interacting with the PostgreSQL subsystem. PHP and CGI scripts running on the Apache web server version 2.0.58 () deliver web content to users via the Internet.

DATABASE CONTENTS
As of August 2006, MiST version 1.0 contains 69 720 predicted signal transduction proteins—22 868 proteins that belong to two-component regulatory systems and 46 402 one-component regulators—from 365 bacterial and archaeal genomes. Excluding chemotaxis proteins, there are almost equal numbers of sensor kinases and response regulators: 9317 (51.8%) and 9025 (49.2%), respectively. These display a strong, positive, linear relationship (Figure 2). Using several chemotaxis-related domains from the Pfam database, we found 4526 chemotaxis proteins, out of which 328 are putative Class II histidine kinases (CheA-like) (37). MiST contains more than twice as many one-component proteins as two-component proteins. When considering the number of systems based on their output activity, the number of one-component systems exceeds two-component systems by a factor of five. This is slightly higher than our initial analysis of the signaling transduction proteins derived from 145 prokaryotic genomes (4). The taxonomic distribution of two- and one-component systems within the MiST database is given in Supplmentary Table S2.

Figure 2 Scatterplot showing the linear, positive relationship (R = 0.92) between the number of predicted histidine kinases and response regulators within representative bacterial genomes (excluding chemotaxis proteins).

In addition to specific information about the signal transduction repertoire for each microbial genome, MiST stores the large amount of pre-computed data used to identify this class of proteins. This includes any predicted Pfam and SMART domains, protein secondary features (e.g. transmembrane regions, signal peptides, coiled-coils, low-complexity regions), chromosomal position of the corresponding gene, statistical information and the complete RefSeq annotation. We have also indexed the NCBI non-redundant database to provide cross-referencing support. Therefore, while MiST focuses primarily on the classification and analysis of signal transduction proteins, this information along with its web interface (see below) enables users to effectively use MiST as a general resource for exploring microbial genomes.

WEB INTERFACE
To provide an effective means for exploring and utilizing the MiST database, we designed and implemented a user-friendly web interface oriented toward experimental biologists (). A separate page for each microbe displays both general and specific information about its signal transduction repertoire including: descriptive data about the genome and proteome (e.g. size, replicons, etc.), a graphical representation of its signal transduction profile, querying options and a summary table of the signal transduction protein counts for each replicon (Figure 3). The profile represents an overall view of the number of signaling domains separated into functional categories. For example, E.coli O157:H7 contains 272 DNA-binding domains, clearly indicating a substantial degree of gene-regulation. Each of the headings and signal transduction counts within the summary table are linked to the particular slice of data it represents. Clicking on a replicon displays the signal transduction proteins associated with that replicon. Selecting the heading ‘Two-component proteins’ displays all the predicted two-component proteins for this organism. Clicking on a signal transduction count will show either the two- or one-component proteins of a particular replicon. The resulting list for each protein contains GenBank Identifier (GI) number, RefSeq description, set of input and/or output domains and visualized domain architecture. Following an individual protein link provides an expanded overview of that particular protein and its associated gene. In addition to basic sequence data, this page contains the full protein/gene RefSeq annotation, domain architectures and graphical view of the gene neighborhood. The domain architecture section presents a linear representation of the protein's domain architecture and other secondary features (e.g. transmembrane regions, low-complexity regions). The chromosome view reveals neighboring genes as arrows pointed toward their transcriptional direction. Each arrow is hyperlinked to its respective gene/protein page for evaluating neighboring genes.

Figure 3 MiST screen-shot: the E.coli 0157:H7 organism page. Clickable links are shown in blue. Queries can be made with various search types and terms (upper right corner).

Relevant data may be accessed in several ways: (i) choosing an organism from the complete list of bacteria in MiST, (ii) searching for a particular organism, or (iii) searching by a GenBank GI number. In addition, multiple species may be selected and simultaneously searched using a sophisticated, taxonomy-driven, selection tool. Users may query for particular Pfam or SMART domains, keyword, locus, GI number or internal MiST identifier.

CONCLUSIONS AND FUTURE WORK
Using sensitive profile HMM searches and utilizing the available knowledge on microbial signal transduction from the literature and on-line resources, we were able to identify more than 69 000 proteins that are predicted to comprise signal transduction networks. We have developed a comprehensive database, which captures, stores and makes this information available to the scientific community. We expect MiST to become a widely-used resource significantly enhancing our knowledge within this important area of research as well as improving current genome annotation and contributing to important medical applications, such as antimicrobial drug development. Future MiST development will focus on incorporating user- and literature-based curation, GO ontology driven annotations (38), addition of new domain models, mapping functional residues, pathway support and improvements to the web interface.

SUPPLEMENTARY DATA
Supplementary Data are available at NAR online.

We thank the BLAST/STIM research community for many valuable discussions and suggestions on the MiST scope and development. We are grateful to Jan Lukens and ‘Friends of the Zhulin lab’ for their contribution toward our Linux cluster. This work was supported in part by a grant GM072285 from the National Institutes of Health to I.B.Z. and funds provided by the UT-ORNL Joint Institute for Computational Sciences. Funding to pay the Open Access publication charges for this article was provided by NIH grant GM072285.

Conflict of interest statement. None declared.
==== Refs
REFERENCES
1 Parkinson J.S.  Kofoid E.C.   Communication modules in bacterial signaling proteins Annu. Rev. Genet. 1992 26 71 112 1482126 
2 Stock A.M.  Robinson V.L.  Goudreau P.N.   Two-component signal transduction Annu. Rev. Biochem. 2000 69 183 215 10966457 
3 Hoch J.A.  Silhavy T.A.   Two-component signal transduction 1995 Washington, DC, USA ASM Press 
4 Ulrich L.E.  Koonin E.V.  Zhulin I.B.   One-component systems dominate signal transduction in prokaryotes Trends Microbiol. 2005 13 52 56 15680762 
5 Lewis M.  Chang G.  Horton N.C.  Kercher M.A.  Pace H.C.  Schumacher M.A.  Brennan R.G.  Lu P.   Crystal structure of the lactose operon repressor and its complexes with DNA and inducer Science 1996 271 1247 1254 8638105 
6 Kolb A.  Busby S.  Buc H.  Garges S.  Adhya S.   Transcriptional regulation by cAMP and its receptor protein Annu. Rev. Biochem. 1993 62 749 795 8394684 
7 Calogero S.  Gardan R.  Glaser P.  Schweizer J.  Rapoport G.  Debarbouille M.   RocR, a novel regulatory protein controlling arginine utilization in Bacillus subtilis , belongs to the NtrC/NifA family of transcriptional activators J. Bacteriol. 1994 176 1234 1241 8113162 
8 Vannini A.  Volpari C.  Gargioli C.  Muraglia E.  Cortese R.  De Francesco R.  Neddermann P.  Marco S.D.   The crystal structure of the quorum sensing protein TraR bound to its autoinducer and target DNA EMBO J. 2002 21 4393 4401 12198141 
9 Pennella M.A.  Arunkumar A.I.  Giedroc D.P.   Individual metal ligands play distinct functional roles in the zinc sensor Staphylococcus aureus  CzrA J. Mol. Biol. 2006 356 1124 1136 16406068 
10 Cavet J.S.  Graham A.I.  Meng W.  Robinson N.J.   A cadmium-lead-sensing ArsR-SmtB repressor with novel sensory sites. Complementary metal discrimination by NmtR and CmtR in a common cytosol J. Biol. Chem. 2003 278 44560 44566 12939264 
11 Ponting C.P.  Aravind L.   PAS: a multifunctional domain family comes to light Curr. Biol. 1997 7 R674 R677 9382818 
12 Taylor B.L.  Zhulin I.B.   PAS domains: internal sensors of oxygen, redox potential, and light Microbiol. Mol. Biol. Rev. 1999 63 479 506 10357859 
13 Aravind L.  Ponting C.P.   The GAF domain: an evolutionary link between diverse phototransducing proteins Trends Biochem. Sci. 1997 22 458 459 9433123 
14 Anantharaman V.  Aravind L.   Cache—a signaling domain common to animal Ca(2+)-channel subunits and a class of prokaryotic chemotaxis receptors Trends Biochem. Sci. 2000 25 535 537 11084361 
15 Anantharaman V.  Aravind L.   The CHASE domain: a predicted ligand-binding module in plant cytokinin receptors and other eukaryotic and bacterial receptors Trends Biochem. Sci. 2001 26 579 582 11590000 
16 Mougel C.  Zhulin I.B.   CHASE: an extracellular sensing domain common to transmembrane receptors from prokaryotes, lower eukaryotes and plants Trends Biochem. Sci. 2001 26 582 584 11590001 
17 Zhulin I.B.  Nikolskaya A.N.  Galperin M.Y.   Common extracellular sensory domains in transmembrane receptors for diverse signal transduction pathways in bacteria and archaea J. Bacteriol. 2003 185 285 294 12486065 
18 Shu C.J.  Ulrich L.E.  Zhulin I.B.   The NIT domain: a predicted nitrate-responsive module in bacterial sensory receptors Trends Biochem. Sci. 2003 28 121 124 12633990 
19 Ulrich L.E.  Zhulin I.B.   Four-helix bundle: a ubiquitous sensory module in prokaryotic signal transduction Bioinformatics 2005 21 Suppl. 3 iii45 iii48 16306392 
20 Grebe T.W.  Stock J.B.   The histidine protein kinase superfamily Adv. Microb. Physiol. 1999 41 139 227 10500846 
21 Galperin M.Y.  Nikolskaya A.N.  Koonin E.V.   Novel domains of the prokaryotic two-component signal transduction systems FEMS Microbiol. Lett. 2001 203 11 21 11557134 
22 Pei J.  Grishin N.V.   GGDEF domain is homologous to adenylyl cyclase Proteins 2001 42 210 216 11119645 
23 Shu C.J.  Zhulin I.B.   ANTAR: an RNA-binding domain in transcription antitermination regulatory proteins Trends Biochem. Sci. 2002 27 3 5 11796212 
24 Maltsev N.  Marland E.  Yu G.X.  Bhatnagar S.  Lusk R.   SENTRA, a database of signal transduction proteins Nucleic Acids Res. 2002 30 349 350 11752334 
25 Kanehisa M.  Goto S.  Hattori M.  Aoki-Kinoshita K.F.  Itoh M.  Kawashima S.  Katayama T.  Araki M.  Hirakawa M.   From genomics to chemical genomics: new developments in KEGG Nucleic Acids Res. 2006 34 D354 D357 16381885 
26 Keseler I.M.  Collado-Vides J.  Gama-Castro S.  Ingraham J.  Paley S.  Paulsen I.T.  Peralta-Gil M.  Karp P.D.   EcoCyc: a comprehensive database resource for Escherichia coli  Nucleic Acids Res. 2005 33 D334 D337 15608210 
27 Makita Y.  Nakao M.  Ogasawara N.  Nakai K.   DBTBS: database of transcriptional regulation in Bacillus subtilis  and its contribution to comparative genomics Nucleic Acids Res. 2004 32 D75 D77 14681362 
28 Eddy S.R.   Profile hidden Markov models Bioinformatics 1998 14 755 763 9918945 
29 Finn R.D.  Mistry J.  Schuster-Bockler B.  Griffiths-Jones S.  Hollich V.  Lassmann T.  Moxon S.  Marshall M.  Khanna A.  Durbin R.    Pfam: clans, web tools and services Nucleic Acids Res. 2006 34 D247 D251 16381856 
30 Letunic I.  Copley R.R.  Pils B.  Pinkert S.  Schultz J.  Bork P.   SMART 5: domains in the context of genomes and networks Nucleic Acids Res. 2006 34 D257 D260 16381859 
31 Mulder N.J.  Apweiler R.  Attwood T.K.  Bairoch A.  Bateman A.  Binns D.  Bradley P.  Bork P.  Bucher P.  Cerutti L.    InterPro, progress and status in 2005 Nucleic Acids Res. 2005 33 D201 D205 15608177 
32 Tatusov R.L.  Fedorova N.D.  Jackson J.D.  Jacobs A.R.  Kiryutin B.  Koonin E.V.  Krylov D.M.  Mazumder R.  Mekhedov S.L.  Nikolskaya A.N.    The COG database: an updated version includes eukaryotes BMC Bioinformatics 2003 4 41 12969510 
33 Pruitt K.D.  Tatusova T.  Maglott D.R.   NCBI Reference Sequence (RefSeq): a curated non-redundant sequence database of genomes, transcripts and proteins Nucleic Acids Res. 2005 33 D501 D504 15608248 
34 Kall L.  Krogh A.  Sonnhammer E.L.   A combined transmembrane topology and signal peptide prediction method J. Mol. Biol. 2004 338 1027 1036 15111065 
35 Lupas A.   Prediction and analysis of coiled-coil structures Methods Enzymol. 1996 266 513 525 8743703 
36 Wootton J.C.  Federhen S.   Statistics of local complexity in amino acid sequences and sequence databases Comput. Chem. 1993 17 149 163 
37 Bilwes A.M.  Alex L.A.  Crane B.R.  Simon M.I.   Structure of CheA, a signal-transducing histidine kinase Cell 1999 96 131 141 9989504 
38 Ashburner M.  Ball C.A.  Blake J.A.  Botstein D.  Butler H.  Cherry J.M.  Davis A.P.  Dolinski K.  Dwight S.S.  Eppig J.T.    Gene ontology: tool for the unification of biology. The Gene Ontology Consortium Nature Genet. 2000 25 25 29 10802651

