cufflinks and non-uniquely mapped reads

[clariet]

Cufflinks appear to take not only uniquely mapped reads. I only filtered out low quality mapped reads and ran cufflinks without any problem. But how cufflinks take care of those reads mapped to multiple regions? Does it only pick the region with highest mapping score? Anyone has ideas?

I am wondering how the results will differ if I use uniquely mapped reads only.

Any input will be highly appreciated.
Clariet

[lpachter]

The correct way to deal with reads that map to multiple places is explained in the paper by Li et al:

http://bioinformatics.oxfordjournals.org/cgi/content/abstract/26/4/493

Mathematically, the ambiguity that arises from multiple hits in alignment, is no different than the ambiguity that arises when reads can belong to multiple isoforms at one location- in fact the statistical deconvolution is identical.

We plan to extend Cufflinks to correctly solve this problem in the near future.
Currently, multiple hits are assigned equal probabilities- details are in the supplement to the Cufflinks paper:

http://www.nature.com/nbt/journal/vaop/ncurrent/extref/nbt.1621-S1.pdf