edgeR vs DESeq vs bayseq

[Azazel]

I want to identify differentially expressed genes between two timepoints, from high-throughput digital gene expression profiling (Helicos), in mammalian cells. Sometimes I have replicates, sometimes not (depending on experiment/dataset).

I read about three programs which seem suitable, edgeR, DESeq and bayseq; however the differences between these three programs seem quite subtle to me.

What are the central differences between these three programs? How can I choose the one which fits my analysis needs best?

[epigen]

edgeR and DESeq use slightly different methods for estimating the mean and variance of the negative binomial distribution by which they model the read number distribution. The DESeq authors claim theirs is better. bayseq I don't know yet.
You could also use cuffdiff from the cufflinks package, it gives FPKM values instead of read counts. I tried this and DESeq. cuffdiff reported many more genes as significantly differently expressed than DESeq, while discarding some of DESeq's candidates, but I didn't have replicates that strengthen DESeq's power. Using at least two different tools and relying on their consensus is what I'd recommend. And of course experimental validation.