Using Deep Learning Models of Gene Regulation to Guide Drug Prioritization

Drug repurposing offers a cost-effective strategy to accelerate therapeutic discovery, but most computational approaches fail to 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. Here, we developed an integrative deep learning framework that links allele-specific enhancer prediction to transcription factor (TF)-centered gene expression changes and drug-induced transcriptional profiles to prioritize candidate therapeutics. Our cell type-specific deep learning enhancer models accurately distinguish active enhancers across seven cell lines. Using breast cancer as a proof-of-concept, we found that GWAS heritability is significantly enriched in MCF7 enhancers, supporting MCF7 as the cellular context for this disease. Allele-specific variant scoring identified breast cancer risk variants with strong allele-dependent effects, and attribution-based motif discovery revealed enrichment of FOXA1-associated motif features, consistent with FOXA1 upregulation in primary tumors. Integration of the FOXA1 knockdown-induced and drug-induced gene expression profiles identified 63 candidate compounds for treatment of breast cancer, including 18 approved drugs, with recovery of the known breast cancer therapy fulvestrant. Among prioritized compounds, 54% showed anti-correlated transcriptional effects across eight core breast cancer pathways, compared to 5.3% of non-prioritized compounds. Integration of drug-gene interaction data further refined these to eight compounds with supporting experimental or clinical evidence. Together, these results establish a regulatory variant-guided drug repurposing framework that connects noncoding genetic variation to therapeutic candidates and provides a generalizable strategy for translating the noncoding genome into pharmacologically relevant hypotheses.Drug repurposing offers a cost-effective strategy to accelerate therapeutic discovery, but most computational approaches fail to 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. Here, we developed an integrative deep learning framework that links allele-specific enhancer prediction to transcription factor (TF)-centered gene expression changes and drug-induced transcriptional profiles to prioritize candidate therapeutics. Our cell type-specific deep learning enhancer models accurately distinguish active enhancers across seven cell lines. Using breast cancer as a proof-of-concept, we found that GWAS heritability is significantly enriched in MCF7 enhancers, supporting MCF7 as the cellular context for this disease. Allele-specific variant scoring identified breast cancer risk variants with strong allele-dependent effects, and attribution-based motif discovery revealed enrichment of FOXA1-associated motif features, consistent with FOXA1 upregulation in primary tumors. Integration of the FOXA1 knockdown-induced and drug-induced gene expression profiles identified 63 candidate compounds for treatment of breast cancer, including 18 approved drugs, with recovery of the known breast cancer therapy fulvestrant. Among prioritized compounds, 54% showed anti-correlated transcriptional effects across eight core breast cancer pathways, compared to 5.3% of non-prioritized compounds. Integration of drug-gene interaction data further refined these to eight compounds with supporting experimental or clinical evidence. Together, these results establish a regulatory variant-guided drug repurposing framework that connects noncoding genetic variation to therapeutic candidates and provides a generalizable strategy for translating the noncoding genome into pharmacologically relevant hypotheses.