Exploring imputation performance of phenotype-associated SNPs for forensic prediction models

Abstract Imputing single nucleotide polymorphisms (SNPs) to expand genotyped SNP panels is commonly applied to study genotype-phenotype correlations in medical and population genetics but is largely unexplored in forensics. Forensic DNA phenotyping, i.e., the SNP-based phenotype prediction, can greatly benefit from imputing missing prediction model SNPs. Unfortunately, most imputation studies investigate the performance of random SNPs with limited focus on phenotype-associated SNPs. Here, we compare the imputation accuracy of SNPs leveraged in phenotype prediction models, phenotype-associated SNPs and all SNPs and explore the performance of phenotype prediction model, HIrisPlex-S, using imputed datasets. We confirm that the number and selection of SNPs in the genotype dataset and minor allele frequencies (MAFs) are major drivers of imputation call and error rates. Phenotypic SNPs show higher imputation error rates than randomly selected SNPs due to MAF differences. Further, we provide novel insights into accurate phenotype prediction for dense SNP panels of high linkage, frequent phenotypes in the reference panel, and lenient imputation thresholds due to the detrimental effect of missing genotypes on trait prediction compared to imputation errors. Overall, we show the importance of imputation performance studies for SNPs of interest and the high applicability of SNP imputations prior to phenotype prediction.