Question with regards to the method's assumptions

[pkmnsandy]

Hi,

Thank you for developing and making this software available, it is very interesting and potentially very useful for my work. I have read the associated paper and would appreciate some clarification regarding the assumptions of the model.

I am considering using this software to infer genetic clusters in our model system, a grass species. This species is highly selfing, such that deviations from Hardy–Weinberg equilibrium are common. It also exhibits a relatively high proportion of rare alleles (we didn't strictly filter them out as we hypothesized that they may have a role to local adaptation). In addition, dispersal can occur over long distances via seeds, although it is not clear to what extent this translates into effective gene flow among populations.

What is not entirely clear to me from the paper is how the model handles the following aspects:
(1) The assumption of Hardy–Weinberg equilibrium within inferred clusters, and how strong inbreeding/selfing might affect inference. (2) The treatment of rare alleles and whether they can influence cluster assignment or model stability. (3) The extent to which long-distance seed dispersal, in the absence of substantial pollen-mediated gene flow, impacts the model’s ability to detect population structure.

Any guidance on whether the conStruct is appropriate under these conditions, or on recommended data filtering or parameter choices, would be greatly appreciated.

Thank you very much for your time.

Best,
Sandy

[gbradburd]

Hi Sandy! Sorry for the slow reply - I'm on sabbatical right now and have been much slower on computer stuff as a result. Anyway, seems like an interesting application! My thoughts are:

1. I don't think a lot of selfing should bias conStruct's inferences. It doesn't actually make use of HWE to, e.g., calculate the probability of different genotypes as a function of allele frequencies. Instead, it's calculating the probability of covariance in allele frequencies (that is, it's taking the genotypes out of the equation and just looking at allele frequencies). So, if local populations tend to be highly inbred, I don't think it should necessarily make conStruct more or less likely to lump them into one large inferred population with a sharp decay of isolation by distance vs. split them out into discrete ones.

2. I don't think rare alleles will bias conStruct's inferences. They'll tend to affect variances within populations rather than covariances between populations, and conStruct is more concerned with the latter. That said, I think the safe choice here would be to run conStruct with different MAF thresholds and see what happens.

3. Long distance seed dispersal will affect conStruct's inferences to the same extent it affects population genetic structure. If seed dispersal sometimes results in successful germination, that will tend to homogenize genetic differences that have arisen locally in different populations due to selection and inbreeding/drift. In turn, that homogenization will make conStruct more likely to (correctly, imo) describe that variation using a single discrete cluster rather than several distinct ones. On the other hand, if long distance seed dispersal is never successful (and you're not genotyping those long distance seed travelers), then it'll be invisible to the genetic record and won't affect conStruct whatosever.

Hope that helps!