At different places (the help page for DESeq2::plotMA(), the apeglm vignette, and here) an s-value threshold of 0.01 or 0.005 is mentioned as a sensible choice for "false sign" (FS) events. However, when specifying a non-zero threshold for the log2 fold change in DESeq2::lfcShrink(), the reported s-values refer to "false sign or smaller" (FSOS) events. In this case, the maximum possible value for the s-values is not 0.5 anymore, but 1 (the maximum possible value was stated here as a reason for the thresholds of 0.01 and 0.005). Thus: Is there a recommendation for what is a sensible choice for an s-value threshold for FSOS events? (I know that thresholding per se is a controversial approach in statistics, but let this discussion be aside.)

Good question. I’d say roughly the same values make sense though. When setting a threshold and calculating FSOS posterior probabilities, you are being more conservative about what should come through as a “discovery” in the final list. Bounding the aggregate rate at say 1% should give many large effect genes in a typical RNA-seq experiment. But it’s always good to do some power analysis first. Since we aren’t doing single gene pvalue based testing it would make sense to do power analysis based on simple simulation motivated from real datasets similar to the one under consideration.