We have a problem with our DE analysis where we are intersecting results and manually removing genes. This is an overview of the samples and the analysis plan.

We have a total of 8 conditions (cell lines): • control • samples with transfected with Gene 1 • samples with transfected with Gene 2 • samples co-transfected with, both, Gene 1 and Gene 2

There are two versions of the above 4 conditions at different temperatures, one set at 32C and one at 37C. When the 2 genes are expressed in parallel at 32C they activate a downstream signalling, which is not activated when each gene is expressed individually or when the 2 genes are expressed in parallel at 37C. It is the gene expression signature of the downstream signalling that we want to identify.

The original analysis plan to identify the gene expression signature promoted by Gene 1 and Gene 2 combination involves comparing every transfected sample to the control at their respective temperature, then keeping only the genes of the co-transfectant that are not found in the individual Gene 1 or Gene 2 results, at both temperatures, and finally from the gene signature of the co-transfectant at 32C keep only the genes that are not also found in the signature of the 37C equivalent.

The issue with the above method is that intersecting results and manually removing genes that are found in other lists is not statistically sound as it erases the false discovery rate control performed by DESeq2, and false positives/negatives will have an effect on what genes remain or are filtered out in the final gene list.

The main question is whether anyone can recommend a more statistically correct way of doing this analysis given the research aim, through the inclusion of multiple factors or interaction terms in the DESeq2 design formula. Right now the design matrix consists of a single column with the 8 separate conditions explained previously, and only this factor is included in the design formula.

Testing an interaction term, depending on how you set up the design, would get at e.g. whether the effect of transfection of gene 1 and gene 2 is non-additive on the log counts. By adding an interaction with temperature, you could test whether the interaction of gene 1 and gene 2 is different across temperatures. It sounds like this would achieve your final set -- to some degree.

However, if you additionally require that features are not DE with gene 1 or gene 2 alone, you cannot obtain this with an interaction design, but you have to do intersection of different FDR sets. You can use altHypothesis="lessAbs" to test for genes being not DE, after specifying lfcThreshold as an LFC of no practical significance.