Beyond Univariate GWAS: Learning Cross-Disorder Representations of Genetic Effects

Genome-wide association studies (GWAS) typically analyze variant--trait associations independently, leaving the shared organization of psychiatric genetic effects largely implicit. We formulate psychiatric GWAS summary statistics as a multi-view representation learning problem in which variant--trait observations are embedded jointly from local association statistics and cross-disorder profiles. Under strict locus-disjoint and held-out-disorder protocols, learned embeddings outperform strong univariate baselines ($|z|$ and $-\log_{10} p$) in 86.4\% of 162 hyperparameter configurations. The learned space also exhibits stable hierarchical organization across a broader multi-trait structural panel, yielding 16 parent modules and 64 submodules with mean purity 0.94. A completed label-removal ablation, trained without trait classification or trait-supervised contrastive losses, retains stable module structure and stronger biological-coherence retrieval than PCA/raw baselines. Biological grounding and perturbation analysis identify a stable brain-related axis. These results suggest that representation learning provides a complementary, structure-aware view of psychiatric GWAS data.