SPATULA: Anchored Bayesian Factorization for Distilling Trait-Specific Genetic Structure in Correlated Phenotypes

Pleiotropy induces pervasive genetic correlation across traits, often producing composite signals that obscure phenotype-specific biological mechanisms and complicate downstream analyses. To address this challenge, I introduce SPATULA (Specific Phenotype Architecture ThroUgh Latent Anchors), a Bayesian matrix factorization framework that anchors one latent factor toward a user-specified focal phenotype using a structured prior governed by a specificity parameter, enabling probabilistic quantification of trait-specific versus shared structure within a correlated phenome. Simulation studies demonstrate recovery of trait-specific loading directions and reasonable inference for the specificity parameter, particularly when the trait-specific component is not aligned with a dominant variance direction, where principal component analysis is less effective. Applications to UK Biobank phenotypes illustrate how anchored latent modeling can refine genetic interpretation. For hormone replacement therapy (HRT)—a treatment used to manage menopausal symptoms—SPATULA produces a trait-specific factor that preserves established reproductive aging loci. In the downstream analyses, association patterns are reweighted, with attenuation of education attainment signals and relative strengthening of reproductive timing-related signals. Integration with heterogeneity testing further suggests partially distinct genetic components underlying overall liability to use HRT from timing of initiation. In a complementary application to major depressive disorder (MDD), a clinically heterogeneous condition often measured using shallow self-reported proxies, SPATULA reduces pleiotropic overlap with correlated psychosocial constructs and isolates a more interpretable depression-aligned genetic axis. Together, these results establish SPATULA as an interpretable probabilistic framework for isolating trait-specific genetic architecture, refining mixture phenotypes, and enhancing downstream heterogeneity detection in large-scale biobank data.