Gemma: a resource for the reuse, sharing and meta-analysis of expression profiling data

Abstract

Summary: Gemma is a database, analysis software system and web site for genomics data re-use and meta-analysis. Currently, Gemma contains analyzed data from over 3300 expression profiling studies, yielding hundreds of millions of differential expression results and coexpression patterns (correlated expression) for retrieval and visualization. With optional registration users can save their own data and securely share it with other users. Web services and integration with third-party resources further increase the scope of the tools, which include a Cytoscape plugin.

Availability: http://chibi.ubc.ca/Gemma, Apache 2.0 license.

Contact: paul@chibi.ubc.ca

1 SCOPE AND DATA SOURCES

The goal of Gemma is to enable the rapid exploration and analysis of large quantities of genomics data, leveraging the extensive data available from other public bioinformatics resources such as the Gene Expression Omnibus (Barrett et al., 2007).

Currently, Gemma contains nearly 4000 expression profiling studies (‘datasets’; in total over 170 000 assays, from eight taxa). Multiple technology types are supported, such as array-based platforms and RNA sequencing. To enable comparisons across platforms, we perform sequence analysis and gene assignment based on the current genome annotations (Barnes et al., 2005). Each public dataset undergoes automated (French et al., 2009) and manual annotation using controlled vocabularies such as the Disease Ontology (Schriml et al., 2012), adding information about the experimental design to allow group comparisons. Additional quality control steps to detect outlier samples or datasets with large batch effects are also performed.

Each dataset is then analyzed for differential expression (e.g. between conditions or tissues) and coexpression (correlation of expression levels across samples). Differential expression is computed using a standard multivariate linear modeling approach (Pavlidis and Noble, 2001) comparing each condition in a dataset with baseline, accommodating complex factorial designs and continuous covariates. Coexpression is computed for each dataset and stored as a set of ‘coexpression links’ that meet stringent statistical criteria (Lee et al., 2004). The results of these analysis are stored in the system for user search and retrieval.

2 FUNCTIONALITY

A main entry point for Gemma is a form that allows users to search for differential expression or coexpression results. The search facilities enable analysis of selected genes [by symbols, key words or Gene Ontology terms (Ashburner et al., 2000)] and experiments (based on free text or our annotations of disease, treatment, tissue, etc.). Users can flexibly organize genes or datasets into groups. With optional registration, these groups persist across sessions and can be securely shared with other users.

Differential expression results are presented in a matrix visualization displaying the genes in rows and individual conditions across studies in columns (Fig. 1). For each gene, a ‘meta-P-value’ is provided. For each condition's differentially expressed genes, Gemma provides information on the enrichment of the user's selected genes in that pattern. The data view can be filtered, sorted and exported as an image. Visualizations of the underlying expression patterns are also readily obtained (Fig. 1).

Fig. 1.

Screen shots from Gemma illustrating (clockwise from top left) the main search form, the output of query for differential expression, a heatmap view of expression profiles and coexpression query results.

For coexpression searches, Gemma applies a user-settable threshold for how many datasets a link must be observed in before it is displayed. Coexpression results are shown in a tabular format and as an interactive network view (Fig. 1) (Lopes et al., 2010). Gemma uses the concept of node degree (how many links a gene has; i.e. ‘hubbiness’) to assist the user in gauging the importance of an observation (Gillis and Pavlidis, 2011). This is important because in a query-driven network view, only a tiny subset of the network is displayed. For each gene, Gemma estimates the overall node degree and indicates low-node degree genes in darker shades (Fig. 1).

Gemma offers many other features for exploration and analysis. For each gene, an overview page shows the datasets in which the gene is differentially expressed, genes with which the gene is reproducibly coexpressed and expression platforms on which the gene is represented (e.g. http://www.chibi.ubc.ca/Gemma/g/?id=14676). Similarly, for each dataset, Gemma provides annotations, summaries of the analyses and visualizations (e.g. http://www.chibi.ubc.ca/Gemma/ee/?id=1570). Registered users can also upload their own expression datasets to be included for meta-analysis and are provided with an extensive suite of administration tools for data management.

Gemma was designed with data reuse and extensibility in mind. Results can be downloaded in tab-delimited formats for external analysis, and web services are available to access Gemma programmatically. As an example of such integration and to allow more advanced visualization and analysis of coexpression data from Gemma, we have developed ‘GemScape,’ a plugin for the popular network analysis tool Cytoscape (Kohl et al., 2011). Data from Gemma are also currently available through the Neuroscience Information Framework (Gupta et al., 2008) (differential expression results) and inSilicoDb (Taminau et al., 2011) (experimental design annotations).

Tutorials, in-line help and a wiki with additional user manuals and system information are available through the Gemma web site.