I'm using the rlog function in the DESeq2 package and I notice a quirk in the transformed data that I do not know what to make of: for genes that are expressed in a small proportion of samples (say, for gene X, 10 samples have non-zero raw counts out of 300 samples), the transformed dataset has no zero values at all; instead, the majority of samples have some other value that is either negative or positive. Negative count doesn't make sense so I could, I suppose, deal with that by zeroing all counts less than 1 in the transformed dataset, but I don't know what to do about the cases where most samples have a positive value, say 3.5, and a small proportion have other higher values- it's as if the zero-level for that gene is shifted to a small positive number. This is seen only with genes expressed in a small proportion of samples, and the amount of shift, positive or negative, varies across genes.  I notice the same with variance-stabilizing transformation and regardless of whether I set blind=FALSE or not.

Have others noticed this with their dataset? If so, how did you deal with it? I don't know how much of an impact this would have on the results of clustering-type exploratory analyses, but I am also not comfortable with seeing that a gene that should not be expressed at all in most samples has positive counts for all of them.

"Negative count doesn't make sense"

First, Ryan is correct that rlog and VST return log2-like values, so negative values are normal, and simply indicate an expected count less than 1. Many samples will have expected counts less than 1 in a very sparse dataset.

Secondly, the rlog and VST may not be optimal for very sparse data. If I were you I would compare to other transformations and pick based on properties such as the stabilization of variance over the mean (see vignette) and preservation of signal (seen for example through a PCA plot).