I use HOMER often for de novo and known motif discovery, but I have found that when the number of input sequences are low the results of motif analysis become inconsistent on repeated runs. This is easily demonstrable on ChIP-Seq data sets after a lot of peak filtering, and I am now running into it with HITS-CLIP motif discovery (in RNA mode) for top CIMS sites of ~200 clusters ('peaks'). Is HOMER still reliable in these cases? How do you deal with this? I am looking into using MEME instead but have not yet figured out all the settings needed to replicate the same level of motif annotation that HOMER gives by default.
So HOMER de novo motif discovery works by comparing peak sequences (binned by length) to background sequences (of matched length) and calculating enrichment. It randomly generates the background sequences with every run, so you would expect some variability over different runs. I worked around this by running the de novo motif analysis multiple times and then mentioning that a couple of positions showed the most variability in my presentations, and I will probably do the same in the paper I'm writing.