Abstract
High-throughput gene expression studies commonly employ pathway analyses to infer biological meaning from lists of differentially expressed genes (DEGs). In toxicology and pharmacology studies, treatment groups are analysed against vehicle controls to identify DEGs and altered pathways. Previously, we empirically quantified false-positive rates of DEGs in gene expression data from pools of vehicle-treated zebrafish embryos to determine appropriate study designs (sample and pool size). Here, the same data were subject to Over-Representation Analysis (ORA) and Gene Set Enrichment Analysis (GSEA) to identify false-positive enriched pathways. As expected, the number of false-positive ORA results was lowest where pool and sample sizes were largest (conditions which also generated the fewest significant DEGs). In contrast, the frequency of GSEA false-positives generated through the fast GSEA (fgsea) algorithm increased with pool and sample size and was highest for simulations that generated 0 DEGs, with ribosomal gene sets significantly enriched with the highest frequency. We describe 2 distinct mechanisms by which GSEA generated these false-positive results, both of which are most likely to generate significant gene sets under conditions where expression differences are particularly low. Finally, GSEA analyses were repeated using 1 alternative GSEA algorithm (CERNO) and 11 different ranking statistics. In almost every analysis, the number of significant results was highest where pool size was highest, with ribosome as the more frequently enriched gene set, suggesting our observations to be generalizable to different implementations of GSEA. These results from zebrafish embryos suggest caution in interpreting any GSEA results in contrasts where there are no DEGs.