CluStrat

CluStrat performs structure-informed clustering to correct population stratification in genome-wide association studies (GWAS) by leveraging linkage disequilibrium (LD)-induced distances between individuals to improve identification of causal variants.

Key Features:

• Agglomerative Hierarchical Clustering: Employs agglomerative hierarchical clustering using the Mahalanobis distance computed from a Genetic Relationship Matrix (GRM) that captures LD-induced covariance among genetic markers.
• Sketching-Based Randomized Ridge Regression: Integrates sketching-based randomized ridge regression on genotype data to compute association statistics while reducing computational overhead for large datasets.
• Genetic Relationship Matrix (GRM) Computation: Computes the GRM via a connection between statistical leverage scores and Mahalanobis distance, avoiding explicit computation and storage of large genome-wide covariance matrices.

Scientific Applications:

• Simulation Validation: Validated on simulations of discrete and admixed sub-populations, identifying two to three times more true causal variants versus Principal Component (PC)-based stratification with a modest increase in spurious associations.
• WTCCC2 Parkinson’s Disease Analysis: Applied to the WTCCC2 Parkinson’s disease dataset, identifying loci associated with BACH2, MAP2, NR4A2, SLC11A1, and UNC5C.

Methodology:

Uses agglomerative hierarchical clustering with Mahalanobis distance derived from a GRM; computes the GRM using statistical leverage scores linked to Mahalanobis distance; and applies sketching-based randomized ridge regression on genotype data to compute association statistics.