I asked this question on SeqAnswers.

Does anyone know how DESeq2 handles genes that have zero counts in one condition and >0 counts in another?

My results output shows that these genes have a positive log2 fold-change value, but I do not understand how DESeq2 arrives at this number if it is taking the log of a ratio, in which the numerator is divided by zero.

Count data (letters are conditions; numbers are replicates):
 

> dds <- DESeqDataSetFromMatrix(countData = countData, colData = colData, design = ~ condition)
> dds$condition <- factor(dds$condition, levels=c("A","B"))
> dds <- DESeq(dds)
> res<-results(dds,independentFiltering = F)

Results table:

I know that DESeq1 gave an 'Inf' value in these cases, but how does DESeq2 arrive at a real number value? 

Thanks

While the maximum likelihood estimate (MLE) of DESeq goes to Inf, the use of a prior distribution on LFCs (log fold changes) in DESeq2 gives us a finite estimate. The way to interpret this is that: zeros might indicate absolute no fragments in samples of A, or more likely that the expected counts of fragments is some positive value below 1. If we were to increase the sequencing depth by 10 or 100, etc., we might start to observe some fragments in A. The prior distribution for LFCs is estimated by looking at the distribution of MLE fold changes observed, including other genes where the sequencing depth is higher, and using this range to give a finite estimate here. (See our paper for full details http://genomebiology.com/2014/15/12/550/abstract .) So the estimate here depends on: the dispersion for this gene, how large the counts are for B, and the distribution of log fold changes for other genes which had finite MLE LFCs.

Thank you for the explanation Michael.

The link you posted is dead, but I assume you mean this paper?

Love, M. I., Huber, W., & Anders, S. (2014). Moderated estimation of fold change and dispersion for RNA-seq data with DESeq2. Genome Biology, 15(12), 550. doi:10.1186/s13059-014-0550-8

Ah. It's because you included the period and parantheses in the hyperlink.