MagicalRsq

MagicalRsq calibrates genotype imputation quality metrics using a machine-learning approach to improve accuracy for lower-frequency genetic variants.

Key Features:

• Machine-learning calibration: Integrates a machine-learning model to recalibrate post-imputation Rsq values.
• Variant-level inputs: Uses variant-level imputation data and population genetics statistics as model inputs.
• Focus on lower-frequency variants: Specifically targets improved calibration for rare and low-frequency variants as well as common variants.
• Validation datasets: Trained and evaluated using whole-genome sequencing (WGS) data from the Cystic Fibrosis Genome Project (CFGP) and whole-exome sequence data from UK Biobank (UKB).
• Cross-ancestry evaluation: Performance validated across European and African ancestry samples.
• Application to TOPMed-imputed array data: Example models trained on 1,992 CFGP sequenced samples and applied to an independent set of 3,103 samples with TOPMed imputation from array genotypes.
• Empirical gains: Demonstrated net identification gains of approximately 1.4 million rare variants, 117,000 low-frequency variants, and 18,000 common variants relative to standard Rsq.
• Post-imputation quality metric: Produces a better-calibrated Rsq to distinguish well-imputed variants from poorly imputed variants.

Scientific Applications:

• Imputation quality assessment: Calibration of Rsq values for more accurate assessment of imputation confidence in genomic studies.
• Variant filtering for association studies: Improved discrimination between well- and poorly-imputed variants to inform downstream association analyses.
• Recovery of rare and low-frequency variants: Enhance detection and inclusion of rare and low-frequency variants in datasets imputed with TOPMed from array genotypes.
• Cross-ancestry analyses: Support for calibration and evaluation of imputation quality across European and African ancestry cohorts.

Methodology:

Apply a machine-learning model trained on variant-level imputation data and population genetics statistics using WGS from CFGP and whole-exome data from UKB, then use the trained model to recalibrate post-imputation Rsq values on TOPMed-imputed array genotype datasets (example: training on 1,992 CFGP samples and applying to 3,103 samples).