WEB-based GEne SeT AnaLysis Toolkit (WebGestalt): update 2013

Gene identifier

As shown in Table 1, the old version of WebGestalt only supported seven gene identifiers from a few public databases for human and mouse, including Entrez Gene ID, Gene Symbol, RefSeq for DNA, RefSeq for protein, Unigene, Ensemble ID and Uniprot ID. In contrast, the new version of WebGestalt not only supports identifiers from more public databases but also recognizes identifiers from major high-throughput platforms for human and seven important model organisms including mouse, rat, worm, fly, yeast, dog and zebrafish. Using human as an example, WebGestalt has increased the number of identifiers from public databases to 13. Meanwhile, WebGestalt has added 35, 4, 5 and 1 identifiers from Affymetrix, Agilent, Illumina and Codelink microarrays, respectively. Although identifiers from gene expression arrays are typically supported by enrichment analysis tools, those from SNP arrays are less well supported. It is increasingly recognized that pathway-based enrichment analysis can greatly complement the single-SNP-based analysis in understanding genetic determinants of common diseases and providing insights into the pathway dysregulation of complex diseases (13,14). WebGestalt has added 21 identifiers from commonly used human SNP array platforms, allowing easy translating of SNP data into biological insights. All 201 supported identifiers for the 8 organisms can be found in Supplementary Table S1. These identifiers will be mapped to Entrez Gene IDs for enrichment analysis in WebGestalt (Figure 1).

Functional category

The old version of WebGestalt only included two types of functional categories (Table 1), whereas the new version covers six types (GO, Pathway, Network, Phenotype, Disease and Drug, and Chromosomal location) (Table 2). For the pathway collection, in addition to the KEGG database, WebGestalt has added data from Pathway Commons and WikiPathways. Because Pathway Commons collects and integrates nine centrally curated biological pathway databases and WikiPathways is a primary source for open community-based curation, adding these two resources has significantly increased the coverage of pathways in WebGestalt. Gene–phenotype association is one of the most important aspects biology. Therefore, we have added phenotype-associated gene sets defined by Mammalian Phenotype Ontology (15) and Human Phenotype Ontology (16) to WebGestalt. WebGestalt has also included gene sets defined by cytogenetic bands, which can help identify alterations related to chromosomal deletions or amplifications, dosage compensation and other regional effects.

One important addition to the new version of WebGestalt includes functional categories defined by network modules, comprising protein–protein interaction network modules and regulatory modules. The former is a recent and unique addition to WebGestalt through computational network analysis. Although a few enrichment analysis tools support protein–protein interaction network-based enrichment analysis, they typically rely on gene sets derived from network decomposition at a single level without considering the hierarchical structure of the network. However, it is well-known that hierarchical organization is a critical intrinsic property of complex systems including biological networks (17). Recent works from the Ideker group (10) suggest that gene networks embed hierarchical structure that is consistent with GO and can extend beyond GO. Along the same line, WebGestalt has added functional categories defined by hierarchical protein interaction network modules. Specifically, for the human network, we combined from seven curated databases (Table 2) all protein–protein interaction data with at least one publication support to build an integrated protein–protein interaction network. For the mouse network, because the seven databases only contained a limited number of interactions, we first used an ortholog-based method (18) to infer mouse interactions from curated human interactions and then combined them with curated mouse interactions. We implemented—with some modifications—the method previously published by Sales-Pardo et al. (19) to identify hierarchical modules from the integrated networks. Although a standard hierarchical clustering is able to reveal hierarchical structure of a network, it does not specify relevant hierarchical levels and modules at different scales. Moreover, it does not assess the statistical significance of the modular organization of a network. Our implementation directly addresses these limitations. Here we briefly describe our network clustering method. A detailed description of our implementation will be included in a separate manuscript. First, using the random walk-based walktrap algorithm (20), we identify the best partition of the network by maximizing the modularity score (21); Secondly, we use the edge switching algorithm (22) to generate 1000 random networks with the same attributes as the protein–protein interaction network and then identify the best partition and corresponding modularity score for each random network; Thirdly, if the modularity score for the interaction network is significantly higher than those for the 1000 random networks (P < 0.05), the interaction network is considered to have a modular organization and is divided into sub-networks (modules) according to the identified best partition. To reveal network modules at different hierarchical levels, we repeat the above three steps iteratively for each sub-network until none of them show a modular organization. Using this method, we identified 987 and 1006 hierarchical modules for the human and mouse protein–protein interaction networks, respectively. Eighty percentage of the modules were enriched for at least one GO term (Fisher’s exact test, FDR < 0.05), suggesting high functional relevance of the network-derived modules. More importantly, network modules without GO enrichment may represent functional categories that are not well captured by existing knowledge. Gene sets defined by these protein interaction network modules have been added to WebGestalt. WebGestalt has also added regulatory modules defined as sets of genes sharing common transcription factor or microRNA binding sites, which were inferred from comparative genomic analysis and made available through MSigDB (8).

Another unique addition to the new version of WebGestalt includes functional categories defined by gene sets associated with diseases or drugs, inferred using our recently published software GLAD4U (11). We first downloaded the disease and drug terms from PharmGKB (23) and then used GLAD4U to derive a gene list for each term. Specifically, for a disease or drug term, GLAD4U queries the term in the MEDLINE database using the eSearch application programming interface (API) developed by the National Center for Biotechnology Information (NCBI), retrieves the corresponding publications, identifies genes associated with the publications based on the gene-to-publication link table provided by Entrez Gene, and returns genes that are significantly associated with the query term. Among all the terms downloaded from PharmGKB, GLAD4U returned at least 5 genes for 2286 disease terms and 758 drug terms. These disease or drug-associated gene sets have been added to WebGestalt. To demonstrate the usefulness of these new gene sets, we uploaded a recently published gene expression signature for colorectal cancer prognosis (24), which includes 487 genes inferred by prioritizing candidate genes identified from genomic and transcriptomic studies on a protein–protein interaction network. The list of all 11 521 genes in the network was used a reference set for the enrichment analysis. As expected, the disease association analysis in WebGestalt correctly identified ‘Colorectal Neoplasms’ as the most enriched disease term (FDR = 7.77e-9). Interestingly, the drug association analysis also correctly identified ‘Fluorouracil’ (FDR = 0.043), a primary drug used in the treatment of colorectal cancer.