Hi,

I am having some problems to interpret REL results.

I am getting very large coefficients of variation for both dS and dN when applying Nonsynonymous and Dual models to two fragment alignments.
For alignment 1 with Dual I get CV(dS)=4.99 and CV(dN)=5.82 and for alignment 2, something maybe more reasonable,  CV(ds)= 2.09 and CV(dN)=2.84.
Do you have an idea what could be behind this values?

Another problem is: I have applied Yang's REL models to alignment 1 and neutral models are always preferred to selection models (namely in the comparisons M2 vs M1a,M8 vs M7).
M3 is much better than M0 (p<0.0001) and reveals a few positively selected sites. When I apply Dual to the same alignment, first I get the big CVs I mentioned above and consequently
a sign that there is a high synonymous rate heterogeneity (p=0.000) and second the dual model detects a few positively selected sites as well.  Could I assume that M8 and M2 fail to detect selection due to having ignored heterogeneity in dS and take the results of Dual as a sign that selection is occurring? Or should Dual be interpreted  more like a test of the heterogeneity of dS in the sequences than like a selection test?

Thank you very much.
 

Dear Dan,

A large CV value by itself is not a problem; it simply tells you that the distribution is "flat" or multimodal, i.e. the standard deviation is large relative to the mean. Could you run your alignment through REL on Multimedia File Viewing and Clickable Links are available for Registered Members only!!  You need to Login (which is basically the Dual model) and paste in the URL of the results?

If you run your alignment through PARRIS on Multimedia File Viewing and Clickable Links are available for Registered Members only!!  You need to Login, this will provide you with a direct comparison against PAML, because PARRIS is effectively an M2 vs M1a comparison allowing for variable dS. The Dual model is a bit more difficult to compare to PAML.

Best,
Sergei

Hi Sergei,

Thank you for your answer.I run my data on datamonkey and here are the results (sorry I couldn't post a url link so I send you a file with the outcome)


Running the data through PARRIS effectively rendered similar results to Paml, namely no positive selection.

I do not really know how to interpret the differences between these groups of tests and how to choose between them. Could you give me your opinion? Thanks a lot.

---

Hi Dan,

For smaller data sets like yours (19 sequences), it may difficult to reliably infer individual sites under selection. PARRIS performs a "global" test (i.e. is there a proportion of sites with dN/dS) on your alignment, and if that comes back negative, then the inference of individual sites under selection (REL) is unlikely to be reliable.

HTH,
Sergei