Abstract
Background: A differential gene expression analysis may produce a set of significantly differentially expressed genes that is too large to easily investigate, so that a means of ranking genes by their biological interest level is desirable. The life-sciences have grappled with the abuse of p-values to rank genes for this purpose. As an alternative, a lower confidence bound on the magnitude of Log Fold Change (LFC) could be used to rank genes, but it has been unclear how to reconcile this with the need to perform False Discovery Rate (FDR) correction. The TREAT test of McCarthy and Smyth is a step in this direction, finding genes significantly exceeding a specified LFC threshold. Here we describe the use of test inversion on TREAT to present genes ranked by a confidence bound on the LFC, while still controlling FDR. Results: Testing the Topconfects R package with simulated gene expression data shows the method outperforming current statistical approaches across a wide range of experiment sizes in the identification of genes with largest LFCs. Applying the method to a TCGA breast cancer data-set shows the method ranks some genes with large LFC higher than would traditional ranking by p-value. Importantly these two ranking methods lead to a different biological emphasis, in terms both of specific highly ranked genes and gene-set enrichment. Conclusions: The choice of ranking method in differential expression analysis can affect the biological interpretation. The common default of ranking by p-value is implicitly by an effect size in which each gene is standardized to its own variability, rather than comparing genes on a common scale, which may not be appropriate. The Topconfects approach of presenting genes ranked by confident LFC effect size is a variation on the TREAT method with improved usability, removing the need to fine-tune a threshold parameter and removing the temptation to abuse p-values as a de-facto effect size.
