Hi Sergei,

I'm curious how one should interpret MEME results when no sites are identified with dN>dS (i.e., in the table where  synonymous and non-synonymous substitution rates over sites are shown), but when a particular site (which we a priori know is a function altering aa change) has post. prob. of 0.995 and empirical Bayes factor of selection >1000 along the lineage where we know the function altering aa change occurred?

In other words does one interpret this situation as no evidence of diversifying selection at this site across all species (from the FEL model fit), but some evidence for an episodic selection event at that site in a particular lineage (from the MEME model fit)?

Thanks
Vinny

(With the obvious caution against placing too much importance on the finding of selection at any particular branch-site combination.)

Hi Vinny,

Am I correct to assume that the p-value from MEME is significant for site X, but the p-value from FEL is NOT significant for the same site?

What MEME tells you is that there is evidence for selection at a proportion of branches, inferred as a part of the model, along that site (i.e. there is episodic but NOT pervasive positive diversifying selection).

The actual branches under selection at that site are really only good for exploratory data analysis; there is no strong statistical significance you could or should derive from this result.

Sergei

Hi Sergei,

Thanks for the reply. Actually I'm curious what it means when the p-value is not significant (in either MEME or FEL), but sites nevertheless have high PP and BF of selection?

We would never use the MEME results to argue that any particular aa change was selected in any particular lineage. But are using the results in exploratory analyses, we have functional data indicating a couple sites are functionally divergent in different species and were curious what MEME might infer was going on there.

Hi Vinny,

The PP and BF for the exploratory MEME analysis are not terribly reliable and that's because Empirical Bayes (used to find them) does not account for variability in point estimates.
For example, if a site has a point estimate of beta- = 0.1 and beta+ = 2, but the variance of beta+ is high (e.g. we can confidently bound it away from 1), EB will take the point estimate of 2 at face value, even though it could have also been anything between 0.7 and 4 (hypothetically speaking).

Exploratory analysis is another matter -- here you probably want to err on the side of sensitivity versus specificity, so you suggested use is probably OK.

Sergei