Empirical versus estimated accuracy of imputation: optimising filtering thresholds for sequence imputation

Area of Science:

• Genomics
• Bioinformatics
• Population Genetics

Background:

• Genotype imputation is vital for cost-effective genomic analyses like genome-wide association studies (GWAS).
• Inaccurate imputation can lead to false positives, necessitating data pre-filtering or accuracy assessment.
• This study benchmarks imputation software and evaluates imputation accuracy in cattle.

Purpose of the Study:

• To benchmark Beagle 5.2, Minimac4, and IMPUTE5 for genotype imputation accuracy.
• To compare empirical imputation accuracy with software-estimated accuracy (Rsq_soft).
• To assess imputation accuracy across different cattle chromosomes (autosomal, X), variant types (SNP, INDEL), and genotype densities (low, high).

Main Methods:

• Benchmarking three imputation programs: Beagle 5.2, Minimac4, and IMPUTE5.
• Comparing empirical imputation accuracy against software-estimated Rsq_soft values.
• Evaluating imputation performance for SNPs and INDELs on autosomal and X chromosomes, using low-density and high-density genotypes.

Main Results:

• Higher imputation accuracy was achieved using high-density genotypes compared to low-density.
• All tested software performed well, with minor accuracy differences.
• Empirical imputation accuracy closely correlated with Rsq_soft, but differed between Minimac4 and Beagle 5.2/IMPUTE5.
• Customizing Rsq_soft thresholds per software is essential for merging data and for regions with poor imputation accuracy (e.g., segmental duplications).
• Indel imputation accuracy was ~6% lower than SNP imputation accuracy.
• Non-PAR X chromosome imputation accuracy was comparable to autosomal, but PAR accuracy was substantially lower, especially with low-density genotypes.

Conclusions:

• An empirically derived approach for applying customized, software-specific Rsq_soft thresholds is proposed for downstream analyses like meta-GWAS.
• The Pseudo-Autosomal Region (PAR) on the X chromosome requires high-density genotypes for accurate imputation, particularly when starting from low-density data.