Maximum likelihood point estimates for improved population genetics statistics

Likelihood-based approaches that incorporate the uncertainty in basecalling have become a standard approach in genotyping. Likewise, the accuracy of population genetic estimators could be improved if these uncertainties could be taken into account. Moreover, the classic asymptotically unbiased estimators for population genetic parameters may yield impossible values with finite data, and introduce a median bias driven by sample size, which complicates the comparison of populations of different sizes. Here we develop an approach (https://github.com/anulin/PiThetic) that more accurately estimates the population genetic parameters and addresses the question of comparison (median) bias. Our approach improves by directly estimating population genetic parameters from basecall accuracies in terms of Phred scores, instead of using allele frequencies. We provide three improvements to classic statistical estimates: (i) we apply Phred scores to improve the accuracy of the estimates, (ii) we show how the pursuit for an unbiased estimator can lead to impossible and/or less accurate values, and provide a corrected estimator, and (iii) we provide a new estimator that minimizes the median bias for nucleotide diversity and another estimator that reduces error for D, because the standard Watterson {theta} and Tajimas D estimators have high error. Overall, this work facilitates the more accurate estimation of population genetics statistics from sequencing data.