Efficiency of different strategies to mitigate ascertainment bias when using SNP panels in diversity studies

Background: Single nucleotide polymorphism (SNP) panels have been widely used to study genomic variations within and between populations. Methods of SNP discovery have been a matter of debate for their potential of introducing ascertainment bias, and genetic diversity results obtained from the SNP genotype data can be misleading. We used a total of 42 chicken populations where both individual genotyped array data and pool whole genome resequencing (WGS) data were available. We compared allele frequency distributions and genetic diversity measures (expected heterozygosity (He), fixation index (F-ST) values, genetic distances and principal components analysis (PCA)) between the two data types. With the array data, we applied different filtering options (SNPs polymorphic in samples of two Gallus gallus wild populations, linkage disequilibrium (LD) based pruning and minor allele frequency (MAF) filtering, and combinations thereof) to assess their potential to mitigate the ascertainment bias. Results: Rare SNPs were underrepresented in the array data. Array data consistently overestimated He compared to WGS data, however, with a similar ranking of the breeds, as demonstrated by Spearman's rank correlations ranging between 0.956 and 0.985. LD based pruning resulted in a reduced overestimation of He compared to the other filters and slightly improved the relationship with the WGS results. The raw array data and those with polymorphic SNPs in the wild samples underestimated pairwise FST values between breeds which had low F-ST (< 0.15) in the WGS, and overestimated this parameter for high WGS F-ST (> 0.15). LD based pruned data underestimated FST in a consistent manner. The genetic distance matrix from LD pruned data was more closely related to that of WGS than the other array versions. PCA was rather robust in all array versions, since the population structure on the PCA plot was generally well captured in comparison to the WGS data. Conclusions: Among the tested filtering strategies, LD based pruning was found to account for the effects of ascertainment bias in the relatively best way, producing results which are most comparable to those obtained from WGS data and therefore is recommended for practical use.