- Open Access
GoMiner: a resource for biological interpretation of genomic and proteomic data
© Zeeberg et al.; licensee BioMed Central Ltd. 2003
- Received: 1 November 2002
- Accepted: 28 February 2003
- Published: 25 March 2003
We have developed GoMiner, a program package that organizes lists of 'interesting' genes (for example, under- and overexpressed genes from a microarray experiment) for biological interpretation in the context of the Gene Ontology. GoMiner provides quantitative and statistical output files and two useful visualizations. The first is a tree-like structure analogous to that in the AmiGO browser and the second is a compact, dynamically interactive 'directed acyclic graph'. Genes displayed in GoMiner are linked to major public bioinformatics resources.
- Gene Ontology
- Directed Acyclic Graph
- Scalable Vector Graphic
- Consortium Database
- Berkeley Drosophila Genome Project
Gene-expression profiling and other forms of high-throughput genomic and proteomic studies are revolutionizing biology. That much is universally agreed. But the new technologies pose new challenges. The first is the experiment itself, the second is statistical analysis of results, the third is biological interpretation. That third challenge is often the most vexing and time-consuming. In gene-expression microarray studies, for example, one generally obtains a list of dozens or hundreds of genes that differ in expression between samples and then asks: 'What does all of this mean biologically?' The work of the Gene Ontology (GO) Consortium  provides a way to address that question. GO organizes genes into hierarchical categories based on biological process, molecular function and subcellular localization. In the past, this GO information was queried one gene at a time. Recently, batch processing has been introduced , but with a flat-format output that does not communicate the richness of GO's hierarchical structure.
We have developed, and present here, the program package GoMiner as a freely available computer resource that fully incorporates the hierarchical structure of the Gene Ontology to automate the functional categorization of gene lists of any length. GoMiner is downloadable free of charge from  or . GoMiner was developed particularly for biological interpretation of microarray data; one can input a list of under- and overexpressed genes and a list of all genes on the array, and then calculate enrichment or depletion of categories with genes that have changed expression. GoMiner thus facilitates analysis and organization of the results for rapid interpretation of 'omic' [5, 6] data. For concreteness, the descriptions in this article will focus on applications to microarray data, but the range of uses is obviously much broader.
The most important parameter for purposes of interpretation is the enrichment (or depletion) of a category with respect to flagged genes (relative to what would have been expected by chance alone). This parameter will be discussed more extensively and more mathematically in the section on 'Statistical considerations'. In Figure 1a, the relative enrichment is indicated by blue numbers for total flagged genes and by red and green numbers for over- and underexpressed genes, respectively. The last number (blue) for each category is a two-sided p-value from Fisher's exact test.
In GoMiner, clicking on a gene of interest in the tree-structure opens a menu that can be used to submit that gene as a query to an external data resource. The number of such links is being expanded rapidly, but currently included are LocusLink , PubMed , MedMiner [9, 10], GeneCards , the NCBI's Structure Database , and BioCarta and KEGG pathway maps as implemented by the NCI Cancer Genome Anatomy Project (CGAP) . These external databases provide GoMiner with a rich set of resources for bioinformatic integration. For example, the links with CGAP and LocusLink provide interaction with pathway maps, chromosome visualizations, a database of single nucleotide polymorphism (SNP), and the Mammalian Gene Collection (MGC).
In GoMiner, clicking on a category instead of a gene brings up a second visualization (Figure 1b), a DAG programmed as a scalable vector graphic (SVG) that can be navigated fluently. Any of its nodes can be moused-over to list the flagged genes or clicked to highlight multiple pathways connecting it to the root. Detailed quantitative and statistical results are downloadable in several tab-delimited formats that can be read directly into a text file or a spreadsheet program for further analysis. For example, the spreadsheet data can be sorted by enrichment factor or p-value to focus attention on potentially interesting categories.
The heart of GoMiner is its processing engine (Figure 2), which parses input gene lists and retrieves database entries for association with GO categories (also called 'terms'). The GO categories and gene associations are stored in a relational database. To enhance the speed of data manipulation, we model the information in memory using a DAG data structure. The root is the topmost node: 'Gene Ontology'. The other nodes represent gene categories, and the connections represent relationships between categories. Each category-node object contains its associated genes, functionality for counting genes, a flag for dereplication during counting, and results of statistical analyses. The gene-category associations are displayed in the form of a tree (Figure 1a) or, alternatively, in the form of a DAG (Figure 1b).
We have developed GoMiner as a client-server application. The client, a Java application, communicates with a server-side database through JDBC. The client can run on platforms with Java run-time environment version 1.3 or higher. The primary client-user GUI, written using the Java Swing API, takes the form of a three-panel window in which the user can inspect GO categories and genes. The left-hand panel lists the genes, the databases from which their identities were derived, and optional up- and down-arrows to indicate under- or over-expression; the middle panel shows a tree visualization of categories in the style of the AmiGO browser  and, in addition, provides a visualization of the flagged genes in the particular microarray experiment. The right-hand panel shows all appearances within the GO hierarchy of any gene selected from the left or middle panel. The gene and category names are implemented as links to facilitate navigation of the data structures and access to public resources.
A second type of visualization, the DAG (programmed as an SVG) shows in compact form the spanning hierarchy for all flagged genes. Optionally, it can include only nodes below a specified level if the entire DAG would be too large for easy visualization. The client application uses several open source components: the Berkeley Drosophila Genome Project (BDGP) Java Toolkit  for utility classes; Browser Launcher  for cross-platform web browser integration; Jakarta-ORO  for text processing; the Jena Semantic Web Toolkit  for manipulating RDF models; MySQL Connector/J  for database connectivity; and Xerces  for parsing XML. The back-end is a relational database server, which stores all gene ontology data. It includes an implementation in MySQL  of the GO Consortium database.
In addition to the deployed components, we have introduced a number of open-source tools to enhance the development environment. In particular, the Concurrent Versions System (CVS) tool  coordinates program development at the Georgia Institute of Technology with that at the NCI, and also coordinates development within each of the groups. jUnit  automates unit- and system-level testing of the application.
Two-by-two contingency table for flagged and unflagged genes in a GO category
n - nf
Not in category
Nf - nf
(N - n) - (Nf - nf)
N - n
N - Nf
The null hypothesis can be formulated as:
H0:p1 - p2 = 0,
where p1 = nf/n and p2 = (Nf - nf)/(N - n). The two-sided p-value for Fisher's exact test is the sum of probabilities of observing tables that give at least as many extreme values as the one actually observed, given that the null hypothesis is true [23–25]. The use of Fisher's exact test implies that we are conditioning on fixed marginal totals (n, N - n, Nf, N - Nf) under the null hypothesis. For a discussion of the implications of fixed marginal values, see for example [23–25].
The following limitations of this statistical formulation should be borne in mind, and the p-values should be interpreted judiciously.
Random experimental and categorization error
Experimental error and any uncertainties in the classification of genes in GO are not included in the statistical model. Perhaps, given enough information (which we essentially never have) about those sources of error, they could be included in the statistical model, for example through a resampling technique.
Gene representation bias
The microarray gene set (or set from some other type of genomic or proteomic experiment) will generally be a biased representation of all genes. Therefore, enrichments and depletions, of necessity defined in terms of the genes studied, may be biased with respect to biological significance as well. An alternative is to replace the list of the total set of genes on the microarray with a list of the total set of genes in the genome (or a representative sample), but that approach introduces another source of bias: genes not on the microarray are counted in determining N and n but have no chance to be flagged.
GO consortium database bias for human gene associations
The GO Consortium  provides a set of flat files that indicate the association between gene names and GO categories for several species . Although the flat files for human are quite comprehensive, we found a low hit rate for GO annotation of human genes using the database created by the GO Consortium's downloaded MySQL script files . The hit rates were low both when the gene names were used in the format of HUGO names and when the gene names were used in the format of 'HUGO_HUMAN.' We tried the latter format because the flat files often contained '_HUMAN' appended to the human gene names. In contrast, when we used a combination of mouse (MGI) and rat (RGD) association files, there were reasonable numbers of hits. Therefore, we now routinely use mouse and rat annotations for human data. We are currently augmenting the human associations in the GO Consortium database to provide a richer annotation of human gene names. This goal will be achieved by using the MatchMiner database to integrate the information in the GO Consortium database  and the Swiss-Prot, TrEMBL and TrEMBLnew databases , and GoMiner will implement this database for human data in the near term. The MySQL script files will be freely available and should represent an improvement over what is currently available to program developers and end-users.
Non-independence of gene data
Gene-expression values within a category may be correlated for any of several reasons. They may represent the same gene, close family members with similar functions, genes in the same pathway or genes in alternative pathways for performing a biological function. Gene classifications in GO may be correlated for analogous reasons. How do such relationships affect the statistics? The answer is most easily seen by imagining a category containing nothing but five instances of the same gene (perhaps because five different identifiers were used and not recognized as representing the same gene). That category might appear either to be strikingly enriched (with five out of five genes flagged) or strikingly depleted (with none out of five genes flagged). But the appropriate value of n for determining statistical significance in those cases would be 1, not 5. GoMiner's companion program MatchMiner [30, 31] handles this problem by identifying replicates of the same gene, even if they are represented by different identifiers.
What about possible sources of correlation other than 'same-gene'? Do we want to dereplicate them as well? Generally, the answer is 'no'. Correlation of genes in the same pathway is precisely the phenomenon we are often trying to identify. We would not want a statistical test to adjust for (and, in effect, null out) the effect of such relationships. Close family members might be considered an intermediate case. The statistical model implemented in GoMiner assumes, as our state of prior knowledge, that we know when two 'genes' are identical but nothing about their relationship if they are not identical. That seems the only available course. However, for each category, GoMiner provides the gene identities and the numbers given in Table 1 – sufficient information for the knowledgeable user to decide to eliminate close family members or pathway partners if desired.
The multiple comparisons problem
If one has not decided before analysis which particular gene category is to be examined, a correction should be made for the multiple opportunities to obtain a p-value indicating statistically significant enrichment or depletion. For example, with 1,000 categories, we would expect approximately 1,000 × 0.05 = 50 false positives simply by chance if we set the critical value at p = 0.05. The most common way to correct for this problem is that of Bonferroni (see, for example ), in which the critical value is divided by the number of trials (in this case, 1,000). However, that approach assumes independence of categories and is so conservative that it becomes extremely hard to detect true positives. A number of less conservative statistical methods have also been developed, but it is beyond the scope of this paper to review them here. An approach based on resampling will be incorporated into GoMiner in the coming months.
Overall, the p-values quoted should be considered as heuristic measures, useful as indicators of possible statistical significance, rather than as the results of formal inference. The p-values can be used, for example, to sort categories to identify those of the most potential interest.
As another useful measure, we have calculated the relative enrichment factor, Re, defined as
Re = (nf/n)/(Nf/N)
and shown as blue numbers in Figure 1a. The analogous quantities for overexpressed (red numbers) and underexpressed (green numbers) are also shown. Depletion is, of course, represented by an enrichment factor less than unity.
As a test, GoMiner was applied to the results of our cDNA microarray study of the molecular mechanisms by which drug resistance develops . The DAG shown in Figure 1a was generated from that study, which used quadruplicate 'Oncochip' microarrays (Microarray Facility, Advanced Technology Center, NCI ) to compare gene expression profiles in a prostate cancer cell line (DU145) and a subline (RC0.1) selected from it for resistance to the topoisomerase 1-inhibitor 9-nitro-camptothecin. The microarray included 1,399 cancer-interesting genes. 181 of those genes differed in expression according to a threshold criterion (>1.5-fold difference). MatchMiner was used to translate IMAGE clone Ids for the 1,399 genes into HUGO names for input to GoMiner. Figure 1a shows that the category 'apoptosis regulator' was enriched 2.4-fold in genes with altered expression levels. More specifically, it was enriched 3.2-fold with underexpressed genes and 2.0-fold with overexpressed genes. Flow cytometric annexin V and TUNEL assays verified important differences in apoptotic potential between the cell lines, and analysis generated a novel hypothesis (the 'permissive apoptosis-resistance' hypothesis) for the relationship between apoptotic and cell-proliferation pathways in the development of drug resistance. Figure 1a provides more detailed information, indicating that these differences were focused in particular subcategories of apoptosis. Thus, GoMiner can help the user in at least two ways: it identifies categories enriched in, or depleted of, genes of interest; and it generates hypotheses to guide further research.
Unfortunately for us, interpretive analysis of the DU145/RC0.1 study was initially done one gene at a time before development of GoMiner (and, in fact, motivated that development). Performing the GO analysis one gene at a time would have taken more than two solid hours at the computer for the 181 genes before getting to the much harder parts of the task: doing the same for the entire array (nominally > 15 hours), then collating and organizing the information for each GO category. In contrast, operating on a 266 MHz PC with 250 MB RAM, it took 90 seconds to browse for and load the files, then 30 seconds for GoMiner to process the entire array of 1,399 genes and display the flagged and unflagged genes in their hierarchical context. In another test, running 900 flagged genes and all of HUGO (15,000 genes) took 4 minutes and 40 seconds on the same computer. Overall, the processing time was essentially linear with respect to the total number of genes (time in minutes = 0.0003 × genes + 0.0656; R2 = 0.998).
GoMiner is being developed jointly by groups from the National Cancer Institute (NCI), the Georgia Institute of Technology, and Emory University. This project has been supported by a contract funded by the NCI's Center for Cancer Research and by The Wallace H. Coulter Biomedical Engineering Department of Georgia Tech and Emory University academic funds for Professor May D. Wang. Its user features, statistical repertoire, and links to external resources will continue to be expanded through the contract funded by the NCI's Center for Cancer Research and through Professor Wang's academic funds.
- Ashburner M, Ball C, Blake J, Botstein D, Butler H, Cherry J, Davis A, Dolinski K, Dwight S, Eppig J, et al: Gene Ontology: tool for the unification of biology. Nat Genet. 2000, 25: 25-29. 10.1038/75556.PubMedPubMed CentralView ArticleGoogle Scholar
- Khatri P, Draghici S, Ostermeier G, Krawetz S: Profiling gene expression using Onto-Express. Genomics. 2002, 79: 266-270. 10.1006/geno.2002.6698.PubMedView ArticleGoogle Scholar
- GoMiner. [http://discover.nci.nih.gov/gominer]
- GoMiner. [http://www.miblab.gatech.edu/gominer]
- Weinstein JN: Fishing expeditions. Science. 1998, 282: 628-629. 10.1126/science.282.5389.627g.PubMedView ArticleGoogle Scholar
- Weinstein JN: 'Omic' and hypothesis-driven research in the molecular pharmacology of cancer. Curr Opin Pharmacol. 2002, 2: 361-365. 10.1016/S1471-4892(02)00185-6.PubMedView ArticleGoogle Scholar
- LocusLink. [http://0-www.ncbi.nlm.nih.gov.brum.beds.ac.uk/LocusLink]
- PubMed. [http://0-www.ncbi.nlm.nih.gov.brum.beds.ac.uk/entrez/query.fcgi?db=PubMed]
- MedMiner. [http://discover.nci.nih.gov]
- Tanabe L, Scherf U, Smith LH, Lee JK, Hunter L, Weinstein JN: MedMiner: an internet text-mining tool for biomedical information, with application to gene expression profiling. BioTechniques. 1999, 27: 1210-1217.PubMedGoogle Scholar
- GeneCards. [http://thr.cit.nih.gov:8081/cards/index.html]
- NCBI Entrez Structure. [http://0-www.ncbi.nlm.nih.gov.brum.beds.ac.uk/entrez/query.fcgi?db=Structure]
- The Cancer Genome Anatomy Project. [http://cgap.nci.nih.gov/Pathways]
- Berkeley Drosophila Genome Project: developers' resources. [http://www.fruitfly.org/developers]
- BrowserLauncher. [http://browserlauncher.sourceforge.net]
- The Apache Jakarta Project. [http://jakarta.apache.org/oro/index.html]
- HP Labs Semantic Web Research. [http://www.hpl.hp.com/semweb/index.html]
- MySQL Connector/J downloads. [http://www.mysql.com/downloads/api-jdbc.html]
- Xerces2 Java Parser Readme. [http://xml.apache.org/xerces2-j/index.html]
- MySQL. [http://www.mysql.com]
- Concurrent Versions System. [http://www.cvshome.org]
- jUnit. [http://www.junit.org]
- Agresti A: Categorical Data Analysis. 1990, New York: John WileyGoogle Scholar
- Agresti A: A survey of Exact inference for contingency tables. Stat Sci. 1992, 7: 131-177.View ArticleGoogle Scholar
- StatXact 5 for Windows. User Manual. 2002, Cambridge, MA: Cytel Software CorporationGoogle Scholar
- Fisher's Exact Test. [http://www.matforsk.no/ola/fisher.htm]
- GO Database. [http://www.geneontology.org/#godatabase]
- GO downloads. [http://www.godatabase.org/dev/database/archive/]
- Swiss-Prot, TrEMBL and TrEMBLnew database. [ftp://ftp.ebi.ac.uk/pub/databases/sp_tr_nrdb/]
- Bussey JK, Kane D, Sunshine M, Narasimhan S, Nishizuka S, Reinhold WC, Zeeberg B, Ajay , Weinstein JN: MatchMiner: a tool for batch navigation among gene and gene product identifiers. Genome Biol. 2003, 4: R27-10.1186/gb-2003-4-4-r27.PubMedPubMed CentralView ArticleGoogle Scholar
- MatchMiner. [http://discover.nci.nih.gov/matchminer/]
- Bonferroni. [http://home.clara.net/sisa/bonhlp.htm]
- Reinhold WC, Kouros-Mehr H, Kohn KW, Maunakea AK, Lababidi S, Roschke A, Stover K, Alexander J, Pantazis P, Miller L, et al: Apoptotic susceptibility of cancer cells selected for camptothecin resistance: gene expression profiling, functional analysis, and molecular interaction mapping. Cancer Res. 2003, 63: 1000-1011.PubMedGoogle Scholar
- NCI human Oncochip genes. [http://nciarray.nci.nih.gov/gi_acc_ug_title.shtml]
- Doniger SW, Salomonis N, Dahlquist KD, Vranizan K, Lawlor SC, Conklin BR: MAPPFinder: using Gene Ontology and GenMAPP to create a global gene-expression profile from microarray data. Genome Biol. 2002, 4: R7-10.1186/gb-2003-4-1-r7.View ArticleGoogle Scholar
- MAPPFinder. [http://www.genmapp.org/MAPPFinder.html]
- FatiGO. [http://fatigo.bioinfo.cnio.es/]
- Onto-Express. [http://vortex.cs.wayne.edu/Projects.html]
- GoSurfer. [http://biosun1.harvard.edu/complab/gosurfer/]
- BDGP Gene Ontology Editor Tool. [http://www.godatabase.org/dev/editor.html]
This article is published under license to BioMed Central Ltd. This is an Open Access article: verbatim copying and redistribution of this article are permitted in all media for any purpose, provided this notice is preserved along with the article's original URL.