Accurate and flexible estimation of effective population size history

Current methods for inferring historical population sizes from DNA sequences often impose a heavy computational burden, or relieve that burden by imposing a fixed parametric form. In addition, they can be marred by sequencing errors or uncertainty about recombination rates, and the quality of inference is often poor in the recent past. We propose “InferNo” for flexible, nonparametric inference of effective population sizes. It requires modest computing resources and little prior knowledge of the recombination and mutation maps, and it is robust to sequencing error and gene conversion. We illustrate the statistical and computational advantages of InferNo over previous approaches using a range of simulation scenarios. In particular, we demonstrate the ability of InferNo to exploit biobank-scale datasets for accurate inference of rapid population size changes in the recent past. Applying InferNo to worldwide human data, we find remarkable similarities in inferences from different populations in the same region. The historical sizes of all non-African populations converge around 40 000 years ago, but they remain distinct from those of African populations for over 800 000 years, long predating estimates for the migration of modern humans out of Africa.