Guidelines on combining genomic data across populations

Abstract

Genomic selection has been widely implemented across several livestock species. Due to reduced genotyping costs, genomic data are being generated at increasing rates. Various genotyping arrays are available with single-nucleotide polymorphisms (SNPs) genotyping densities ranging from few hundred to hundreds of thousands of single-nucleotide polymorphisms (SNPs). Moreover, advancements in genome sequencing enable the generation of newer and more accurate reference genomes. Genotyping arrays are typically updated to remain close to newer reference genomes, thereby increasing the accuracy of genome-wide association studies and, potentially, genomic predictions. Thus, genomic data have been generated using different genotyping chips, i.e., arrays, and assemblies, creating the need to merge them into a unified map for downstream analyses. However, SNPs do not completely overlap across chips. The number of overlapping SNPs decreases even further when chips come from different manufacturers, from different organisations, which may use customised chips, or when the corresponding assembly versions differ. Moreover, even when chips refer to the same genome assembly, SNPs do not always point to the same physical positions, and metadata, such as SNP names, may differ. Therefore, merging genomic data across populations presents significant challenges which are further complicated by the discontinuation of public mapping resources. The aim of this study is to provide guidelines for combining SNP genotyping data collected across different genotyping chips and populations based on the experience of the M3GE project. More than 3.1 million individual genotypes from purebred and crossbred cattle were collected across six national breeding organisations from three countries: Ireland, Great Britain, and Italy. Genotypes were collected on 11 different genotyping chips with densities ranging from 30K to 777K SNPs. The corresponding assemblies were either UMD3.1 or ARS-UCD1.2. Genotypes were either in raw format or imputed to a custom set of SNPs as used in routine evaluations. Genotypes were collected through the Interbull Centre GenoEx-GDE database following the Illumina A/B coding and counting the number of copies of the A allele. All genomic data were converted to the UMD3.1 assembly, validated for inconsistencies or errors in allele coding, physical positions on the genome, and allele frequency distributions, both within and across breeds. For maps alignment, we advise storing and exchanging maps together with their metadata (rsID, assembly version, and manifest), so that SNP positions can be linked or converted across chips and assemblies. For genotypes alignment, we advise using a consistent A/B allele coding and using allele frequencies as a diagnostic tool, combining correlation and root mean squared error and comparing each chip with at least two other chips. These alignment steps, together with the required file formats, are needed to ensure valid downstream analyses of combined genomic data.