Hi All,

Is there documentation for the various ways HyPhy can test for episodic selection? I understand it can do a branch-sites test a la PAML, a new an improved branchsiteREL, TestBranchDNDS.bf, and perhaps lineagedual would work. Are there others?

My specific problem is that I am testing for episodic selection in a *very* conserved protein. There are four amino acid changes along the branch that I care about; two of these aa changes have functional consequences (and are fixed in descendant lineages) and two do not  functional consequences (these two changes appear to be evolving neutrally). There also appears to be synonymous rate variation across the tree.

Ideally I would like to test for an elevated dN along this branch accounting for syn rate variation (though that may not matter for testing dN), and in the best of all possible worlds, use a branch-site like to see if the functionally important aa changes were positively selected while the non-functional aa changes fall into some other site class.

I suspect this means that the improved branchsiteREL and TestBranchDNDS.bf are the best methods (????)

Thanks
Vinny

Hi Vinny,

I'll email you the manuscript (currently in review at MBE) where we describe BranchSiteREL.bf

An example of use can be found in (Multimedia File Viewing and Clickable Links are available for Registered Members only!!  You need to Login)

Let me ask you another question, however:

Do you want to know which lineages are under episodic selection, or which sites? The former question can be answered by (BranchSiteREL) losing the resolution in sites, and the latter -- by losing the resolution in branches. We are currently developing a method to answer the latter question, and I can point you to which (unpublished) analyses to use for this purpose.

Sergei

Hi Sergei (and others),

Thanks for the info/reference. Since I am interested in selection along one (or two) specific lineages I am most interested in the identification of sites with evidence for directional selection.

I have been using the dNdSRateAnalysis.bf with dual and lineage dual because there is significant site-to-site variation in dS among codons. I am curious what your (and every else how might want to chime in) thoughts are about using the lineage dual model (when it is significantly better than dual) to ID sites under selection, and then mapping amino acid changes at those sites onto the phylogeny to ID lineages in which they were selected?

Best,
Vinny

Hi Vinny,

I've never properly benchmarked the 'Lineage Dual' model, but we have one under development which combines the features of BranchSiteREL and FEL to find sites under episodic selection. The code isn't quite polished, but I'd be keen to see if it helps you if you are interested.

Sergei

I am definitely interested.

The lineage dual model actually seems to give meaningful results (when interpreted in a biological context), but if it hasn't been bench marked are the results likely to be unreliable?

It is interesting that the dual model does not identify any sites under selection, but that lineage dual does. My naive interpretation is that this suggests no evidence of diversifying selection, but evidence for episodic directional selection at the sites ID's from lineage dual.

Best,
Vinny

Hi Vinny,

Not necessarily unreliable, it's just that I've never performed power/accuracy simulations using this model, so I don't have a good feeling when it would work well for individual sites.

What happens in your data if you just fit the free-ratio model -- no site-to-site rate variation, but branch-to-branch variation?

Sergei

Hi Segei,

Fitting a free ratio (in PAML at least) suggests there is a lot of variation in the magnitude of the rates, but it is limited to 2-3 branches with most others having a low rate. The free rate model has essentially the same likelihood as the one rate model. This is not terribly surprising since it is a small very conserved protein; ~150aa, with only 10 aa changes in the part of the tree I care about (composed of 8 mammals with a basal divergence of ~40MYA).

What would be involved in accuracy testing? (I assume it can be done within hyphy???).

Best,
Vinny

Hi Vinny,

In this setting I can believe that the lineage dual model will work well -- it basically tunes the mean of the branch dN/dS ratio for each branch, and sounds like for your data set the means are elevated along the branches of interest.

Accuracy testing would involve simulating neutral data under your tree and seeing if Lineage Dual picks up false positives. If you send me your alignment/tree by e-mail, I can run the simulations on the cluster without too much effort.

Sergei

Hi Sergei,

I've been looking through the various output files from the dNdSRateAnalysis.bf trying to get a better understanding of what the models are doing. Are the "r" values in the .fit file for Lineage Dual (for example: givenTree.Node6.r=5.242482728475014) the branch correction parameters (omega_b) in this model?

Thanks,
Vinny

Hi Vinny,

Yes, the 'r' parameter is the ratio of the means (not the mean of the ratio, as the standard omega) of the beta (non-syn) and alpha (syn) rates.

Sergei