CARLIS: covariate-assisted replicability analysis for genome-wide association studies via triplet hidden Markov models

Abstract Replicability analysis is a cornerstone for identifying genuine genetic associations in genome-wide association studies (GWAS), yet existing methods are constrained by their failure to account for linkage disequilibrium (LD) structure among single nucleotide polymorphisms (SNPs) or underuse of auxiliary information, limiting their reliability and statistical power. We develop CARLIS, a comprehensive covariate-assisted replicability analysis method to enhance both statistical rigor and biological interpretability while maintaining asymptotic false discovery rate control. CARLIS innovatively leverages a triplet hidden Markov model (HMM) to jointly characterize heterogeneous LD structures across two primary studies and an integrated auxiliary covariate (obtained via the Cauchy combination rule). The derived CARLIS statistic enables more efficient ranking of replicable SNPs by synthesizing cross-study and cross-SNP information through forward and backward probabilities. Computational scalability to genome-wide analyses is achieved through semiparametric estimation of composite null proportions and heterogeneous non-null density functions embedded in the HMM forward-backward algorithm. Extensive simulations demonstrate that CARLIS outperforms competing methods in statistical power while maintaining asymptotic false discovery rate control. Applications to replicability analysis of Parkinson’s disease GWAS and pleiotropy analysis of bipolar disorder and schizophrenia GWAS show that CARLIS identifies more biologically relevant replicable variants, highlighting its potential to accelerate functional genomics discovery and advance precision medicine.