Considerations of our approach to infer adaptive gene-flow

Whilst admixture appears to be an almost universal characteristic of the populations that we use in this analysis, we only find evidence of substantial adaptive gene-flow in three populations: the Fula and Jola from the Gambia and the Ju/hoansi from Namibia. Importantly, we infer significant recent admixture involving diverged source populations in these three groups, from Eurasian-like sources in the Gambians, and East African Asiatic-like source in the Ju/hoansi [Supplementary Fig. \ref{fig:admOverview}]\cite{Busby2016AdmixtureAfrica}. Our method is likely to have greater power to detect deviations in these scenarios. As such, it is perhaps not surprising that across populations from the Central West African ancestry region, where we previously inferred admixture events involving mostly closely related African sources\cite{Busby2016AdmixtureAfrica} we infer a single example of adaptive gene-flow in the Namkam from Ghana at FZD1 on chromosome 8 [Fig. \ref{fig:figSharedAncestry}]. It is important to note that even though all Gambian populations in our analysis had similar sample sizes and were painted using exactly the same procedure as the Jola and Fula, we do not see similar numbers of local ancestry changes in them, suggesting that such deviations are not linked to our chromosome painting procedure.
In addition to working with populations that have substantial recent admixture, there are further considerations to the results that come from our approach. The local ancestry inference procedure that we have used\cite{Lawson2012InferenceData} is based on a fully probabilistic framework. To access uncertainty in local ancestry with the paintings, we generated 10 realisations (sampled paintings) of the painting algorithm and inferred ancestry proportions from these paintings. Because of the stochastic nature of such an approach, we are unlikely to capture the full uncertainty in local ancestry, but chose this approach based on computational requirements and the use of a similar approach in previous admixture inference work\cite{Hellenthal2014AHistory}. Nevertheless, future work to optimise the computation issues of storing the \(n*n\) posterior copying probability matrix at each locus in the genome will likely provide a clearer assessment of local ancestry uncertainty. We used a binomial likelihood model, which takes each ancestry separately and models genome-wide proportions against local deviations. Using this model, we found no evidence that more than one ancestry increases or decreases in concert at a locus, at least in the Fula [Fig. \ref{fig:figVarCovar}]. Modelling changes across all ancestries at the same time, for example by using a multi-variate normal model, might provide further insights into part of the genome where we observe co-deviations.