Extending Gene Set Enrichment Analysis with Cancer Immunology Collections

23 Feb 2016

Gene Set Enrichment Analysis (GSEA) is a well-known and widely-used method in Computational Biology and Bioinformatics. GSEA uses a sorted list of genes (obtained by comparing gene expression levels between groups of patients) and a database of gene sets as input, and it checks whether members of a particular gene set have a non-random ordering, biasing them towards the top or the bottom of the list (i.e. enrichment). We leave the details of the method out of this blog post, but the original paper describing GSEA is a good starting point for newcomers.

Gene sets often represent biologically meaningful groupings: for example, all genes that are active during controlled cell death. It is common to see these gene sets being referred as pathways and it is for this reason that GSEA is also known as pathway analysis. This name, unfortunately, leads to the following misconceptions regarding the whole approach: 1) Pathways are well-defined gene sets and they do not change; 2) There is no room for customization of the database of gene sets used in the enrichment analysis.

First, there is no clear consensus over the contents of a pathway across many of the databases, let alone the researchers, and therefore our knowledge on the list of genes that are members of a particular pathway is constantly evolving. This means that our database of gene sets should evolve with our knowledge, but in reality, it is far behind our current knowledge due to the slow nature of the literature-based curation process.

Second, the enrichment analysis does not necessarily depend on default data sets and is flexible enough to work against custom gene set definitions when needed. As long as a researcher is confident that the gene sets used in the enrichment analysis reflect our current knowledge regarding biological processes and they are uniformly curated, there is no need to worry about using custom data sets as gene sets.

New Cancer Immunology Collections in GSEA

As part of one of our research projects focusing on immunotherapy in cancer, we needed to run GSEA on a gene expression data set to explore enrichments in patients with different outcomes (phenotypes). Our sample size was on the order of tens, which meant that we had low statistical power for the analysis. To account for this, we decided to stick to only one of the curated gene sets; but this collection was lacking gene sets that were highly interesting to us. Specifically we were interested in including tumor-associated antigens recognized by CD4+/8+ T-cells and gene signatures associated with immune cell infiltration to the tissue.

Our solution to this was to collect curated gene lists from relevant resources that are not part of the MSigDB, extract the official gene symbols and include them in our analysis as additional gene set collections. For the former, we used the up-to-date list of tumor-associated antigens from Cancer Immunity’s Peptide Database; and for the latter, we used gene sets inferred by Senbabaoglu et al that builds on the work of Bindea et al and infers gene expression signatures associated with immune cell filtration in kidney cancer:

Creating GSEA-compatible gene set collections

You will see from the examples above that the default Gene Matrix Transposed (GMT) file format features a separate gene set on each row, two metadata columns for each gene set followed by an arbitrary number of gene identifiers that belong to the corresponding gene set. The file format, therefore, is an extension of tab-separated values (TSV), meaning that almost any text editor or other third party software that can handle TSV format can be used to curate these files.

One way to tackle this problem is to generate these files programmatically and often researchers are interested in adding new gene sets to the analysis based on lists of genes published as supplemental material to papers. A practical solution that works nicely for us is to import such gene sets into Google Spreadsheets (for collaborative editing, taking care to ensure that gene names aren’t converted to dates) and then reduce them to GMT files to be used in the analysis. You can find a Python-based solution to this common task in this notebook as an example that converts gene lists extracted from the papers we mentioned above and turns them into individual GMT files.

Some words of caution about custom gene sets

Although adding one or more custom gene sets into the analysis is pretty straightforward, it still requires careful planning and some basic bioinformatics skills. It is, for example, really important to be consistent with the use of common identifiers across default and custom collections. Not all gene sets are curated with gene identifiers that are consistent across databases; therefore, gene identification mapping from one provider to another often becomes essential. Note that problems in such mappings can easily cause over- or under-representation of gene sets due to cross identification mismatches.

Furthermore, once these custom gene sets start accumulating, it becomes important to check the degree of overlap across multiple gene set collections to prevent redundant enrichment tests. Although this is an imperfect science, one way to check this would be to compare all pairs of gene sets across two GMT files and calculate an asymmetric score that represents to what extent the first collection is represented within the second collection.

Another thing to consider when using custom gene sets in the analysis is to run the analysis on the combined collection of gene sets (including both the default and the custom ones) and not individually to properly account for false discovery rate. With that, it often does not make sense to restrict the analysis to only the custom gene set, but it is good practice to combine it with another collection as a good background for the enrichment tests.

Final words

It is crucial for the community to share such custom gene sets and explain the rationale behind them for encouraging others to build on such efforts. We wanted to share our experience and solutions that worked for us so that not everybody has to reinvent the wheel.