Project Summary The ability to accurately predict the effect of genetic variation on phenotypes at multiple scales would radically transform our ability to apply genomic technologies in order to understand human health and disease. This predictive ability would significantly improve the effectiveness of a broad spectrum of genomic analyses ranging from genome-wide association studies for common diseases to diagnostic odysseys searching for genetic causes of rare diseases. To address this challenge, we propose to develop a trainable approach for predicting the phenotypic impact of genetic variants. This approach will support predictions for a broad range of genetic variations, phenotypes, and biological contexts. It will incorporate and exploit mechanistic knowledge of pathways where available, but augment this pathway knowledge with learned models where it is not. This approach will consist of a synthesis of (i) methods that link genomic variants to their effect on expression or function of individual gene products, (ii) methods that link those relationships into the subnetworks involved in cellular responses of interest, (iii) machine-learning approaches that infer models pertaining to a variety of genotype-phenotype relations from large training sets. We will also develop and apply active learning algorithms to identify the most informative experiments for subsequent analysis by IGVF Consortium. Additionally, we will develop and apply a statistical framework for elucidating genetic modifiers, through probabilistic, network-informed inference of common variants identified in GWAS that modify the impact of rare variants implicated in sequencing-based association studies. Throughout the project, we will work closely with other IGVF Centers to guide experimental data collection, benchmark methods from across Centers, and contribute to the variant-element-phenotype catalog which will have broad applications by the community.