How to identify gene expression signatures from gene expression data?

Question

I have TCGA gene expression data. I'm interested in identifying gene expression signatures using the data.

I would like to know whether there are any tools or R packages for identifying gene signatures.

How are gene signatures different from GSEA?

Answer 1

There is no golden/standard way to define a gene signature but they are completely different from a gene set enrichment analysis (GSEA). I will start with how to obtain a gene signature:

Usually this kind of signatures are defined by comparing a group against another ie doing a differential gene expression analysis. One then selects the most up or down regulated genes and defines them as the signature of the group of cells. That's how the gene sets of the MSigDB which are not from (or derived from) pathway databases or gene ontologies are defined. There is no gold threshold of a fold change or a specific number of genes, it (highly) depends on the experiment you run and how good are the isolation of the cells.
Tools: DESeq2, limma, edgeR,...

As suggested by gringer, another way is to find which genes are expressed together (independently of how do they relate to other groups). This method relays on having a good enough number of samples from the same condition to be able to identify which genes are expressed the same way in those cells. If these genes are not expressed that way in other cell lines, then you have your signature for this group. Again, there isn't a golden method, different types of co-expression networks can be build and depends on how specific do you want to be or which is the data provided.
Tools: WGCNA, diffcoexp, TReNA,...

This is different from a gene set enrichment analysis¹, which consists on evaluating if a group of genes (usually a gene set, a pathway or genes from a specific gene ontology) are not randomly sorted in a sorted list of genes (usually this sorted list is the fold-change of a differential analysis).

There are several methods developed depending on the null hypothesis, the type of sub-method and what do they take into account...
Tools: GSEA, PLAGUE, GSVA, limma, fgsea,...

TLDR: To perform a gene set enrichment you don't need a gene signature but a gene set (that can be a gene signature or a pathway or a random group).

1: Note that gene set enrichment analyisis is different from gene enrichment analysis. For a good review on the methods I recommend this article

Answer 2

If you're interested in an unbiased approach you could have a look at Gene Correlation Network Analysis. WGCNA is a fairly old R package, with a somewhat confusing workflow and results that can be difficult to interpret. It also needs a large sample set, preferably from fairly different biological conditions, to produce meaningful results.

Asking how "gene expression signatures" differ from GSEA is like asking how surface textures differ from trademarks. Gene set enrichment analysis is for looking at gene rankings and seeing how they compare to a known / existing gene set. A gene set is considered "enriched" if the majority of the gene set appears in high-ranked genes (or alternatively in low-ranked genes), rather than a more even distribution of the set across the ranked gene list. Most commonly people will rank genes based on expression (or fold change), but the method is agnostic to the nature of the ranking.

Answer 3

Gene Signatures are sets of genes to report a consensus signal or function. They are not defined in a particular way but can be defined by any chosen group depending on the biological function that you are interested in. Typically these are defined by differential gene expression analysis or principal component analysis, although other approaches such as network analysis or time-course data could also be used.

There are plenty of examples of sets of genes published as gene signatures (especially in the field of cancer research). Those that have been cross-validated in other datasets or replicated by other research groups are ideal. If you wish to generate such a “signature” for a particular biological function, you could follow the approaches demonstrated in these published signatures (especially those that have been widely used and cited).

A gene expression signature is used to generate a consensus expression signal for these groups of genes (representing the given biological function). A widely accepted way to do this across samples of bulk RNASeq (or microarray data) is a “meta-gene” from the first eigenvector (v[,1]) of the singular value decomposition (SVD). This dimension reduction identifies a major contributor to variation across samples (in these genes) without negatively correlated genes averaging out. Dimension reduction can also be achieved by PCA and tSNE, depending on the requirements of your analysis.