Neural inference of population size history from noisy genotype data and recombination maps

Population size history can be estimated from genomes; methods from PSMC to phlash fit a coalescent model. Sequencing errors and recombination-rate variation are hard to write into such a model but easy to simulate, so we trained a neural estimator, NeCast, on simulations containing both. From unphased genotypes of 32 individuals in a 30-megabase window and its recombination map, it estimates sizes 50 to 500,000 generations ago. On error-free simulated windows, NeCast was about as accurate as phlash, more accurate than SMC++ and FitCoal, and took 7.4 s on a 16-core CPU, against their 5 min to 7.6 h. Sequencing errors raised its RMSE (log10 Ne) from 0.150 to 0.175, and between low-coverage and 30× data for the same individuals its estimates differed less than phlash's. Without a map it underestimated recent sizes; the map largely corrected this in simulations and raised recent estimates in all 26 1000 Genomes populations.