Failure to replicate a genetic signal for sex bias in the steppe migration into central Europe.

Goldberg et al. (1) used genome-wide ancient DNA data (2) from central European Bronze Age (BA) populations and their three ancestral sources of steppe pastoralists (SP), Anatolian farmers (AF), and European hunter-gatherers (HG) to investigate whether the SP migration into central Europe after 5,000 years ago (3, 4) was sex-biased. By estimating a lower proportion of SP ancestry on the X chromosome (36.6%), which is primarily carried in females, than on the autosomes (61.8%), they suggest that the migration involved a ratio of 5–14 SP males for every female. We attempted to replicate this finding using qpAdm (3), which leverages allele frequency correlations between the admixed (BA) and source (SP, AF, and HG) populations with distant outgroups to eliminate potential biases due to genetic drift between the true source populations and the ones used as surrogates for them. Our outgroups are Mota (5), Ust_Ishim (6), Kostenki14 (7), GoyetQ116-1 (7), Vestonice16 (7), MA1 (8), AfontovaGora3 (7), and Levantine Neolithic farmers (9). We ran qpAdm with allsnps: YES and Mota as the basis population. The model fits the data (P = 0.072 autosomes and P = 0.747 chromosome X). For the BA population, we estimate 61.4 ± 2.9% SP, 31.0 ± 1.2% AF, and 7.6± 2.9%HG ancestry using all autosomal SNPs and 67.5 ± 17% SP, 26.5 ± 6.9% AF, and 6.0 ± 16.4% HG using all X-chromosome SNPs; thus, we do not find less SP ancestry on the X chromosome. To diagnose why we do not replicate the finding of ref. 1 we simulated pseudo-BA individuals with known admixture proportions and then used supervised ADMIXTURE (10)—the method used in ref. 1 on the same 4,605 X-chromosome SNPs. We started our simulations using all possible combinations of (20 AF) × (9 HG) × (8 SP) individuals. For each of these 1,440 triples we simulated 16 individuals, randomly sampling alleles from the source individuals according to the qpAdmestimated proportions of the 16 real BA individuals and removing simulated individuals with fewer than 1,000 SNPs as in ref. 1 (results were little changed with no filtering or a 2,000-SNP threshold). We estimated ancestry for each individual by performing 10 random-seeded supervised ADMIXTURE replicates and averaging the results (we used the remaining 19 AF, 8 HG, and 7 SP individuals as source populations) (Fig. 1A). Supervised ADMIXTURE predicts real ancestry poorly in this setting (Fig. 1B). The estimation error (estimated – real ancestry) is strongly correlated (r = 0.91) with the estimated SP ancestry, allowing us to predict it by a regression (Fig. 1C), which indicates upward bias for high SP ancestry estimates and downward bias for low ones. For the pool of 16 individuals that the authors of ref. 1 estimated had 36.6% ancestry the error is predicted to be −19.5%, largely eliminating the discrepancy between theX-chromosomeestimate of ref. 1 and the autosomal ancestry estimates of ADMIXTURE (1) and qpAdm. These results show that bias in the estimation of admixture proportions, rather than sex bias in the steppe migration, drives the findings of ref. 1.