what you are measuring is differentiation based on a set of loci between populations. In other words you are comparing some measurements between populations. Thus, the family of comparisons should be the number of comparisons between populations as far as the multiple testing procedure is discussed.

Dziękuję Ireneusz! (unfortunately this is as far as my Polish goes :)

If I understood you correctly, and I must say I tend to agree with you after doing some more research since posting this question yesterday, the alpha value (0.05 before Bonferroni Correction) should be revised by dividing the total number of comparisons made, e.g. at a single locus, if n number of population datasets are to be compared, then the total number of pairwise comparisons made would then be (n x (n-1))/2, etc. Now considering that comparisons were made for each locus independently, then the same revised alpha value should be used for p-values at each different locus, of course provided that the same number of population datasets were used for comparisons at each of the loci analyzed.

Kind regards

how about making per each population a vector of these genetic distances, and compare these vectors treating them as individual observations for given population. By that comparing such 2 vectors between 2 populations it would be like comparing for example height measurements of 2 sets of individuals (so to put is simply a t-test between means). The loci should be independent of each other in terms of LD. But that is treating comparisons between populations as a family of comparisons. 

Other way is doing what you say, that is takin each locus compare between populations and correct for the number of loci. Depends what you are looking for: the loci that differentiate populations, or average differences between populations. You can do both and then correct p-values respectively as discussed, as those 2 scenarios should form distinct comparisons families, and with that you get 2 answers: which populations differ, and what drives this difference (although regression might be better)

You can check SGoF methodology for multitest adjustments, it do not allow as lower rate of false positives as FDR, but it has the advantage of increasing statistical power when increasing sample size, such method can be applied for each comparison separately, considering the particular number of test used (just ask SGoF in google). Alternatively, you can just use chi-square (or G test) to compare the number of positive-negative results per pair of populations...

Many thanks to Ireneusz Stolarek and Emilio Rolán-Alvarez for their kind feedbacks!

There is a problem with the queston

Is the p value of 0.05 the correct value in the first place. 

This is the conventional level but there is a lot of debate about p values should be given the status that they are frequently accorded.

If you want to go down the route of p values the a Bonferroni correction is a good idea.

A Bayesian approach might be an alteernativ method.