What have you tried so far? One sensible solution might be to try:
- whatever you would try if you didn't have to worry about low counts; maybe K-means
- some out-of-the-box method such as PAGODA. http://hms-dbmi.github.io/scde/pagoda.html
If those don't get you to where you need to be, then where's the gap? Do you think the estimates are off, and if so, why? Or maybe you just need more reliable quantification of uncertainty?