Extending gene set variation analysis with a reference dataset to stabilize scores.

Background: Biological pathways are sets of genes that jointly drive biological processes. Rather than analyzing genes individually, it is common practice to summarize sets of related genes using gene set variation analysis (GSVA). In short, GSVA summarizes a set of genes into a single score bounded between -1 and 1, where negative values suggest downregulation and positive values suggest upregulation. Although this interpretation is simple in theory, it depends on unbiased estimation of individual gene distributions. In the current version of GSVA, gene distributions are estimated using the input dataset (i.e., the scores are calculated based on the gene distributions from the same dataset). This becomes a major issue when study data does not adequately represent the full distribution of the population. For example, if RNA-seq data was collected on an imbalanced sample (e.g., more disease samples than healthy controls), it would be difficult to discern abnormalities in pathway activity since the gene distributions were estimated on a biased population. Therefore, we propose reference stabilizing GSVA (rsGSVA), a solution to this commonly ignored limitation by using reference datasets to estimate the gene distributions for a more stable GSVA score.

Results: rsGSVA shows comparable power to classic GSVA, singscore, and ssGSEA under ideal settings while demonstrating stable scores on sample subsets. An application on irritable bowel disease highlights interpretational advantages of rsGSVA to other methods in up/down regulation of inflammation signatures.

Conclusions: The rsGSVA technique enhances the GSVA functionality by incorporating a reference dataset. This integration of a reference dataset makes the enrichment scores independent of the input distribution and ensures their stability and reproducibility, even as samples are added or removed.