Hello to all,

It is frequently mentioned that maximum likelihood based methods are not suitable to detect selection on intraspecific sequence data. Why is that, is it simply because the level of variation is not enough? Is it because of recombination?

I have data set composed of 28 sequences from 14 diploid individuals from different populations and 2 other sequences from an outgroup species. Recombination is moderate. I run both REL and FEL methods and I get some signal of positive selection with both.  

Is this a valid approach?
Can I trust this results

Thanks a lot!

Hi Dan,

Wow, I've never heard that opinion.  It's certainly possible - if you're operating outside of the microbial world - that there may simply be insufficient phylogenetic signal in an alignment comprised of intraspecific sequences.  Given that you've got only 30 sequences from 16 individuals, I would be inclined to use REL only, as there is most probably insufficient signal to rely on FEL (where a unique rate is estimated for each site in the alignment).  

When you say that recombination is moderate, is this based on other evidence or have you run GARD?  You can always run REL on multiple partitions..

Cheers,
- Art.

I think the opinion about intraspecific sequences comes from the Kryazhimskiy & Plotkin paper (below). They simulated population genetic data with defined selection co-efficients and calculated pairwise dN/dS and showed these were inconsistent with the simulated selection coefficients. I guess the major crit is the use of pairwise dN/dS calcs. I agree with Art in that "if there ain't no signal to infer the tree, there ain't no signal for selection detection." cheers ./w


KRYAZHIMSKIY, S., and J. B. PLOTKIN, 2008 The population genetics of dN/dS. PLoS Genet 4: e1000304

Yeah, but if I remember correctly, those simulations were based on a classic biallelic model, and it doesn't make sense to look for an inherently directional selective signal using methods that were designed to detect diversifying selection (dN/dS)..

:)
- Art.

Thank you.

I said that maximum likelihood based method are very often said to be unsuitable for detecting positive selection based on two reviews: (often was maybe not the right word to use)

"molecular signatures of natural selection", rasmus nielsen, annu.rev. genet. 2005:39:197-218.
 "approaches for identifying targets of positive selection" Jensen et al. 2007, Trends in Genetics vol.23 No.11

My impression, is that most studies using intraspecific sequences  (outside the microbial world), very OFTEN :), do not use these methods and rely only on classic molecular population genetics tests.

I guess from your answer that people just lack enough phylogenetic signal and there is nothing intrinsically wrong about using these methods if one takes care of recombination.


When I said I have moderate recombination, my estimates were based on the minimum number of recombination events (Hudson 1985) estimate which is supposed to be conservative. The result was 7 minimum recombination events (similar to other estimates from other genes). Anisimova, Nielsen and Yang 2003 simulations, suggest that LRT is robust to levels of recombination of fewer than 3 recombination events in a sample of 10 sequences, "effect of recombination on teh accuracy of the likelihood method for detecting positive selection at amino acid sites" Genetics 164:1229-1236.

After your post, I run GARD and I found evidence for only 1 breakpoint.
...?...I need to think more about this difference to make  sense out of it.

But, when running REL and FEL, and REL using partitions I come to similar results. So I was (am) wondering how likely is it that these are false positives.


Thanks lot!