Finding function: evaluation methods for functional genomic data
© Myers et al; licensee BioMed Central Ltd. 2006
Received: 10 May 2006
Accepted: 25 July 2006
Published: 25 July 2006
Accurate evaluation of the quality of genomic or proteomic data and computational methods is vital to our ability to use them for formulating novel biological hypotheses and directing further experiments. There is currently no standard approach to evaluation in functional genomics. Our analysis of existing approaches shows that they are inconsistent and contain substantial functional biases that render the resulting evaluations misleading both quantitatively and qualitatively. These problems make it essentially impossible to compare computational methods or large-scale experimental datasets and also result in conclusions that generalize poorly in most biological applications.
We reveal issues with current evaluation methods here and suggest new approaches to evaluation that facilitate accurate and representative characterization of genomic methods and data. Specifically, we describe a functional genomics gold standard based on curation by expert biologists and demonstrate its use as an effective means of evaluation of genomic approaches. Our evaluation framework and gold standard are freely available to the community through our website.
Proper methods for evaluating genomic data and computational approaches will determine how much we, as a community, are able to learn from the wealth of available data. We propose one possible solution to this problem here but emphasize that this topic warrants broader community discussion.
Recent advances in experimental methods have enabled the development of functional genomics, a genome-wide approach to understanding the inner workings of a cell. While such large-scale approaches will undoubtedly be instrumental in extending our knowledge of molecular and cellular biology, they produce enormous amounts of heterogeneous data of varying relevance and reliability. A key challenge in interpreting these data is separating accurate, functionally relevant information from noise.
Here we focus on using noisy genomic datasets to associate uncharacterized genes or proteins with biological processes. Recent literature on protein function prediction focuses on integrating multiple sources of evidence (e.g. physical interactions, genetic interaction, gene expression data) to assign proteins to processes [1–4] or to predict functional associations or interactions between related proteins [5–10]. Individual high-throughput datasets are typically noisy, but effective integration can yield precise predictions without sacrificing valuable information in the data. All of these methods require a gold standard, which is a trusted representation of the functional information one might hope to discover. Such a standard, coupled with an effective means of evaluation, can be used to assess the performance of a method and serves as a basis for comparison with existing approaches. Beyond methods for predicting protein function or interactions, evaluation against gold standards can be used to directly measure the quality of a single genomic dataset, a necessary step in developing and validating new experimental technology.
In this paper, we describe these problems with current evaluation standards with the hope of instigating a community dialog on proper approaches to comparing genomic data and methods. As noted above, there are two typical approaches to using genomic data for analyzing protein function: methods that directly associate proteins with particular processes or functional classes, and methods that focus on predicting functional associations or interactions between pairs of proteins. We focus our attention toward standards for the latter, evaluating pairwise associations between genes produced by either experimental or computational techniques. Many of the problems we describe, however, apply to both approaches, and we suggest an alternative standard for evaluation that is appropriate in both settings. We provide both a trusted set of functional associations between proteins as well as a specific set of biological processes that maps proteins to well-defined functional classes. Both standards are based on curation by a panel of biological experts. Furthermore, we propose several guidelines for using these standards to perform accurate evaluation of methods and data. The resulting evaluation framework can be used to directly measure and compare the functionally relevant information present in raw high-throughput datasets as well as to evaluate or train computational genomics methods.
Our gold standard and evaluation methodology have been implemented in a web-based system  to facilitate community use for comparison among published datasets or methods. We demonstrate the use of our approach on genomic data from Saccharomyces cerevisiae. Accurate evaluation methods are particularly critical for this model organism, because yeast is widely used as a platform for the development of both high-throughput experimental techniques and computational methods. However, the weaknesses we identify in existing evaluation methodologies as well as the solution we propose are applicable to data from other model organisms and humans.
Results and discussion
We first discuss commonly used gold standards and several fundamental issues with current approaches to evaluation of functional genomic data and methods. To address these problems, we propose a new gold standard based on expert curation and recommend appropriate uses of the standard that ensure accurate evaluation. Finally, we describe a web-based implementation of our evaluation framework, which is available for public use by computational and experimental biologists.
Challenges to effective functional evaluation
Existing gold standards
A number of different gold standards for evaluating yeast functional genomic data or methods have been proposed in the literature. Each standard generally consists of sets of gene or protein pairs grouped as either "positive" or "negative" examples. This is due in large part to the fact that some high throughput data takes the form of associations between genes or gene products (e.g. physical or genetic interactions). Furthermore, a pairwise approach to analysis is a natural way to view biological systems, which are composed of networks, or groups of interactions between gene products. Although this is a commonly adopted approach, others have trained classifiers for specific functional classes where individual proteins or genes are directly associated with functional classes or processes [1, 4]. While we focus on data and methods for pairwise associations between proteins here, many of the issues described are equally problematic for such non-pairwise approaches, and we propose an alternative gold standard appropriate for both settings (see details in "Defining a new gold standard" in Methods).
Most functional genomics evaluations derive gold standard positives from functional classification schemes that capture associations of genes or proteins with specific biological processes as reported in the literature [7, 10, 12, 14–18]. Such classifications are available from multiple sources including the Gene Ontology (GO) (and associated annotation repositories such as the Saccharomyces Genome Database), KEGG , the Munich Information Center for Protein Sequences (MIPS) , and the Yeast Protein Database (YPD) . A common source of gold standard negatives is cellular localization data [6, 7, 23, 24]. Most of these methods utilize a localization study in which 75% of the yeast proteome was GFP-tagged and classified into 22 different cellular compartments  and they assume that two proteins localizing to distinct compartments do not interact. Random pairs of proteins sampled from the proteome provide another common gold-standard negative, relying on the assumption that the expected number of functionally related or interacting pairs is much less than the total number of possible pairwise protein-protein combinations [5, 26, 27].
Inconsistencies among and within different standards
Example depth five biological process GO terms. GO term depth is a commonly used metric for biological specificity in the Gene Ontology. 5 of the smallest depth 5 GO terms and 4 of the largest depth 5 GO terms are listed above. The processes described range from very specific behaviors (e.g. contractile ring contraction) to less informative groupings (e.g. cellular protein metabolism), suggesting depth is a poor measure of specificity. The size distribution for all depth 5 GO terms is plotted in Fig. 3.
lipoic acid metabolism (GO:0000273)
cytokinesis, contractile ring contraction (GO:0000916)
DNA ligation (GO:0006266)
lysosomal transport (GO:0007041)
regulation of sister chromatid cohesion (GO:0007063)
cytoskeleton organization and biogenesis (GO:0007010)
protein biosynthesis (GO:0006412)
cellular protein metabolism (GO:0044267)
Functional biases in prediction performance
The majority of current evaluation approaches are performed without regard to which biological processes are represented in the set of true positives (correctly predicted examples), and thus they are often unknowingly skewed toward particular processes. We illustrate this bias with an example using the KEGG pathways gold standard to evaluate genomic data (Fig. 1). In this evaluation, the estimated reliability of microarray co-expression drops dramatically when a single pathway ("Ribosome" or sce3010) is excluded from the analysis. The substantial drop in precision suggests that a large fraction of the true positives predicted by co-expression are exclusively ribosome relationships. In fact, of the positive examples in the 1% most co-expressed pairs, 86% (~8500 of 9900) are due to co-annotation to the ribosome pathway. This bias becomes even more pronounced at higher co-expression level cutoffs: of the 0.1% most co-expressed positive pairs, 99% (842 of 848) are from the ribosome pathway. We find a similar bias in evaluations using the GO and MIPS gold standards.
Thus, the traditional approach of using a general ROC curve (or related measure) without regard to which processes are represented can be misleading (see Methods for a discussion of ROC curves). This is particularly true when the data or computational predictions have process-dependent reliability as is often the case with genomic or proteomic data. The problem is magnified when the gold standard examples themselves are heavily skewed towards specific functional categories. While the general precision-recall characteristics such as those portrayed in Figure 1 are technically correct, they generalize poorly to non-ribosomal protein relationships. Thus, such an evaluation would be misleading for a scientist hoping to use these data to generate new hypotheses about proteins unrelated to the ribosome. We address this problem in our process-specific evaluation framework.
Gold standard negatives
Another shortcoming of current standards for gene/protein function prediction is the nature of the gold standard negative examples. In yeast, one proposed source of gold standard negatives is based on protein localization data [23, 25] because pairs of proteins localizing to different cellular compartments are highly enriched for non-interacting proteins. However, localization data is likely not representative of "typical" unrelated protein pairs. For instance, Ben-Hur and Noble found the performance of SVM classifiers trained with localization negatives artificially inflated because this negative set is composed entirely of high-confidence pairs [5, 33]. Using such a non-representative "easy" set of negatives will overestimate prediction accuracy, and the resulting classifier will generalize poorly to real biological problems.
Thus, although protein localization data is a strong negative indicator of functional relationships or interactions, we caution against its use as a general negative gold standard. This is particularly problematic for higher-level questions such as function prediction, because proteins co-involved in some biological processes span cellular compartments. Perhaps a safer role for localization data is as the input to computational approaches. We suggest an alternative negative standard based on the biological process Gene Ontology that can provide representative negative examples (see "Suggestions for representative functional evaluation of data and methods").
Relative size of gold standard positive/negative sets
A final issue common among many evaluation standards in the literature is the relative size of the positive and negative example sets. The expected number of proteins involved in any particular biological process is a small percentage of the proteome, which should be reflected in evaluation standards. This imbalance is particularly problematic in methods based on pairwise associations between proteins, where the expected number of protein pairs sharing functional relationships is an even smaller fraction of all possible protein combinations. For instance, of the 18 million possible protein pairs in yeast, it is expected that less than 1 million are functionally related. This large difference makes the typical reporting of sensitivity and specificity misleading. For instance, a recently published method for predicting protein-protein interactions from several genomic features showed seemingly impressive 90% sensitivity and 63% specificity in evaluations , but would make correct predictions only 1 out of every 9 times when applied on a whole-genome scale, rendering the method impractical in many experimental contexts (details in additional file 3: Supplementary discussion).
Given this imbalance, an appropriate measure of functional relevance of genomic data or predictions is the precision or positive predictive value (PPV). This measure rewards methods that generate firm positive predictions, without regard to the accuracy of negative predictions, which are less helpful in guiding laboratory experiments. Direct application of precision may be misleading, though, because this measure is only correct under the assumption that the ratio of positive to negative examples in the gold standard matches that in the application domain. If the ratio of positive to negatives in the gold standard is much larger than in whole-genome data, as is often the case in published evaluations, then the number of false positive predictions will be small and will artificially inflate the precision statistic. For instance, the 90%-63% sensitivity-specificity example above used an approximately equal number of positive and negative examples (1500 and 2000 respectively), leading to 65% precision. However, application of this method on a whole-genome scale, where the ratio of positive to negative examples is roughly 20 times smaller, would lead to an expected precision of just 11% (details in additional file 3: Supplementary discussion).
To avoid such misleading evaluations, the balance of positives and negatives in the gold standard should match that of the application domain as closely as possible. Precision, or PPV, then becomes a direct, representative measure of how well one could expect a dataset or method to perform on whole-genome tasks. Of course, precision alone does not convey all of the important information, only the quality of the predictions made by a dataset or method. It must be reported in tandem with some measure of the quantity of true predictions made. A standard measure for this is the recall, or sensitivity , which is what is used in our evaluation framework (for more details, see Methods).
Suggestions for representative functional evaluation of data and methods
In light of these problems with current gold standards and approaches to evaluation, we have compiled a new functional genomics gold standard and suggest several strategies for accurate comparative evaluation of genomic datasets and methods.
Defining a new gold standard
As discussed previously, a major issue with the current state of the community is inconsistency among the variety of standards used. Evaluations based on different standards (e.g. derived from KEGG versus GO) are often not comparable, even in a qualitative sense. Deriving a standard from these hierarchies is further complicated due to varying levels of biological specificity of curated biological knowledge. Furthermore, each of the sources of curated information has inherent functional biases that can lead to incorrect estimates of accuracy.
To develop a unified standard for general application in functional genomics, several key criteria must be met. The standard must be cross-organismal to ensure relevance to a broad audience. Secondly, the standard should cover a wide variety of biological functions or processes to facilitate comprehensive evaluations. Finally, the standard should adapt quickly as biological knowledge expands. Although there are several sources of annotation that satisfy these criteria to varying extents (eg. KEGG, MIPS, and GO), GO is arguably the best option to serve as a foundation for the standard, as it is well-curated and was designed for complete coverage.
This curation can also be used to obtain a negative standard which addresses some issues with currently used methods. Specifically, our standard includes a set of negatives more broadly representative than sources such as localization while excluding likely positive examples (a shortcoming of approaches that use random sampling). Further, the standard approximates the correct relative balance of positive and negative sets enabling biologically relevant evaluations (see Methods for details).
Evaluating genomic methods and data
In addition to defining a unifying standard, it is critical to use the standard in a manner that accurately reflects the biological reliability of datasets or methods. To expressly address the process-specific variability in accuracy, we developed an evaluation framework that facilitates identification of functional biases in current general evaluations. To accomplish this, we propose that two complementary modes of analysis accompany any evaluation of functional genomic data: (1) a genome-wide evaluation that estimates general reliability but also reports the functional composition of the results and (2) a process-specific evaluation in which the data or method is independently evaluated against a set of expert-selected processes.
Figure 5 illustrates an example of a genome-wide evaluation of several high-throughput datasets using our framework. At first glance, a general evaluation indicates that the Gasch et al. microarray data is the second most reliable source for functional data (Fig. 5a). However, an analysis of the processes represented in the set of correctly classified pairs reveals that approximately 60% of the correct predictions by the co-expression data are related to the process of ribosome formation (Fig. 5a, bottom chart). This type of analysis is included for any evaluation done with our system and interactive visualization allows for quick and accurate detection of any biases that might be present.
In addition to identifying biases in genome-wide evaluations of datasets or methods, our evaluation framework provides a way to normalize these biases out of the analysis. A user can choose to exclude all positive examples related to one or more biological processes. Figure 5b illustrates an example of this functionality for the evaluation discussed above. Based on the bias we observed, we excluded all proteins involved in ribosome biogenesis and assembly (GO term GO:0042254) and re-evaluated the same set of datasets. While none of the interaction datasets change significantly with this process excluded, both gene expression datasets show substantial decay in their precision-recall characteristics, suggesting they are generally less reliable at predicting functional relationships over a broad range of processes. This result is quite different from what we might have concluded had we not been able to discover and correct this process-specific bias.
Many biological laboratories focus on specific processes or domains of interest, even when using high throughput data/methods. In such situations, a targeted, process-specific evaluation is often more appropriate than a genome-wide evaluation. Our framework facilitates convenient and representative process-specific evaluations by performing independent precision-recall analysis for each process of interest.
We have identified a number of serious issues with current evaluation practices in functional genomics. These problems make it practically impossible to compare computational methods or large-scale datasets and also result in conclusions or methods that generalize poorly in most biological applications. We have developed an expert-curated functional genomics standard and a methodological framework that address the problems we have identified. We hope these can serve as an alternative to current evaluation methods and will facilitate accurate and representative evaluation. Furthermore, we hope our analysis will initiate a broader community discussion about appropriate evaluation techniques and practices.
In recent years, the computational community has played an influential role in the field of genomics by contributing many valuable computational methods that facilitate discovery of biological information from high-throughput data. However, without an accurate understanding of how well the computational methods perform, the role of bioinformatics in directing experimental biology will remain limited. Lack of accurate assessment of the experimental methods themselves hinders both interpretation of the results and further development of genomic techniques. Thus, representative evaluation of computational approaches and high throughput experimental technologies is imperative to our ability as a community to harness the full potential of biological data in the post-genome era.
GO-based functional gold standard
With the Gene Ontology and corresponding annotations in hand, the main issue in generating a standard for evaluation is deciding which terms are specific enough to imply functional associations between gene products. As noted in Results and discussion, the typical approach to this problem has been to select a particular depth in the ontology, below which all co-annotated genes are taken to be positive examples. This has obvious problems in that biological specificity varies dramatically at any given depth in the ontology (see Fig. 3 and Table 1 for details). Another approach reported in the literature is to use term size (i.e. the number of gene product annotations) as a proxy for biological specificity. Using this approach, gene products co-annotated to terms smaller than a certain threshold are considered positive examples. The number of annotation genes, however, is not only a function of how specific a particular term is, but often how well-studied the area is. Thus size is not always an accurate indicator of specificity, and this problem only becomes worse in organisms that are less well-studied.
To address the issue of biological specificity of positive examples, we chose the less automated but more direct and biologically consistent approach of expert curation. For this task, we chose six biological experts with doctorate degrees in yeast genomics. This group contains a cumulative total of more than 40 years of post-doctoral experience working with yeast in a research setting. Instead of using characteristics of the GO term (e.g. depth in the hierarchy, number of annotations) to determine specificity, we instructed our expert panel to formally assess which GO terms are specific enough to imply a meaningful biological relationship between two annotated proteins. More precisely, we instructed the experts to select terms with enough specificity that predictions based on them could be used to formulate detailed biological hypotheses, which could be confirmed or refuted by laboratory experiments. This curation was performed for all GO terms from the biological process branch of the ontology without information of their hierarchical relationships, and each set of resulting responses was corrected for hierarchical inconsistencies. Responses for all experts were then merged by counting the number of votes for each GO term and terms that received more than three votes were selected for the positive evaluation standard. The final counts for all GO terms can be obtained from Biological expert voting results.
Given this set of specific GO terms, we can generate a positive pairwise gold standard by considering all proteins co-annotated to each term as positives. This set of specific functional classes can also be used to directly evaluate or train computational approaches that explicitly associate proteins with particular biological processes as well. For this, we start with the set of specific terms and obtain a non-redundant set by removing any terms whose ancestors are also in the set. This set of terms can be obtained from additional file 2: Non-redundant set of specific GO terms.
We can also use the results of this voting procedure to define a representative set of negative examples. We expect that GO terms receiving 1 or fewer votes are too general to imply meaningful functional relationships between co-annotated proteins. Furthermore, GO terms with a very large number of direct and indirect annotations (i.e. a substantial fraction of the genome) are most certainly too general to imply meaningful functional relationships between co-annotated members. Thus, we obtain a set of gold standard negatives by finding pairs of proteins in which both members have annotations (other than "biological process unknown") but whose most specific co-annotation occurs in terms with more than 1000 total annotations (~25% of the annotated genome) and with one or fewer votes from our panel of six experts. The resulting negative set is more accurate than random pairs of proteins but is still large enough to reflect our understanding of the relative size of functionally related to unrelated pairs in the genome. Furthermore, this set of negative examples is more representative of the presumed distribution of biological negatives than alternate sources of negative evidence such as co-localization. The final gold standard based on this analysis can obtained from http://avis.princeton.edu/GRIFn/data/GO_curated_gold_standard.txt.gz: GO-based yeast functional gold standard. This file contains the final pairwise gold standard set of positive and negatives resulting from our expert curation. Yeast protein pairs classified as positives are labeled with a Â“1Â” and pairs classified as negative in the standard are indicated with a -1.
The resulting set of gold standard positive and negative examples is quite different from previously used GO standards based on size or depth as a measure of biological specificity. Figure 4 illustrates this, plotting a histogram of GO term depth and size for both the excluded and included GO term sets based on the biological expert voting procedure described above. Because our gold standard is based on direct re-evaluation of the gene ontology with respect to functional genomics, there are a number of non-specific GO terms excluded based on the voting results that appear relatively deep in the ontology, and conversely, a number of relevant GO terms included that appear near the root (Fig. 4). A similar trend is true of the GO term sizes of the selected and excluded set: many of the GO terms excluded on the basis of expert voting have relatively few annotations. This confirms our earlier observation that neither size nor depth in the ontology serve as good measures of biological specificity. Basing the criteria for generating a GO-based gold standard instead on expert knowledge ensures that the standard is consistent in terms of the biological specificity of the relationships it is capturing and can therefore provide a meaningful basis for evaluation.
Other efforts have previously aimed to derive summary terms from the GO hierarchy, most notably the Saccharomyces Genome Database's (SGD) GO Slim set . This set, however, is not generally appropriate for the purposes of functional evaluation as it was constructed to be a set of "broad biological categories" meant to span the entire range of processes . The functional relationships captured by such broad terms are often too general to provide a meaningful basis for data evaluation. For example, protein biosynthesis (GO:0006412) is one such term included in the GO Slim set, which has approximately 800 annotated genes. A prediction of an uncharacterized protein's involvement in "protein biosynthesis" would not be specific enough to warrant further experimental investigation in most cases. Furthermore, from the perspective of defining an accurate pairwise evaluation standard, clearly not every pair of genes within this set (over 300,000 possible pairwise combinations) has a specific functional relationship.
Metrics for evaluation: ROC and precision-recall curves
Sensitivity-specificity and precision-recall analysis are two approaches to measuring the predictive accuracy of data from two classes given the class labels (referred to here as positive and negative). Sensitivity and specificity are typically computed over a range of thresholds (for multi-valued data) and plotted with respect to one another. Such an analysis is known as a Receiver Operating Characteristic (ROC) curve and portrays the trade-off between sensitivity and specificity. Each threshold yields one point on the curve by considering protein pairs whose association in the data exceeds the threshold value to be positive predictions and other pairs to be negative. Precision-recall analysis is done in the same way, but with precision (or PPV) replacing specificity. Each of these quantities is calculated as follows:
True positives (TP): protein pairs associated by data and annotated as positives in gold standard
False positives (FP): protein pairs associated by data and annotated as negatives in gold standard
True negatives (TN): protein pairs not associated by data and annotated as negatives in gold standard
False negatives (FN): protein pairs not associated by data and annotated as positives in gold standard
ROC and precision-recall curves can be summarized with a single statistic: the area under the curve. For ROC curves, we refer to this statistic as the AUC, which is equivalent to the Wilcoxon rank-sum (Mann-Whitney) statistic. Precision-recall characteristics can be summarized with a similar measure which we refer to as the AUPRC. For all plots shown here, we have used AUPRC because precision is more informative than specificity for the typical sizes of positive and negative example sets as discussed in the "Relative size of gold standard positive/negative sets" section of Results and discussion.
Implementation of web-based evaluation framework
To facilitate community use of the standard, we have implemented our evaluation framework in a public, web-based system available at . All evaluations are based on the standard described in "Defining a new gold standard", which is also available for download at http://avis.princeton.edu/GRIFn/data/GO_curated_gold_standard.txt.gz: GO-based yeast functional gold standard and additional file 2: Non-redundant set of specific GO terms. The website allows users to upload genomic datasets for evaluation and includes several widely used high throughput datasets (including those described here) for comparative evaluation. The methods for presenting evaluation results, including all graphs and interactive components, were implemented in SVG (Scalable Vector Graphics), which can be viewed on most browsers with freely available plugins (see Help at  for details). The web interface was implemented in PHP, with a back-end MySQL database and C++ evaluation server.
The authors would like to gratefully acknowledge Matt Brauer, Kara Dolinski, Maitreya Dunham, Rose Oughtred, and Charlotte Paquin for their help in constructing the Gene Ontology-based standard. We also thank John Wiggins and Mark Schroeder for excellent technical support. CLM and CH are supported by the Quantitative and Computational Biology Program NIH grant T32 HG003284. MAH is supported by NSF grant DGE-9972930. OGT is an Alfred P. Sloan Research Fellow. This research was partially supported by NIH grant R01 GM071966 and NSF grant IIS-0513552 to OGT and NIGMS Center of Excellence grant P50 GM071508.
- Barutcuoglu Z, Schapire RE, Troyanskaya OG: Hierarchical multi-label prediction of gene function. Bioinformatics. 2006Google Scholar
- Clare A, King RD: Predicting gene function in Saccharomyces cerevisiae. Bioinformatics. 2003, 19 (Suppl 2): II42-II49.PubMedView ArticleGoogle Scholar
- Lanckriet GR, Deng M, Cristianini N, Jordan MI, Noble WS: Kernel-based data fusion and its application to protein function prediction in yeast. Pac Symp Biocomput. 2004, 300-311.Google Scholar
- Pavlidis P, Weston J, Cai J, Noble WS: Learning gene functional classifications from multiple data types. J Comput Biol. 2002, 9 (2): 401-411. 10.1089/10665270252935539.PubMedView ArticleGoogle Scholar
- Ben-Hur A, Noble WS: Kernel methods for predicting protein-protein interactions. Bioinformatics. 2005, 21 (Suppl 1): i38-i46. 10.1093/bioinformatics/bti1016.PubMedView 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 (5644): 449-453. 10.1126/science.1087361.PubMedView ArticleGoogle Scholar
- Lee I, Date SV, Adai AT, Marcotte EM: A probabilistic functional network of yeast genes. Science. 2004, 306 (5701): 1555-1558. 10.1126/science.1099511.PubMedView ArticleGoogle Scholar
- Lin N, Wu B, Jansen R, Gerstein M, Zhao H: Information assessment on predicting protein-protein interactions. BMC Bioinformatics. 2004, 5 (1): 154-10.1186/1471-2105-5-154.PubMedPubMed CentralView ArticleGoogle Scholar
- Troyanskaya OG, Dolinski K, Owen AB, Altman RB, Botstein D: A Bayesian framework for combining heterogeneous data sources for gene function prediction (in Saccharomyces cerevisiae). Proc Natl Acad Sci USA. 2003, 100 (14): 8348-8353. 10.1073/pnas.0832373100.PubMedPubMed CentralView ArticleGoogle Scholar
- Wong SL, Zhang LV, Roth FP: Discovering functional relationships: biochemistry versus genetics. Trends Genet. 2005, 21 (8): 424-427. 10.1016/j.tig.2005.06.006.PubMedView ArticleGoogle Scholar
- Kanehisa M, Goto S: KEGG: kyoto encyclopedia of genes and genomes. Nucleic Acids Res. 2000, 28 (1): 27-30. 10.1093/nar/28.1.27.PubMedPubMed CentralView ArticleGoogle Scholar
- Yamanishi Y, Vert JP, Kanehisa M: Protein network inference from multiple genomic data: a supervised approach. Bioinformatics. 2004, 20 (Suppl 1): I363-I370. 10.1093/bioinformatics/bth910.PubMedView ArticleGoogle Scholar
- GRIFn Home Page. [http://function.princeton.edu/GRIFn]
- Gavin AC, Bosche M, Krause R, Grandi P, Marzioch M, Bauer A, Schultz J, Rick JM, Michon AM, Cruciat CM: Functional organization of the yeast proteome by systematic analysis of protein complexes. Nature. 2002, 415 (6868): 141-147. 10.1038/415141a.PubMedView ArticleGoogle Scholar
- Sprinzak E, Sattath S, Margalit H: How reliable are experimental protein-protein interaction data?. J Mol Biol. 2003, 327 (5): 919-923. 10.1016/S0022-2836(03)00239-0.PubMedView ArticleGoogle Scholar
- Uetz P, Giot L, Cagney G, Mansfield TA, Judson RS, Knight JR, Lockshon D, Narayan V, Srinivasan M, Pochart P: A comprehensive analysis of protein-protein interactions in Saccharomyces cerevisiae. Nature. 2000, 403 (6770): 623-627. 10.1038/35001009.PubMedView ArticleGoogle Scholar
- von Mering C, Krause R, Snel B, Cornell M, Oliver SG, Fields S, Bork P: Comparative assessment of large-scale data sets of protein-protein interactions. Nature. 2002, 417 (6887): 399-403. 10.1038/nature750.PubMedView ArticleGoogle Scholar
- Lee SG, Hur JU, Kim YS: A graph-theoretic modeling on GO space for biological interpretation of gene clusters. Bioinformatics. 2004, 20 (3): 381-388. 10.1093/bioinformatics/btg420.PubMedView ArticleGoogle Scholar
- Ashburner M, Ball CA, Blake JA, Botstein D, Butler H, Cherry JM, Davis AP, Dolinski K, Dwight SS, Eppig JT: Gene ontology: tool for the unification of biology. The Gene Ontology Consortium. Nat Genet. 2000, 25 (1): 25-29. 10.1038/75556.PubMedPubMed CentralView ArticleGoogle Scholar
- Ball CA, Dolinski K, Dwight SS, Harris MA, Issel-Tarver L, Kasarskis A, Scafe CR, Sherlock G, Binkley G, Jin H: Integrating functional genomic information into the Saccharomyces genome database. Nucleic Acids Res. 2000, 28 (1): 77-80. 10.1093/nar/28.1.77.PubMedPubMed CentralView ArticleGoogle Scholar
- Mewes HW, Frishman D, Guldener U, Mannhaupt G, Mayer K, Mokrejs M, Morgenstern B, Munsterkotter M, Rudd S, Weil B: MIPS: a database for genomes and protein sequences. Nucleic Acids Res. 2002, 30 (1): 31-34. 10.1093/nar/30.1.31.PubMedPubMed CentralView ArticleGoogle Scholar
- Costanzo MC, Crawford ME, Hirschman JE, Kranz JE, Olsen P, Robertson LS, Skrzypek MS, Braun BR, Hopkins KL, Kondu P: YPD, PombePD and WormPD: model organism volumes of the BioKnowledge library, an integrated resource for protein information. Nucleic Acids Res. 2001, 29 (1): 75-79. 10.1093/nar/29.1.75.PubMedPubMed CentralView ArticleGoogle Scholar
- Jansen R, Gerstein M: Analyzing protein function on a genomic scale: the importance of gold-standard positives and negatives for network prediction. Curr Opin Microbiol. 2004, 7 (5): 535-545. 10.1016/j.mib.2004.08.012.PubMedView ArticleGoogle Scholar
- Patil A, Nakamura H: Filtering high-throughput protein-protein interaction data using a combination of genomic features. BMC Bioinformatics. 2005, 6 (1): 100-10.1186/1471-2105-6-100.PubMedPubMed CentralView ArticleGoogle Scholar
- Huh WK, Falvo JV, Gerke LC, Carroll AS, Howson RW, Weissman JS, O'Shea EK: Global analysis of protein localization in budding yeast. Nature. 2003, 425 (6959): 686-691. 10.1038/nature02026.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 (5): 349-356. 10.1074/mcp.M100037-MCP200.PubMedView ArticleGoogle Scholar
- Qi Y, Klein-Seetharaman J, Bar-Joseph Z: Random forest similarity for protein-protein interaction prediction from multiple sources. Pac Symp Biocomput. 2005, 531-542.Google Scholar
- Bork P, Jensen LJ, von Mering C, Ramani AK, Lee I, Marcotte EM: Protein interaction networks from yeast to human. Curr Opin Struct Biol. 2004, 14 (3): 292-299. 10.1016/j.sbi.2004.05.003.PubMedView ArticleGoogle Scholar
- Xenarios I, Rice DW, Salwinski L, Baron MK, Marcotte EM, Eisenberg D: DIP: the database of interacting proteins. Nucleic Acids Res. 2000, 28 (1): 289-291. 10.1093/nar/28.1.289.PubMedPubMed CentralView ArticleGoogle Scholar
- Breitkreutz BJ, Stark C, Tyers M: The GRID: the General Repository for Interaction Datasets. Genome Biol. 2003, 4 (3): R23-10.1186/gb-2003-4-3-r23.PubMedPubMed CentralView ArticleGoogle Scholar
- Alfarano C, Andrade CE, Anthony K, Bahroos N, Bajec M, Bantoft K, Betel D, Bobechko B, Boutilier K, Burgess E: The Biomolecular Interaction Network Database and related tools 2005 update. Nucleic Acids Res. 2005, D418-424. 33 Database
- Lord PW, Stevens RD, Brass A, Goble CA: Investigating semantic similarity measures across the Gene Ontology: the relationship between sequence and annotation. Bioinformatics. 2003, 19 (10): 1275-1283. 10.1093/bioinformatics/btg153.PubMedView ArticleGoogle Scholar
- Ben-Hur A, Noble WS: Choosing negative examples for the prediction of protein-protein interactions. BMC Bioinformatics. 2005, 7 (Suppl61): S2-Google Scholar
- 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 U S A. 2001, 98 (8): 4569-4574. 10.1073/pnas.061034498.PubMedPubMed CentralView ArticleGoogle Scholar
- Tong AH, Drees B, Nardelli G, Bader GD, Brannetti B, Castagnoli L, Evangelista M, Ferracuti S, Nelson B, Paoluzi S: A combined experimental and computational strategy to define protein interaction networks for peptide recognition modules. Science. 2002, 295 (5553): 321-324. 10.1126/science.1064987.PubMedView ArticleGoogle Scholar
- Ho Y, Gruhler A, Heilbut A, Bader GD, Moore L, Adams SL, Millar A, Taylor P, Bennett K, Boutilier K: Systematic identification of protein complexes in Saccharomyces cerevisiae by mass spectrometry. Nature. 2002, 415 (6868): 180-183. 10.1038/415180a.PubMedView ArticleGoogle Scholar
- Gasch AP, Huang M, Metzner S, Botstein D, Elledge SJ, Brown PO: Genomic expression responses to DNA-damaging agents and the regulatory role of the yeast ATR homolog Mec1p. Mol Biol Cell. 2001, 12 (10): 2987-3003.PubMedPubMed CentralView ArticleGoogle Scholar
- Spellman PT, Sherlock G, Zhang MQ, Iyer VR, Anders K, Eisen MB, Brown PO, Botstein D, Futcher B: Comprehensive identification of cell cycle-regulated genes of the yeast Saccharomyces cerevisiae by microarray hybridization. Mol Biol Cell. 1998, 9 (12): 3273-3297.PubMedPubMed CentralView ArticleGoogle Scholar