Assessing Ancient DNA Sampling Strategies for Natural Selection Inference in Humans Using Allele Frequency Time Series Data.

The increased availability of genomic data from ancient humans allows estimating the strength of natural selection at a given locus using time series data. Several methods have been developed for this purpose and were originally validated through simulations using mostly large sample sizes. However, human ancient DNA (aDNA) data typically have high missingness and include only a small uneven number of individuals per time point, making estimations of selection challenging. Here, we benchmark the inference of selection with aDNA-like time series datasets using extensive simulations and four methods: ApproxWF, BMWS, Slattice, and Sr. We test several sampling schemes of time series data and selection coefficients (s) and focus on whether one can infer selection using time series datasets with sparse data and small sample sizes. We show that detecting selection with aDNA data is possible for strong s with a sample size of ~100 when assuming constant Ne. While ApproxWF performs best across simulations, the other methods present more variable results and do not perform well for typical aDNA datasets. Importantly, generally low false positive rates (<6%) highlight low risks to falsely detect selection when the loci are evolving neutrally. Moreover, relatively high power (>90%) for s ≥ 0.02 (Nes ≥ 200) shows that strong selection can generally be detected with confidence. We also show that more homogenous sampling improves the accuracy of the estimations. Finally, we provide recommendations for future research aiming to estimate selection with aDNA, noting the importance of spreading data evenly across time and avoiding time points with extremely small sample sizes.