LKOM: Longitudinal Knockoff with Masked Likelihood Ratio for FDR-Controlled Association Discovery

Abstract

Repeated phenotypic measurements in biobanks can reveal genetic effects on both average trait levels and within-subject variability. However, existing longitudinal genome-wide association methods generally test variants marginally and analyze the two components separately, limiting the localization of signals in linkage disequilibrium and the detection of variants with shared effects. Here we introduce Longitudinal KnockOff with Masked likelihood ratio (LKOM), a multivariable framework that jointly models mean and variability signals while controlling the variant-level false discovery rate for the joint null of no association with either component. LKOM uses model-X knockoffs and masked likelihood-ratio statistics to distinguish original variants from synthetic controls, and assigns selected variants to mean-specific, variability-specific or shared association states. A genetic-relationship-matrix-based whitening extension accommodates related individuals. In simulations, LKOM maintained false discovery rate control across diverse genetic architectures, heavy-tailed errors, reduced longitudinal follow-up and sample sizes up to 100,000. It generally improved power over separate-channel knockoff and marginal association methods, with the largest gains for shared effects and rare variants, particularly when the overall genetic contribution to phenotypic variation was modest. In related samples, whitening restored calibration under dense pedigree structures while preserving power gains over SPAGRM(CCT). LKOM provides an interpretable and scalable approach for identifying conditionally associated variants from longitudinal phenotypes.