4 weeks ago by
Using the latest version of DESeq2 (v1.16), the maximum likelihood estimate of the LFC will be something like log2 of the mean of normalized counts in the group with positive counts. We include a threshold on how low the expected value of the counts can go, which stabilizes the methods and prevents the LFC from going to +/- infinity. So the MLE LFCs can grow very large, e.g. if the group has a mean of 1024, you would get an LFC close to 10.
We recommend using shrinkage after running DESeq() to produce LFCs that are more suitable for ranking or visualization, e.g.:
res <- lfcShrink(dds, coef=2, res=res)
Then the LFC is a posterior estimate, given the counts for that gene, and the distribution of observed LFCs from the experiment. This "moderates" or pulls in the noisy LFC estimates from genes that have a lot of variability or low counts. You can read more about this in the DESeq2 paper: