Dear Sarah,

Infinite ratios are typically an artifact of low divergence sequence where there are a few substitutions per branch, hence dS is inferred to be 0.
An infinite ratio is technically not a failure of the method, but rather the effect of data undersampling.

Another possibility is that your sequences are too short relative to the number of sequences. As set up by default, c-AIC model selection requires there to be at least 2N+14 for N sequences (because the number of model parameters, including frequencies must be less than the total number of sites less one). If your sequences are shorter than that, inference can become meaningless. I can modify the analysis to count the number of samples more liberally (as the number of characters in the alignment) to avoid this issue.

HTH,
Sergei

Dear Sarah,

An 'inf' estimate is not really a failure of the method. There could be two possibilities:

1). Imagine that at some branch the estimate of dN is 100 but that of dS is 0. The numerical estimate of the ratio is 100/0 = inf. Clearly, '0' is not biological, and simply suggests that we haven't sampled enough data to find the small but non-zero estimate of dS. This may be better understood, if one computed confidence intervals (e.g. profile likelihood) on dS and dN. Suppose for example that the 95% CI for dS is (0-0.1). In that case you can crudely approximate the CI for the ratio to be (1000-Inf) and the estimate of inf is effectively correct (i.e. a very large dN/dS).

2). Now imagine that the estimates are dN = 0.1 and dS = 0 and confidence intervals are 0.05-0.15 and 0-0.1, respectively. Now the point estimate of dN/dS = inf can be very wrong indeed, because the ratio can credibly lie in 0.5-Inf.

In the latter case there is not much one can do with any method - the variance of dN/dS estimates will be just too high.

Hope this helps,
Sergei

Dear Sarah,

The best way around the dS = 0 problem is to re-parameterize the model, to use separate alpha (dS) and beta (dN) rates on each branch, instead of defining beta = omega*alpha, running into problems when alpha = 0 and beta>0. It's a bit tedious to do (because the representation of the problem for the GA changes somewhat), but it can be done.

Cheers,
Sergei