MaCH-Admix

MaCH-Admix improves genotype imputation accuracy in admixed human populations by constructing individual-specific reference panels and evaluating IBS-based and ancestry-weighted reference selection within a hidden Markov model framework to address complex linkage disequilibrium patterns.

Key Features:

• Piecewise Reference Selection Method: Constructs tailored reference panels specific to each target individual to improve imputation quality in admixed populations.
• Separation of Model Parameter Estimation and Imputation: Decouples parameter estimation from the imputation step, allowing imputation with standard reference panels combined with pre-calibrated parameters.
• Identity-by-State (IBS)-Based and Ancestry-Weighted Approaches: Evaluates both IBS-based and ancestry-weighted strategies for constructing effective individual-specific reference panels.
• Hidden Markov Model Framework: Implements imputation and reference-panel evaluation within a hidden Markov model framework.
• Extension of MaCH 1.0: Builds upon the algorithms and capabilities of MaCH 1.0.
• Performance Improvement: Demonstrates up to a 5.1% information gain over BEAGLE and IMPUTE2 with statistical significance by a Wilcoxon signed rank test (P-value < 0.0001), with particular gains for uncommon variants.

Scientific Applications:

• Large-scale genomic studies of admixed populations: Supports genotype imputation in cohorts with complex LD, such as admixed African American and Hispanic American populations.
• Empirical validation and benchmarking: Validated on Women's Health Initiative datasets comprising 8,421 African Americans and 3,587 Hispanic Americans to assess imputation quality, especially for uncommon variants.

Methodology:

Evaluates various reference panel construction methods within a hidden Markov model framework, compares large, medium, and small reference panels across genome regions with varying linkage disequilibrium, employs a piecewise IBS method for individual-specific panel construction, and separates model parameter estimation from imputation.