Performance of <i>qpAdm</i> -based screens for genetic admixture on admixture-graph-shaped histories and stepping-stone landscapes

Abstract

<i>qpAdm</i> is a statistical tool that is often used for testing large sets of alternative admixture models for a target population. Despite its popularity, <i>qpAdm</i> remains untested on two-dimensional stepping-stone landscapes and in situations with low pre-study odds (low ratio of true to false models). We tested high-throughput <i>qpAdm</i> protocols with typical properties such as number of source combinations per target, model complexity, model feasibility criteria, etc. Those protocols were applied to admixture-graph-shaped and stepping-stone simulated histories sampled randomly or systematically. We demonstrate that false discovery rates of high-throughput <i>qpAdm</i> protocols exceed 50% for many parameter combinations since: 1) pre-study odds are low and fall rapidly with increasing model complexity; 2) complex migration networks violate the assumptions of the method, hence there is poor correlation between <i>qpAdm p</i>-values and model optimality, contributing to low but non-zero false positive rate and low power; 3) although admixture fraction estimates between 0 and 1 are largely restricted to symmetric configurations of sources around a target, a small fraction of asymmetric highly non-optimal models have estimates in the same interval, contributing to the false positive rate. We also re-interpret large sets of <i>qpAdm</i> models from two studies in terms of source-target distance and symmetry and suggest improvements to <i>qpAdm</i> protocols: 1) temporal stratification of targets and proxy sources in the case of admixture-graph-shaped histories; 2) focused exploration of few models for increasing pre-study odds; 3) dense landscape sampling for increasing power and stringent conditions on estimated admixture fractions for decreasing the false positive rate.