Dear Sergei,

I have come across a recent paper which used GARD for recombination detection (Multimedia File Viewing and Clickable Links are available for Registered Members only!!  You need to Login). The authors were able to detect recombination breakpoints in the genes of interest, but they claimed that the alternate topologies based on the putatively recombinant segments were not statistically supported by the Shimodaira-Hasegawa (SH) test. They then suggested that these potentially recombinant areas might be due to substitution rate variation and not recombination.

They've also made a similar claim in their previous paper (Multimedia File Viewing and Clickable Links are available for Registered Members only!!  You need to Login).

My questions are:
1. Is it reasonable and appropriate to use the SH test to "verify" the recombination breakpoints detected by GARD? To me, it seems that the SH test is too conservative for this purpose.

2. Indeed, GARD is able to detect "recombination breakpoints" not detected by other programs/ methods. However, could this be due to substitution rate variation, as these authors have claimed?

3. Following the above question, given that GARD seems to be one of the most sensitive recombination detection methods available, do you have any suggestions or ideas on how to "verify" the recombination breakpoints results by GARD?

Many thanks.

Regards,
olorin

Dear Sergei,

Thanks. Your reply is most helpful. I will put further thoughts in this matter.

olorin

Dear Gregonomic,

Traditional rate variation along the sequence (i.e. the kind that affects all branches in the tree) can be easily modeled with existing models (e.g. gamma rate variation); one can however imagine a more involved process where only parts of the tree are affected by rate/model variation (e.g. heterotachy) - if enough branches are involved, or if the variation shortens some branches and lengthens others at a given site, then the method can be confused.

HTH,
Sergei

Dear Greg,

You are quite right in saying that there exist substitution rate variation scenarios which may confuse any recombination detection method: e.g. convergent evolution along a sequence region can be considered one of them (i.e. changes in substitution rates which modify the nature, not just the speed, of the substitution process). That said, in recombination detection literature, 'rate variation' is usually assumed to be the typical site-to-site rate variation (e.g. in the Posada/Crandall PNAS simulation paper). I should have been more specific in the GARD paper discussion and qualified substitution rate variation

SH and KH tests are identical when only two fixed trees are involved - sorry for loose terminology on my part.

Going back to the JVI paper; it's indeed difficult to verify the claims not having the data at hand; however, really short or low sequence divergence fragments that do not support any phylogeny particularly well (i.e. have similar likelihood scores on a number of phylogenies) also will not likely be the primary contributors to the c-AIC improvement, i.e. if there had been any strong c-AIC derived signal, it would have likely come from fragments that support a particular type of phylogeny.

Cheers,
Sergei