The Signal of Admixture Can Decay Rapidly When Using Clustering-Based Methods

Abstract

Gene flow shapes evolutionary trajectories by introducing novel alleles, facilitating or retarding adaptation or eroding divergence among populations. Studies commonly infer gene flow through estimates of genetic admixture from methods that cluster individuals by allele-frequency similarity. However, the ability of these methods to reliably detect historical admixture remains poorly understood, particularly under spatially restricted dispersal and across a range of migration rates. Here, we evaluate how signals of admixture arise and decay during and after gene flow across a range of within- and between-population dispersal rates, using three common inference methods: ADMIXTURE, sNMF and PopCluster. Using forward-time simulations in a two-dimensional landscape, we modelled two metapopulations that diverged in isolation, experienced a pulse of secondary contact and subsequently returned to isolation. All three methods systematically underestimated admixture, with the bias increasing once gene flow ceased: inferred admixture averaged 56% of the true value at the end of gene flow, declined to 23% after 1000 generations and fell below 5% by 2000 generations, even where true admixture exceeded 20%. This underestimation was accompanied by corresponding overestimation of barrier strength, assessed using a metric analogous to cline width. Both patterns reflect the erosion of allele-frequency differences upon which these methods depend, which coincides with the rapid re-establishment of Hardy–Weinberg and linkage equilibrium following secondary contact. Our simulations show that, at least in some scenarios, the clustering-based admixture methods evaluated here cannot recover the signal of historical gene flow from present-day samples.