Performance of qpAdm -based screens for genetic admixture on admixture-graph-shaped histories and stepping-stone landscapes.

Abstract

qpAdm is a statistical tool that is often used in exploratory archaeogenetic studies for finding optimal admixture models of population history. Despite its popularity, qpAdm remains untested on histories in the form of admixture graphs of random topology or stepping-stone landscapes. We analyzed data from such simulations and found that while for admixture-graph-shaped histories there exist simple solutions (temporal stratification) for minimizing false findings of gene flow, in the case of stepping-stone landscapes the method generates results that do not appear suspect but are misleading: feasible qpAdm models are either accurate but simplistic in the context of landscapes, or highly inaccurate in the case of multi-component models. This is largely is due to two reasons: 1) because of complex migration networks that violate the assumptions of the method, there is poor correlation between qpAdm p -values and model optimality in many sections of the parameter space; 2) admixture fraction estimates between 0 and 1 are largely restricted to symmetric source configurations around targets, hence popular [0, 1] model plausibility criteria confound analyses of landscape-type demographies, unless their interpretations are explicitly spatial. For many species/regions/periods archaeogenetic sampling is very sparse and may be random with respect to population density of ancient individuals. In this situation only a specific combination of landscape properties and feasibility criteria allows to efficiently reject highly asymmetric non-optimal models most abundant in random deme sets. This problem may obscure useful signal (rare optimal models) and might be responsible for some claims about rapid long-distance migrations in the literature.