The UniFrac authors' stated justification for its tree-based design is that OTUs with higher sequence similarity should be treated as more similar when comparing samples because they tend to fill similar ecological niches. I don't know if this is a valid argument from the point of view of microbiology -- if anyone can point me at evidence either way, please let me know. With OTUs generated by QIIME, which are generally very noisy, it is essential to use UniFrac to suppress differences which are due to noise. However, with the much more accurate OTUs generated by UPARSE or UNOISE, I believe that UniFrac is generally a bad choice of beta diversity metric because it has very low sensitivity to differences in OTUs with high sequence similarity. Such differences are surely biologically significant even if the OTUs fill similar niches. My recommendation is generally to use weighted Jaccard unless you have a good reason to use something else.