Using Deep Learning Models of Gene Regulation to Guide Drug Prioritization.

Background: Drug repurposing offers a cost-effective strategy to accelerate therapeutic discovery, but most computational approaches do not model noncoding genetic variation. Because over 90% of genome-wide association study (GWAS) risk variants reside in noncoding regions, linking regulatory variation to therapeutic hypotheses remains a major challenge. Methods: We developed an integrative deep learning framework that links allele-specific enhancer prediction to candidate therapeutics through two complementary prioritization strategies, a transcription factor (TF)-based and a gene-based approach. We used MCF7-breast cancer context as a proof-of-concept system. Results: GWAS heritability was significantly enriched in MCF7 enhancers. Allele-specific variant scoring identified 1537 breast cancer risk variants with strong predicted regulatory effects, and attribution-based motif discovery revealed enrichment of FOXA1-associated motif features, consistent with FOXA1 upregulation in primary tumors. TF-based prioritization, integrating FOXA1 knockdown-induced and drug-induced gene expression profiles, identified 63 candidate compounds, including 18 approved drugs, and recovered fulvestrant, an established breast cancer therapy. Gene-based prioritization, mapping candidate regulatory variants to 347 target genes, identified 140 candidate compounds, including approved breast cancer drugs toremifene and raloxifene. Both strategies identified compounds with anti-correlated transcriptional signatures across core breast cancer hallmark pathways, and integration of pathway anti-correlation, drug-gene interactions, and supporting experimental or clinical evidence yielded 15 high-confidence repurposing candidates. Conclusions: Recovery of approved breast cancer therapeutics supports the biological relevance of deep learning-predicted regulatory variants. This study establishes a regulatory variant-guided drug repurposing framework that connects noncoding genetic variation to candidate therapeutics and provides a scalable strategy for generating pharmacologically relevant hypotheses from the noncoding genome.