Gene Expression Atlas at the European Bioinformatics Institute

Atlas web interface

The Gene Expression Atlas interface (Figure 1) allows the user to query for condition-specific gene expression across multiple datasets. There are three basic types of queries: (i) for a gene, or a set of genes, by name or various gene attributes, including synonyms, Ensembl identifiers and Gene Ontology terms; (ii) for a ‘biological condition’, such as, disease name, developmental stage, as well as tissue or cell type; and (iii) for a combination of genes and biological conditions. Biological conditions, also referred to as EFs, are organized using an application ontology called EFO, which is described in more detail in the next section.

If a query matches one gene uniquely, the ‘Gene page’ for that gene is displayed (Figure 2). This page summarizes the behaviour of the selected gene across all Atlas datasets, providing easy access to both statistical analysis results and expression data. All gene pages can be linked directly, using links of the form http://www.ebi.ac.uk/gxa/gene/IDENTIFIER, where IDENTIFIER is any one of annotated gene attributes (e.g. Ensembl, UniProt and other accessions). Direct links can also be made to Atlas experiments, e.g. http://www.ebi.ac.uk/gxa/experiment/E-AFMX-5. Full details for linking and other Atlas use are available in online Atlas documentation.

The thumbnail plots provide a direct link to individual experiment pages where the gene expression profile for the selected gene can be viewed in detail (Figure 3). In the experimental page, multiple ‘search’ options allow the user to retrieve genes of interest and add their expression profiles to the main plot (Figure 3, right). The search options available are: (i) search for any gene by name or attribute; (ii) search up to 10 most similar genes, based on Pearson correlation, to any of the genes currently plotted; and (iii) choose any gene from a list of top 10 differentially expressed genes for the selected study. For each gene, a P-value of significance of differential expression is provided.

It is also possible to query for such a condition as a particular disease, either over all genes or for those matching specified attributes, such as belonging to a pathway. Figure 4 is an example of a summary view of transcriptional activity among members of the ‘Wnt signalling pathway’ in ‘carcinoma’. Both ‘Genes’ and ‘Conditions’ boxes provide auto complete functionality to help the user formulate queries. Condition queries (Figure 4, top) are expanded using our application ontology (EFO, see next section) to include all child terms available for the original query, so that, for instance, brain queries are expanded to all brain parts. The query results are displayed as a heat map (Figure 4, middle); genes matching the query are listed in the first column and the conditions go across the remaining columns grouped under ‘Ontology’ (when the query matches an ontology term) and ‘Keywords’ (when the query matches a keyword). All conditions mapping to Ontology terms are shown as nodes in the EFO tree (Figure 4, bottom; see next section). As in the ‘gene page’, the red colour corresponds to up-regulation of the selected gene in the selected condition, and the blue corresponds to down-regulation. The numbers in the cell correspond to the number of independent studies where differential expression has been observed, and the colour intensity represents the best P-value of this observation (i.e. the brighter the colour, the more significant the P-value). Results can be downloaded in tab-delimited format. Clicking on a cell in the heat map opens a window, showing information about the expression of the selected gene in the associated experiments including thumbnail plots of the gene expression profile(s) and links to the experimental details in the ArrayExpress Archive of Functional Genomics Data, as shown in Figure 2.

A query can be refined by using terms enriched in the results (‘refine your query’ link; Figure 4, middle), by adding conditions to the original query (using the ‘advanced search’ link), as well as by adding various filters through the advanced interface, which allows user to formulate complex queries by combining several conditions, gene property and organism filters in one query.

Biological conditions—the concepts of EFs and their values

High-throughput gene expression experiments are typically used to compare gene expression in different biological conditions. Our approach to describe biological conditions is based on the concept of EF and EF value (EFV). An EF is defined as the experimental variable that is tested for gene expression variation, and EFVs are the values of this variable. For instance, in an experiment that compares gene expression in leukemic to normal blood, EF is a disease state, which has two values, leukemia and normal.

For a study to be included in the Atlas, it must have at least one defined EF with at least two different EFVs, and each EFV should be tested in at least two replicates. Experiments can have multiple EFs, for instance normal and leukemic gene expression can be profiled in peripheral blood or bone marrow. Such an experiment has two EFs, disease state and organism part.

Using free text keywords to describe EFs and their values limits their utility. For instance, to identify genes differentially expressed in ‘cancer’, we would also like to find genes that are studied in an experiment profiling ‘leukemia’. One way to achieve this is to use a disease ontology linking ‘leukemia’ to ‘cancer’ as a type of cancer, and then expand the query to all cancer types. Since in Gene Expression Atlas we are dealing with EFs mapping to a wide range of biological concepts such as organism parts, diseases or treatments, we require multiple source ontologies to describe these conditions. Initial mapping to existing ontologies identified the NCI Thesaurus (9) as providing best coverage due to the large amount of gene expression data performed on cancer samples. However, as no external ontology covered all our EFs and EFVs, we developed our own application ontology called EFO (7).

Mappings from EFO to external ontologies are maintained as identifiers from the external resource into a ‘denition citation’ annotation property in EFO. Equivalent classes are thus mapped from EFO into multiple other ontologies, for example a ‘neoplasia’ in EFO maps to ‘neoplasia’ in the NCI Thesaurus (9). This appears in the EFO ‘denition citation’ property, which has the value ‘NCI Thesaurus:C3262’. Typically, there are multiple maps to external ontologies as many biomedical ontologies and controlled vocabularies are non-orthogonal. EFO is released synchronously with the Atlas, new terms being added to each release where needed to describe Atlas data. EFO is mapped to 14 external ontologies including the NCI Thesaurus, the Foundational Model of Anatomy (10) and Chemical Entities of Biological Interest (ChEBI) (11) (for full list see http://www.ebi.ac.uk/efo/overview). As illustrated in the previous section, EFO is used in the Atlas to expand queries as outlined above, as well as a means to browse the Atlas content.

Where datasets are submitted directly to ArrayExpress, the submission tools guide the users through annotation of their EF and EFV, which later are checked and curated by the ArrayExpress staff. For data imported into ArrayExpress from GEO, we use text mining tools using EFO as the dictionary to identify the potential EFs and EFVs, and then introduce them via curation.

Statistical computations to rank query results

The meta-analysis approach taken in the construction of the Atlas can be outlined as follows. For each experiment:

1. identify differentially expressed genes for each EF;

2. for each gene found, identify EFVs where the gene's; mean expression level is significantly different from its overall mean across all factor values; and

3. perform multiple comparisons correction and aggregate the identified Gene-EFV associations for storage and retrieval.

If an experiment has several factors, i.e. a multi-factorial experiment design, we treat each factor independently from all others, and identify Gene–EFV associations separately for each EF in an experiment. This represents a simplification, and assumes that there are no interactions between EFs (see ‘Discussion’ section). We observed that in most cases (data not shown) of multi-factorial experiment designs, even if the factors are treated independently, once one statistically significant factor is identified for a gene, secondary factors’ effects become readily visible to the investigator when visualizing the data. For this reason, to construct a basic platform for further data analysis, we initially limited statistical analysis to the factor independence scenario.

The detailed description of the data analysis procedure outline is below:

1. If y_gj are expression values for genes g = 1,…,G and arrays j = 1,…,J, pre-processed, background-corrected and normalized, we describe systematic expression effects for each gene by a linear model E(y_g) = Xβ_g, where y_g = (y_g1}, …,y_gJ)^T is the vector of expression values for gene g, X is a known design matrix with full column rank K and β_g = (β_g1,…,β_gK)^T is a gene-specific vector of regression coefficients. The design matrix depends on the experimental design and choice of parameterization, and the regression coefficients represent comparisons of interest between RNA sources in the experiment. These coefficients are estimated with a least squares linear model fitting procedure in the Bioconductor package limma (12) and tested for differential expression (i.e. testing any particular β_gk equal to 0) with moderated Student's; t-statistic via the empirical Bayesian statistics developed by Smyth (12). For each EF in the experiment, we set up a complete pairwise-comparisons contrast matrix (equivalent to one-way Analysis of Variance) as an omnibus test of differential expression across all factor values in the selected factor. We then can accept or reject the ‘equal means for all groups’ null hypothesis on the basis of P-values computed for the omnibus F-statistic via limma as described above, at a specified significance level. These P-values, with appropriate multiple testing adjustment to control the False Discovery Rate (FDR) at 5% (13), allow us to identify differentially expressed genes.

2. Given a set of expression values for a gene under k different conditions, hence k group means μ_g1,…,μ_gk and a significant F-statistic, we look at k differences μ − μ_i, of each group mean to the global mean, and seek to identify which ones are significant, and in what direction. The problem of multiple comparisons with the overall mean is known in statistical literature as an Multiple Comparisons with the Mean procedure. We follow Hsu (14) to make direct inference on multiple comparisons of the means. For each gene we compute k simultaneous confidence intervals governed by a multivariate t-distribution and computed from quantiles of the Studentized maximum modulus statistic and look at their directionality.

3. Steps (i) and (ii) for each factor in a given experiment produce a matrix of up/down calls (−1, 0 or 1) according to the directionality of the confidence interval and respective P-values: one call/P-value for each Gene–EFV combination. Multiple testing adjustment is performed on these P-values to control the global (across genes and contrasts, i.e. EFVs) FDR at 5%, following recommendations by Smyth in limma (12). Taking advantage of the robustness of the tests performed, we use the P-values significance-based calls to aggregate differential expression results into ‘votes’: each time a gene has been observed as differentially expressed in a particular EFV, we use that as a vote for up/down activity for that Gene–EFV combination. These are the numbers displayed in the heatmap view of the Gene Expression Atlas. All the P-values and computed statistics as well as the aggregated votes are stored in a database and indexed for fast retrieval in the interface.

Currently, all data included in the Atlas are based on microarray assays. The statistical method and the data integration framework developed are generic and can be extended to many other technologies, such as RNA-seq and in situ hybridizations. The R-code for this procedure is available as Supplementary Data and will be released as a separate Bioconductor package (in preparation).

Gene expression data

Data for the Atlas are selected from ArrayExpress Archive and selection is based on various criteria outlined earlier. As currently we are using only microarray data, our first consideration is whether sufficient array annotation is given to enable us to map the array design elements to existing gene identifiers. We use two routes for this mapping: we preferentially map array probe sequences to Ensembl genomes (15) or we attempt to map the design element annotation identifiers to gene annotation in UniProt database (16). Where re-annotation fails, experiments that are performed on such arrays cannot be included in the Atlas. The array re-annotation pipeline will be released as a software package, described and published separately (Sarkans et al., in preparation).

Experiments in ArrayExpress Archive that are performed on well-annotated arrays, which have high MIAME scores (2,17), where the EF/EFV annotation and sufficient replication criteria (as well as some other technical criteria not described here), and where normalized data are present, are annotated as ‘suitable for Atlas’. When all basic criteria are satisfied, experiment selection for the Atlas is motivated by the quality of annotation, use of standard platforms and large sample sizes, without any preference for any biological conditions. Recently, we started to produce themed Atlas data releases, e.g. species oriented or addressing a specific research domain, or by curating user-requested studies. Experiments selected for Atlas are then exported from the Archive. The submitter's; normalized data are used, hence we do not perform any renormalization. Prior to loading into the Atlas, annotations are harmonized, experimental descriptions checked for consistency and non-standard terms are standardized. Maps to EFO are added where the term required is present in the ontology. If terms are not in EFO, we examine source ontologies and provide a term name, definition and maps to external ontologies. The term is then placed in the EFO hierarchy that is optimized for the Atlas visualization.

Once data are loaded, statistical computations, as described in the previous section, are performed and for each new experiment, for each EF and EFV, for each gene the P-value is computed.

Currently, the Atlas contains data from nine species. Table 1 shows the number of assays and the number of studies (experiments) included from each. The experiments included in the Atlas together have more than 40 different EFs, covering over 4500 different EFVs. The distribution of the number of assays for the most frequently studied (at least 50 experiments for each factor) EFs and EFVs are given in Table 2.

The method used in Gene Expression Atlas analytics allows us to examine trends in differential gene expression across all Atlas data. Figure 5A shows the distribution of proportions of differentially expressed genes across all experiments. There are approximately 400 experiments (from over 1000) with fewer than 10% of all genes showing differential expression; the mean proportion of genes differentially expressed in an experiment, according to our FDR criteria, is 25%. Further, when we examine the number of differentially expressed genes per factor (Figure 5B), we observe that the numbers are highest in the factors ‘observation’, ‘histology’, ‘cell line’, ‘generation’ and ‘organism part’. It appears that, broadly, across species, transcriptional activity is strongly driven by its context: by tissue (‘histology’, ‘organism part’ and, by extension, ‘cell line’), followed by developmental stage and then cell type, while the main extrinsic drivers of transcriptional activity such as xenobiotic responses (‘compound treatment’) and disease states contribute to differential expression to a smaller extent. We can also observe that the number of differentially expressed genes is largely independent of the number of EFVs (the median factor value count is around 3 EFVs).

Programmatic access

REST (Representational State Transfer) is a simple technology that allows users to retrieve data in an easy-to-parse format by going directly to a web address Universal Resource Identifier (URI). For instance, all information available on gene matching ‘aspm’ anywhere in gene property fields can be obtained by entering a URI: http://www.ebi.ac.uk/gxa/api?geneIs=aspm.

This will produce a simple output either in JSON or XML format (the latter will be used if ‘&format=xml’ is appended to the URI). Example programmatic queries are provided in Table 3 below.

These APIs allow advanced users to search and retrieve complete information on any gene or experiment from the Atlas, including all gene and sample attributes, details of experimental design, meta-analysis statistics and gene expression values. Additionally, the Atlas provides a gene view-based Distributed Annotation System (18) track at http://www.ebi.ac.uk/gxa/das that can be viewed with any Distributed Annotation System client such as the Ensembl genome browser.

Examples of output for these queries are available from the Atlas documentation at http://www.ebi.ac.uk/gxa/help/AtlasApis.