Improving Genotype Imputation in High-Dimensional Pharmacogenomics Using Multiple Imputation: Evaluation with Machine...

Area of Science:

• Genetics
• Bioinformatics
• Statistical Genetics

Background:

• Handling missing data in high-dimensional genetic datasets is challenging.
• Traditional imputation methods often fall short in complex genetic analyses.
• Multiple imputation (MI) is established but underutilized in this context.

Purpose of the Study:

• To compare machine learning (ML) and traditional methods for imputation and covariate selection in high-dimensional genetic data.
• To develop and evaluate an MI framework incorporating genotype probabilities and imputation uncertainty.
• To assess the performance of MI in recovering pharmacogenomic associations and improving discovery.

Main Methods:

• Developed a novel multiple imputation framework using genotype probabilities, INFO scores, and missingness percentages.
• Employed dimensionality reduction with random forest and penalized regression for scalable covariate selection.
• Validated methods using pharmacokinetic tuberculosis simulations, SNP data from the 1000 Genomes Project, and clinical warfarin datasets (War-PATH, IWPC, UK Biobank).

Main Results:

• Multiple imputation significantly improved confidence interval coverage (up to 94%) compared to single imputation (0%) in simulations.
• MI successfully recovered known pharmacogenomic associations (e.g., CYP2C9, VKORC1) and identified novel signals (e.g., rs4697699) in clinical datasets.
• Penalized regression excelled in high-effect SNP selection (F1=0.897), while GWAS + random forest performed better in low-effect scenarios (F1=0.657).

Conclusions:

• Multiple imputation enhances reliability and discovery power in high-dimensional pharmacogenomic studies.
• ML methods show potential for SNP selection but require further investigation for consistent benefits.
• Scalability to biobank-scale analyses and generalizability remain key areas for future research.