Kexin Huang

Understanding how disease-associated variants impact their target genes in cell-type-specific programs enables disease target discovery and effective therapies. However, building a comprehensive map of disease-critical variant-to-gene-to-program links requires testing a vast number of genes and variants across cell types using expensive experimental tools (e.g. CRISPRi-FlowFISH, Perturb-Seq) and is thus highly unscalable. I propose a novel class of graph neural networks that can predict variants, their target genes, and the cellular programs they disrupt. This framework will significantly reduce the required number of experiments and enable the generation of variant-to-gene-to-program maps for all diseases across all possible cellular contexts.