I am testing DESeq2 1.10.1 version and find some issue with cleanContrast function.

I use a standard 2x2  A*B design with interaction terms. 

DESeqDataSetFromMatrix(countData = round(datacounts), colData = designtable, design = ~ Gender*Site)

dds<- DESeq(dds, modelMatrixType="standard")

output <- results(dds, independentFiltering=T, contrast=c(0, 0, 1, 0))

When I extract the contrast results to get the site M vs site K when gender is Female with the contrast vector, I found that some of the all zero contrasts are not properly filtered.  I debug the code using RStudio and find that in CleanContrast function , when a numeric contrast vector is used, contrastAllZero <- contrastAllZeroNumeric(object, contrast) is called. If I get the logic right, this function is used to detect genes whose counts for samples used in contrasts are all zero. However, after checking the source code of contrastAllZeroNumeric, it will do nothing for my contrast vector above. 

### in contrastAllZeroNumeric function ###

if (all(contrast >= 0) | all(contrast <= 0)) {
        return(rep(FALSE, nrow(object)))
    }

I kind of feel this is a bug.

hi John,

Thanks for the report. The extra leg work in DESeq2 to zero out the log2FoldChange, lfcSE, and set the p-value to 1 only is implemented for simple multi-factor designs and not for designs with interactions. This was a feature request from a user a year or so ago, who had an experiment with size factors differing by more than an order of magnitude and confounded with group (this is problematic for many reasons). Here, comparing a group with all zeros against another group with all zeros gives a paradoxical non-zero LFC. You can observe the same with glm() and providing a non-zero offset. So it's not really a bug, but I knowingly only implemented this for the simpler designs. I will add some comments in the code to state this explicitly.