Consolidating the set of known human protein-protein interactions in preparation for large-scale mapping of the human interactome
© Marcotte et al.; licensee BioMed Central Ltd. 2005
Received: 20 December 2004
Accepted: 11 March 2005
Published: 15 April 2005
Extensive protein interaction maps are being constructed for yeast, worm, and fly to ask how the proteins organize into pathways and systems, but no such genome-wide interaction map yet exists for the set of human proteins. To prepare for studies in humans, we wished to establish tests for the accuracy of future interaction assays and to consolidate the known interactions among human proteins.
We established two tests of the accuracy of human protein interaction datasets and measured the relative accuracy of the available data. We then developed and applied natural language processing and literature-mining algorithms to recover from Medline abstracts 6,580 interactions among 3,737 human proteins. A three-part algorithm was used: first, human protein names were identified in Medline abstracts using a discriminator based on conditional random fields, then interactions were identified by the co-occurrence of protein names across the set of Medline abstracts, filtering the interactions with a Bayesian classifier to enrich for legitimate physical interactions. These mined interactions were combined with existing interaction data to obtain a network of 31,609 interactions among 7,748 human proteins, accurate to the same degree as the existing datasets.
These interactions and the accuracy benchmarks will aid interpretation of current functional genomics data and provide a basis for determining the quality of future large-scale human protein interaction assays. Projecting from the approximately 15 interactions per protein in the best-sampled interaction set to the estimated 25,000 human genes implies more than 375,000 interactions in the complete human protein interaction network. This set therefore represents no more than 10% of the complete network.
The past few years have seen a tremendous development of functional genomics technologies. In particular, the yeast proteome has been the subject of considerable effort, including genome-wide protein interaction assays using yeast two-hybrid technology [1, 2], affinity chromatography/mass spectrometry [3, 4], synthetic lethal assays [5, 6], and genome context methods [7–10]. Success in these areas, even given the limited accuracy of these technologies [11–15], has led to the application of the yeast two-hybrid method for the fly  and the worm proteomes , providing initial steps toward maps of the fly and worm interactomes.
Only minimal progress has been made with respect to the human proteome. The existing protein interaction data are largely composed of small-scale experiments collected in the BIND  and DIP  databases, as well as a set of approximately 12,000 interactions recovered by manual curation from Medline articles  and interactions transferred from other organisms on the basis of orthology . The Reactome database  has around 11,000 interactions  that have been manually entered from articles focusing on core cellular pathways. Large-scale interaction assays among human proteins have yet to be performed, although a medium-scale map was created for the purified TNFα/NFκB protein complex  and the proteins involved in the human Smad signaling pathway . This situation is in stark contrast to the abundant data available for yeast and calls for the application of high-throughput interaction assays for mapping the human protein interaction network.
One lesson from the yeast interactome research is clear: it is critical that such upcoming interaction assays be accompanied by measured error rates, without which the utility and interpretability of the data is jeopardized. To establish a basis for future interaction mapping we sought to consolidate existing human protein interaction data and to establish quantitative tests of data accuracy. We also sought to use data-mining approaches to extract additional known interactions from Medline abstracts to add to the existing interactions.
Most of the current biological knowledge can be retrieved from the Medline database, which now has records from more than 4,800 journals accounting for around 15 million articles. These citations contain thousands of experimentally recorded protein interactions. However, retrieving these data manually is made difficult by the large number of articles, all lacking formal structure. Automated extraction of information would be preferable, and therefore, mining data from Medline abstracts is a growing field [26–29].
In this paper, we present two quantitative tests (benchmarks) of the accuracy of large-scale human protein interaction assays, test the existing sets of interaction data for their relative accuracy, then apply these benchmarks in order to recover protein interactions from the approximately 750,000 Medline abstracts that concern human biology, resulting in a set of 6,580 interactions between 3,737 proteins of accuracy comparable to manual extraction. Combination of the interaction data creates a consolidated set of 31,609 interactions between 7,748 human proteins. On the basis of this initial set of interactions, we estimate the scale of the human interactome.
Assembling existing public protein interaction data
The initial list of the interactions and proteins represented in each of the existing human protein interaction datasets with total interactions, unique self-interactions and unique non-self interactions
Total interactions (number of proteins)
Unique self (A-A) interactions (number of proteins)
Unique (A-B) interactions (number of proteins)
Orthology transfer (all)
Orthology transfer (core)
Benchmarking of protein interaction data
To measure the relative accuracy of each protein interaction dataset, we established two benchmarks of interaction accuracy, one based on shared protein function and the other based on previously known interactions. First, we constructed a benchmark in which we tested the extent to which interaction partners in a dataset shared annotation, a measure previously shown to correlate with the accuracy of functional genomics datasets [13, 14, 21]. We used the functional annotations listed in the Kyoto Encyclopedia of Genes and Genomes (KEGG)  and Gene Ontology (GO)  annotation databases. These databases provide specific pathway and biological process annotations for approximately 7,500 human genes, assigning human genes into 155 KEGG pathways (at the lowest level of KEGG) and 1,356 GO pathways (at level 8 of the GO biological process annotation). KEGG and GO annotations were combined into a single composite functional annotation set, which was then split into independent testing and training sets by randomly assigning annotated genes into the two categories (3,792 and 3,809 annotated genes respectively). For the second benchmark based on known physical interactions, we assembled the human protein interactions from Reactome and BIND, a set of 11,425 interactions between 1,710 proteins. Each benchmark therefore consists of a set of binary relations between proteins, either based on proteins sharing annotation or physically interacting. Generally speaking, we expect more accurate protein interaction datasets to be more enriched in these protein pairs. More specifically, we expect true physical interactions to score highly on both tests, while non-physical or indirect associations, such as genetic associations, should score highly on the functional, but not the physical interaction, test.
For both benchmarks, the scoring scheme for measuring interaction set accuracy is in the form of a log odds ratio of gene pairs either sharing annotations or physically interacting. To evaluate a dataset, we calculate a log likelihood ratio (LLR) as:
where P(D|I) and P(D|~I) are the probability of observing the data (D) conditioned on the genes sharing benchmark associations (I) and not sharing benchmark associations (~I). By Bayes theorem, this equation can be rewritten as:
where P(I|D) and P(~I|D) are the frequencies of interactions observed in the given dataset (D) between annotated genes sharing benchmark associations (I) and not sharing associations (~I), respectively, while P(I) and P(~I) represent the prior expectations (the total frequencies of all benchmark genes sharing the same associations and not sharing associations, respectively). This latter version of the equation is simpler to compute. A score of zero indicates interaction partners in the dataset being tested are no more likely than random to belong to the same pathway or to interact; higher scores indicate a more accurate dataset.
Among the literature-derived interactions (Reactome, BIND, HPRD), a total of 17,098 unique interactions occur in the public datasets. Testing the existing protein interaction data on the function benchmark reveals that Reactome has the highest accuracy (LLR = 3.8), followed by BIND (LLR = 2.9), HPRD (LLR = 2.1), core orthology-inferred interactions (LLR = 2.1) and the non-core orthology-inferred interaction (LLR = 1.1). The two most accurate datasets, Reactome and BIND, form the basis of the protein interaction-based benchmark. Testing the remaining datasets on this benchmark (that is, for their consistency with these accurate protein interaction datasets) reveals a similar ranking in the remaining data. Core orthology-inferred interactions are the most accurate (LLR = 5.0), followed by HPRD (LLR = 3.7) and non-core orthology inferred interactions (LLR = 3.7).
Recognizing protein names with a conditional random field (CRF) algorithm
To expand the list of human interactions, we turned to literature mining. We adopted the strategy of separately identifying the protein names in the abstracts and then matching up the interacting protein partners. This process was made difficult by the fact that unlike other organisms, such as yeast or Escherichia coli, the human genes have no standardized naming convention, and thus present one of the hardest sets of gene/protein names to extract. For example, human proteins may be named with typical English words, such as 'light', 'map', 'complement', and 'Sonic Hedgehog'. Names may be alphanumeric, may include Greek or Roman letters, may be case sensitive, and may be composed of multiple words. Names are frequently sub-strings of each other, such as 'epidermal growth factor' and 'epidermal growth factor receptor', which refer to two distinct proteins. It is therefore necessary that an information-extraction algorithm be specifically trained to extract gene and protein names accurately.
Extracting functional interactions via co-citation analysis
The co-citation algorithm is remarkably robust to variations in the minimal accuracy with which the protein names are identified by the CRF algorithm (Figure 3b). This robustness is presumably due to the fact that co-citation requires proteins to be named repeatedly across many abstracts, thereby tolerating occasional errors in the name extraction process. With a threshold on the estimated extraction probability of 80% (as computed by the CRF model) in the protein name identification, around 15,000 interactions are extracted with the co-citation approach that score comparably or better on the independent functional annotation test benchmark than the manually extracted interactions from HPRD, which serves to establish a minimal threshold for our mined interactions.
A comparison of the contributions of each dataset to the composite human protein interaction map, with network properties of each of the datasets
Number of interactions
Number of proteins
Orthology inferred (core)
The ID-Serve database of annotation and interactions
We have incorporated the results of this analysis into a web-based server , which can be queried for interactions of specific proteins. Genes are cross-listed under a variety of naming conventions, including LocusLink/EntrezGene, RefSeq, and Swiss-Prot, and are accompanied by links to other databases and GO and KEGG functional annotations. Protein interactions derived from the co-citation/Bayesian analysis are hyperlinked to the co-citing Medline abstracts, where they can be directly manually verified.
Features of the network
A quantitative analysis of the network clustering and connectivity distribution (reviewed in Barabasi and Oltvai ) is presented in Table 2. The clustering coefficient (<C>) captures the modularity of the network. A comparison of our final network (<C> = 0.24) with 10 randomly generated networks with the same number of interactions and proteins (<C> = 9 × 10-3 ± 3 × 10-5) shows the clustering in the human protein interaction network is considerably above that expected at random, in spite of the incompleteness of the network. The 'degree' of the network is defined as the average number of links per protein and captures the connectivity of the network. Except for Reactome, each of the datasets indicated in Table 2 show low connectivity. The combined network is intermediate in both connectivity and modularity. Projecting from the approximately 15 interactions per protein in the best sampled interaction dataset (Reactome) to the 25,000 or so estimated in the human genome  implies more than 375,000 interactions in the complete human protein interaction network. Note that any overestimates in the average number of interactions per protein will be counterbalanced by the effect of alternative splicing in increasing the number of actual proteins, making this estimate at least a reasonable ballpark estimate. The current set of interactions therefore represents no more than 10% of the complete network.
Advantages of the log likelihood benchmarks
A good accuracy measure is of tremendous importance, impacting on the reliability of all downstream analysis. The log likelihood analysis eases comparison and assessment of diverse datasets. The score indicates the probability that the identified interactions are correct based on enrichment of positive interactions over background expectations. Note that this approach is distinct from simply measuring the intersection with the benchmark associations - because enrichment of positive to negative associations is measured, rather than just recovery of positive associations, even datasets with small intersections to the benchmark set can be evaluated for accuracy. Note also that the benchmarks themselves are not likely to be 100% correct - protein annotations are subjectively assigned, many proteins belong to multiple pathways, and even hand-curated protein interaction data can be mis-entered. Nonetheless, the log likelihood framework is tolerant of errors and merely requires that the benchmark data are generally correct among true interaction partners. Figure 4a shows the accuracy of each of the datasets. While the existing datasets have a single accuracy value, the mined interactions can be adjusted for accuracy based on the CRF threshold and the co-citation probabilities. New datasets can be incorporated using the log likelihood scoring scheme, and the ultimate strength of these benchmarks will be their utility in integrating data from diverse experiments .
Shortcomings and strengths of literature mining via the co-citation/Bayesian classifier approach
From our previous work , we realized that directly identifying protein interactions would be a difficult task if we were unable to differentiate proteins and genes from the rest of the text. We therefore concentrated on building protein name extractors and interaction extractors in parallel so that the results of the former analysis could be fed into the latter.
Crucial to this process was the creation of a high-quality dictionary of human protein names and synonyms with mappings back to database entries. We therefore decided to start by creating a set of unambiguous gene names along with their synonyms that could all be mapped to a single unified gene identifier (LocusLink identifiers, now maintained through EntrezGene). The dictionary had to have very few spurious entries to ensure minimal false positives. The resulting ID-Serve database captures the various identifiers for a given gene and creates a repository for the retrieval of these genes along with their mined interactions. Building on this dictionary, the CRF algorithm then analyzed the context in which likely protein names appeared in order to identify the protein names more accurately. In the approach we describe, protein interaction partners are identified from among these protein names by a filtered version of co-citation.
The co-citation approach [14, 26, 40] calculates the random probability of co-occurrence of two protein names. The assumption is that if the co-citation is statistically unlikely under the random model, then there is a true underlying reason for the proteins to be co-cited - that is, they are interacting at either the functional, pathway level, or are co-localized or physically interact. The method has both advantages and disadvantages. It does not extract all interactions, but only those with statistically significant co-citations. By using the Bayesian estimator  we enrich further for physical interactions, but at the expense of coverage. Among the disadvantages are that the algorithm enriches for certain types of errors (for example, 'A does not interact with B', dictionary errors leading to synonyms being wrongly enriched, and so on). However, we feel the advantages outweigh the disadvantages: In particular, the probabilistic ranking, combined with the Bayesian filter, minimizes systematic errors, and at the left side of Figure 4b, it can be seen that errors in the co-citation data are no more extensive than errors introduced in transferring annotation from other organisms, or those errors introduced by human curators reading Medline abstracts. The method is easily applied, and currently outperforms other publicly available protein interaction extraction algorithms [34, 35]. Finally, the precise nature of the interaction can be directly checked from the linked Medline abstracts. Thus, the mined interactions will be ideal for manual validation by curators of protein interaction databases (for example, DIP and BIND).
In conclusion, to prepare for attempts to map the set of human protein interactions we sought to consolidate known interactions and to establish measures of accuracy that are useful for the evaluation and integration of upcoming datasets. We established two benchmarks for assessing the quality of large-scale human protein interaction datasets, providing quantitative measures useful for the testing and integration of interaction data. Using these benchmarks, along with available and mined interactions, we assembled an integrated dataset of 31,609 interactions between 7,748 human proteins, forming a framework for the interpretation of human functional genomics data. These data are collected in the ID-Serve database , which can be queried for protein interactions and their corresponding Medline citations. We estimate these interactions form less than 10% of the human interactome, setting the stage for future efforts to map the complete human network of protein interactions.
Materials and methods
Identification of human protein names and interactions in Medline abstracts
The training datasets used for the literature mining are as in . The dictionary of human protein names was assembled from the LocusLink and Swiss-Prot databases by manually curating the gene names and synonyms (87,723 synonyms between 18,879 unique gene names) to remove genes that were referred to as 'hypothetical' or 'probable' and to omit entries that referred to more than one protein identifier. From the Medline database of approximately 11 million abstracts (1951-2002) we retrieved 753,459 abstracts containing the word 'human' either in the title or the text to use as our corpus for extracting protein interactions.
We have previously described  effective protein and gene name tagging using an algorithm based on maximum entropy. Conditional random fields (CRF)  are new types of probabilistic models that preserve all the advantages of maximum entropy models and at the same time avoid the label bias problem by allowing a sequence of tagging decisions to compete against each other in a global probabilistic model. In this paper, we show that CRF outperforms our best previous maximum entropy tagger.
In both training and testing the CRF protein-name tagger, the corresponding Medline abstracts were processed as follows: text was tokenized using white space as delimiters and treating all punctuation marks as separate tokens. The text was segmented into sentences, and part-of-speech tags were assigned to each token using Brill's tagger . For each token in each sentence, a vector of binary features was generated using the feature templates employed by the maximum entropy approach described in . Each feature occurring in the training data was associated with a parameter in the CRF model. We used the CRF implementation from McCallum . To train the CRF's parameters, we used 750 Medline abstracts manually annotated for protein names . We then tagged predicted protein names in the entire set of 753,459 Medline abstracts using the version of the CRF algorithm that utilizes the dictionary as part of the learned model (Figure 2), and in this way linked each tagged name to a dictionary entry. The Medline abstracts with marked-up protein names are available on request.
The model assigns each candidate phrase a probability of being a protein name. We selected all names scoring higher than a given threshold (testing thresholds between 40% and 95%), retaining the proteins' LocusLink identifiers along with the PubMed identifiers (PMID) of the associated abstracts. The significance of co-citation of two protein names across a set of Medline abstracts was calculated from the hypergeometric distribution [14, 26] as:
and N equals the total number of abstracts, n of which cite the first protein, m cite the second protein, and l cite both.
The top-scoring 15,000 co-cited protein pairs were then re-ranked according to the tendency of the co-citing abstracts to discuss protein-protein interactions. Specifically, the likelihood of a co-citing abstract to discuss physical protein interactions was evaluated using the naive Bayesian classifier as described in , which scores Medline abstracts according to usage frequencies of discriminating words relating to protein-protein interactions. For each co-cited protein pair, we calculated the average of the scores of the co-citing Medline abstracts, then re-ranked the co-cited protein pairs by these average scores.
Analysis of network properties
We evaluated the clustering of genes in an interaction network  by calculating the average clustering coefficient (<C>) of the N genes as:
where C i is the clustering coefficient of gene i, evaluated over the set of genes with at least two interactions and measured as the number of links, n, among the gene's k neighbors, divided by the number of maximum possible linkages, k(k -1)/2.
Construction of the functional annotation benchmark
The specific GO and KEGG annotations for the functional benchmarks were downloaded from the Gene Ontology database  and the KEGG database . Within the GO process annotation hierarchy (more strictly, a directed acyclic graph (DAG)), the number of distinct annotation terms is maximal at level 8, where the level is defined as the number of nestings from the root node (level 1), as given in the Gene Ontology DAG file . KEGG functional annotations were constructed as the sets of numerical codes for the KEGG pathway diagrams associated with each gene. The functional annotation benchmark is composed of all pairs of human genes sharing annotation from either source (KEGG or GO). For training and testing sets, annotated genes were randomly assigned into two categories and associations were only considered between genes of the same category.
The ID-Serve database
ID-Serve is a relational mySQL database of human proteins created to simplify comparison of datasets with differing protein identifiers. The database maps 42,232 LocusLink (now EntrezGene) identifiers to their corresponding Genecard, Swiss-Prot, Ensembl, OMIM, Unigene, NCBI GI codes and Accession numbers and to the GO and KEGG pathway annotations. Protein interaction data can be retrieved from ID-Serve, with co-citation derived interactions hyperlinked to the supporting Medline abstracts.
Additional data files
The following additional data relevant to the analysis, training and testing carried out in this work are available with the online version of this paper and can also be obtained from the ID-Serve database . Additional data files 1 and 2 contain tables of protein 'tagger' training sets. Additional data file 3 contains a dictionary of human protein names and synonyms indexed to LocusLink identifiers. Additional data file 4 contains the final set of 31,609 protein interactions between 7,748 proteins derived from this analysis. Additional data file 5 contains the final set of co-citation/Bayesian classifier-derived interactions with the PubMed identifiers of co-citing abstracts. Additional data file 6 contains the benchmark training set of functional annotations. Additional data file 7 contains the benchmark test set of functional annotations. Additional data file 8 contains the benchmark set of physical interactions. Additional data file 9 contains the discriminating word list used by the Bayesian classifier to estimate the likelihood of Medline abstracts to discuss protein interactions.
We thank Insuk Lee for critical comments and Zack Simpson for critical comments and help with network visualization. We also thank Ewan Birney's group at the European Bioinformatics Institute for providing us with the interaction data from Reactome. This work was supported by grants from the NSF. (IIS-0325116, EIA-0219061), NIH. (GM06779-01), Welch (F1515), and a Packard Fellowship (E.M.M.).
- Ito T, Chiba T, Ozawa R, Yoshida M, Hattori M, Sakaki Y: A comprehensive two-hybrid analysis to explore the yeast protein interactome. Proc Natl Acad Sci USA. 2001, 98: 4569-4574. 10.1073/pnas.061034498.PubMedPubMed CentralView ArticleGoogle Scholar
- Uetz P, Giot L, Cagney G, Mansfield TA, Judson RS, Knight JR, Lockshon D, Narayan V, Srinivasan M, Pochart P, et al: A comprehensive analysis of protein-protein interactions in Saccharomyces cerevisiae. Nature. 2000, 403: 623-627. 10.1038/35001009.PubMedView ArticleGoogle Scholar
- Gavin AC, Bosche M, Krause R, Grandi P, Marzioch M, Bauer A, Schultz J, Rick JM, Michon AM, Cruciat CM, et al: Functional organization of the yeast proteome by systematic analysis of protein complexes. Nature. 2002, 415: 141-147. 10.1038/415141a.PubMedView ArticleGoogle Scholar
- Ho Y, Gruhler A, Heilbut A, Bader GD, Moore L, Adams SL, Millar A, Taylor P, Bennett K, Boutilier K, et al: Systematic identification of protein complexes in Saccharomyces cerevisiae by mass spectrometry. Nature. 2002, 415: 180-183. 10.1038/415180a.PubMedView ArticleGoogle Scholar
- Tong AH, Evangelista M, Parsons AB, Xu H, Bader GD, Page N, Robinson M, Raghibizadeh S, Hogue CW, Bussey H, et al: Systematic genetic analysis with ordered arrays of yeast deletion mutants. Science. 2001, 294: 2364-2368. 10.1126/science.1065810.PubMedView ArticleGoogle Scholar
- Tong AH, Lesage G, Bader GD, Ding H, Xu H, Xin X, Young J, Berriz GF, Brost RL, Chang M, et al: Global mapping of the yeast genetic interaction network. Science. 2004, 303: 808-813. 10.1126/science.1091317.PubMedView ArticleGoogle Scholar
- Gabaldon T, Huynen MA: Prediction of protein function and pathways in the genome era. Cell Mol Life Sci. 2004, 61: 930-944. 10.1007/s00018-003-3387-y.PubMedView ArticleGoogle Scholar
- Eisenberg D, Marcotte EM, Xenarios I, Yeates TO: Protein function in the post-genomic era. Nature. 2000, 405: 823-826. 10.1038/35015694.PubMedView ArticleGoogle Scholar
- Huynen MA, Snel B, von Mering C, Bork P: Function prediction and protein networks. Curr Opin Cell Biol. 2003, 15: 191-198. 10.1016/S0955-0674(03)00009-7.PubMedView ArticleGoogle Scholar
- Mellor JC, Yanai I, Clodfelter KH, Mintseris J, DeLisi C: Predictome: a database of putative functional links between proteins. Nucleic Acids Res. 2002, 30: 306-309. 10.1093/nar/30.1.306.PubMedPubMed CentralView ArticleGoogle Scholar
- Jansen R, Yu H, Greenbaum D, Kluger Y, Krogan NJ, Chung S, Emili A, Snyder M, Greenblatt JF, Gerstein M: A Bayesian networks approach for predicting protein-protein interactions from genomic data. Science. 2003, 302: 449-453. 10.1126/science.1087361.PubMedView ArticleGoogle Scholar
- Deane CM, Salwinski L, Xenarios I, Eisenberg D: Protein interactions: two methods for assessment of the reliability of high throughput observations. Mol Cell Proteomics. 2002, 1: 349-356. 10.1074/mcp.M100037-MCP200.PubMedView ArticleGoogle Scholar
- von Mering C, Krause R, Snel B, Cornell M, Oliver SG, Fields S, Bork P: Comparative assessment of large-scale datasets of protein-protein interactions. Nature. 2002, 417: 399-403. 10.1038/nature750.PubMedView ArticleGoogle Scholar
- Lee I, Date SV, Adai AT, Marcotte EM: A probabilistic functional network of yeast genes. Science. 2004, 306: 1555-1558. 10.1126/science.1099511.PubMedView ArticleGoogle Scholar
- Mrowka R, Patzak A, Herzel H: Is there a bias in proteome research?. Genome Res. 2001, 11: 1971-1973. 10.1101/gr.206701.PubMedView ArticleGoogle Scholar
- Giot L, Bader JS, Brouwer C, Chaudhuri A, Kuang B, Li Y, Hao YL, Ooi CE, Godwin B, Vitols E, et al: A protein interaction map of Drosophila melanogaster. Science. 2003, 302: 1727-1736. 10.1126/science.1090289.PubMedView ArticleGoogle Scholar
- Li S, Armstrong CM, Bertin N, Ge H, Milstein S, Boxem M, Vidalain PO, Han JD, Chesneau A, Hao T, et al: A map of the interactome network of the metazoan C. elegans. Science. 2004, 303: 540-543. 10.1126/science.1091403.PubMedPubMed CentralView ArticleGoogle Scholar
- Bader GD, Betel D, Hogue CW: BIND: the Biomolecular Interaction Network Database. Nucleic Acids Res. 2003, 31: 248-250. 10.1093/nar/gkg056.PubMedPubMed CentralView ArticleGoogle Scholar
- Xenarios I, Salwinski L, Duan XJ, Higney P, Kim SM, Eisenberg D: DIP, the Database of Interacting Proteins: a research tool for studying cellular networks of protein interactions. Nucleic Acids Res. 2002, 30: 303-305. 10.1093/nar/30.1.303.PubMedPubMed CentralView ArticleGoogle Scholar
- Peri S, Navarro JD, Kristiansen TZ, Amanchy R, Surendranath V, Muthusamy B, Gandhi TK, Chandrika KN, Deshpande N, Suresh S, et al: Human protein reference database as a discovery resource for proteomics. Nucleic Acids Res. 2004, 32 (Database): D497-501. 10.1093/nar/gkh070.PubMedPubMed CentralView ArticleGoogle Scholar
- Lehner B, Fraser AG: A first-draft human protein-interaction map. Genome Biol. 2004, 5: R63-10.1186/gb-2004-5-9-r63.PubMedPubMed CentralView ArticleGoogle Scholar
- Joshi-Tope G, Gillespie M, Vastrik I, D'Eustachio P, Schmidt E, de Bono B, Jassal B, Gopinath GR, Wu GR, Matthews L, et al: Reactome: a knowledgebase of biological pathways. Nucleic Acids Res. 2005, 33 (Database): D428-432. 10.1093/nar/gki072.PubMedPubMed CentralGoogle Scholar
- Reactome database. [http://www.reactome.org/download]
- Bouwmeester T, Bauch A, Ruffner H, Angrand PO, Bergamini G, Croughton K, Cruciat C, Eberhard D, Gagneur J, Ghidelli S, et al: A physical and functional map of the human TNF-alpha/NF-kappa B signal transduction pathway. Nat Cell Biol. 2004, 6: 97-105. 10.1038/ncb1086.PubMedView ArticleGoogle Scholar
- Colland F, Jacq X, Trouplin V, Mougin C, Groizeleau C, Hamburger A, Meil A, Wojcik J, Legrain P, Gauthier JM: Functional proteomics mapping of a human signaling pathway. Genome Res. 2004, 14: 1324-1332. 10.1101/gr.2334104.PubMedPubMed CentralView ArticleGoogle Scholar
- Jenssen TK, Laegreid A, Komorowski J, Hovig E: A literature network of human genes for high-throughput analysis of gene expression. Nat Genet. 2001, 28: 21-28. 10.1038/88213.PubMedGoogle Scholar
- Rzhetsky A, Iossifov I, Koike T, Krauthammer M, Kra P, Morris M, Yu H, Duboue PA, Weng W, Wilbur WJ, et al: GeneWays: a system for extracting, analyzing, visualizing, and integrating molecular pathway data. J Biomed Inform. 2004, 37: 43-53. 10.1016/j.jbi.2003.10.001.PubMedView ArticleGoogle Scholar
- Liu H, Wong L: Data mining tools for biological sequences. J Bioinform Comput Biol. 2003, 1: 139-167. 10.1142/S0219720003000216.PubMedView ArticleGoogle Scholar
- Hirschman L, Park JC, Tsujii J, Wong L, Wu CH: Accomplishments and challenges in literature data mining for biology. Bioinformatics. 2002, 18: 1553-1561. 10.1093/bioinformatics/18.12.1553.PubMedView ArticleGoogle Scholar
- Kanehisa M, Goto S, Kawashima S, Okuno Y, Hattori M: The KEGG resource for deciphering the genome. Nucleic Acids Res. 2004, 32 (Database): D277-280. 10.1093/nar/gkh063.PubMedPubMed CentralView ArticleGoogle Scholar
- Ashburner M, Ball CA, Blake JA, Botstein D, Butler H, Cherry JM, Davis AP, Dolinski K, Dwight SS, Eppig JT, et al: Gene ontology: tool for the unification of biology. The Gene Ontology Consortium. Nat Genet. 2000, 25: 25-29. 10.1038/75556.PubMedPubMed CentralView ArticleGoogle Scholar
- Bunescu R, Ge R, Kate R, Marcotte EM, Mooney RJ, Ramani AK, Wong YW: Comparative experiments on learning information extractors for proteins and their interactions. Artificial Intell Med. 2005, doi:10.1016/j.artmed.2004.07.016Google Scholar
- Franzen K, Eriksson G, Olsson F, Asker L, Liden P, Coster J: Protein names and how to find them. Int J Med Inform. 2002, 67: 49-61. 10.1016/S1386-5056(02)00052-7.PubMedView ArticleGoogle Scholar
- Fukuda K, Tamura A, Tsunoda T, Takagi T: Toward information extraction: identifying protein names from biological papers. Pac Symp Biocomput. 1998, 707-718.Google Scholar
- Tanabe L, Wilbur WJ: Tagging gene and protein names in biomedical text. Bioinformatics. 2002, 18: 1124-1132. 10.1093/bioinformatics/18.8.1124.PubMedView ArticleGoogle Scholar
- Marcotte EM, Xenarios I, Eisenberg D: Mining literature for protein-protein interactions. Bioinformatics. 2001, 17: 359-363. 10.1093/bioinformatics/17.4.359.PubMedView ArticleGoogle Scholar
- ID-Serve. [http://bioinformatics.icmb.utexas.edu/idserve]
- Barabasi AL, Oltvai ZN: Network biology: understanding the cell's functional organization. Nat Rev Genet. 2004, 5: 101-113. 10.1038/nrg1272.PubMedView ArticleGoogle Scholar
- International Human Genome Sequencing Consortium: Finishing the euchromatic sequence of the human genome. Nature. 2004, 431: 931-945. 10.1038/nature03001.View ArticleGoogle Scholar
- Stapley BJ, Benoit G: Biobibliometrics: information retrieval and visualization from co-occurrences of gene names in Medline abstracts. Pac Symp Biocomput. 2000, 529-540.Google Scholar
- Lafferty J, McCallum A, Pereira F: Conditional Random Fields: Probabilistic models for segmenting and labeling sequence data. Proc 18th Int Conf Machine Learning (ICML 2001). Edited by: Danyluk A. 2001, San Francisco: Morgan KaufmanGoogle Scholar
- Brill E.: Transformation-based error driven learning and natural language processing: A case study in parts of speech tagging. Comput Linguistics. 1995, 21: 543-565.Google Scholar
- McCallum AK: MALLET: A Machine Learning for Language Toolkit. 2002, [http://mallet.cs.umass.edu]Google Scholar
- Gene Ontology database. [http://www.geneontology.org]
- KEGG Encyclopedia. [http://www.genome.jp/kegg/kegg2.html]
- Adai AT, Date SV, Wieland S, Marcotte EM: LGL: creating a map of protein function with an algorithm for visualizing very large biological networks. J Mol Biol. 2004, 340: 179-190. 10.1016/j.jmb.2004.04.047.PubMedView ArticleGoogle Scholar
This article is published under license to BioMed Central Ltd. This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.