A genetically informed atlas of transcriptomic programs from massive human RNA-seq compendia

Abstract

Genetic risk for complex disease acts through coordinated gene programs in specic tissues and cellular contexts, which single-gene analyses cannot resolve. Interpretable gene modules can capture these programs, but prior-informed methods have not scaled to large expression compendia, andgenetic associations have not been integrated with modules learned at this scale. Here we present CLAMP, a fast, memory-ecient prior-informed framework that learns interpretable gene modules from such compendia and integrates them with genetic associations and drug signatures. CLAMP recovered cell-type composition and tissue organization more accurately than existing decomposition methods. Training on 605,614 ARCHS4 samples increased pathway and genetic-trait recovery, and projection into these modules recovered programs missed by study-specic models. ARCHS4 modules also ranked known drug-disease pairs better than modules learned from smaller compendia or individual genes. In reproducible consensus modules, cell-type, pathway, drug-response and genetic evidence converged on shared processes, placing CRISPR-validated lipid regulators missed by singlegene tests in disease-associated programs. CLAMP is available through Bioconductor, and its trained models provide a genetically informed atlas of gene programs for interpreting complex disease.