Hi Tessa,

FEL infers dN and dS values using data from a single site. With 13 sequences, this means that the estimates are very imprecise -- they have very large variances associated with those values.
So your dN = 100 could really be dN = 100+/- 100, which is why the p-value is so high (since the likelihood ratio test accounts for the lack of information). You should NOT put too much trust into these point estimates.

Sergei

Hi Tessa,

First, to reiterate Sergei's point: one should not put much trust in point estimates obtained from small data sets. The purpose of the significance test is to incorporate information about the reliability of those estimates. The bottom line is that the small number of sequences should not be a concern, provided you accept the idea that lack of significance means the data are not informative enough to make a call either way (absence of evidence rather than evidence of absence).

Re alternative methods:
All of these methods should be robust against small sample sizes, in the sense of not giving you overly confident conclusions when the data are uninformative. So both REL and MEME should also be fine; I would  expect REL to also not give significant results, because empirically it tends to give results very similar to FEL. The situation where REL is likely to differ from FEL is when there are many sites under positive selection: this would make REL more likely to think that a borderline site is also under positive selection, because the underlying assumption is that different sites are likely to experience similar selective pressure. The FEL vs REL debate really boils down to the validity of this assumption.

MEME is a FEL method over sites (and REL over branches), so should be comparable to FEL except that it can also detect evidence for one or more short bursts of selection - so if the data contain evidence of this you might still get a positive result.

Hope this helps,
Konrad