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

Area of Science:

• Genomics
• Computational Biology
• Pharmacology

Background:

• Most drug repurposing computational methods overlook noncoding genetic variations, a significant challenge as over 90% of genome-wide association study (GWAS) risk variants are noncoding.
• Linking regulatory variations to therapeutic hypotheses is crucial for effective drug discovery.

Purpose of the Study:

• To develop an integrative deep learning framework for drug repurposing that models noncoding genetic variation.
• To connect allele-specific enhancer predictions with transcription factor (TF)-centered gene expression changes and drug-induced transcriptional profiles.
• To prioritize candidate therapeutics by integrating regulatory variant data with drug response profiles.

Main Methods:

• Developed cell type-specific deep learning models for enhancer prediction.
• Performed allele-specific variant scoring and attribution-based motif discovery.
• Integrated TF knockdown-induced and drug-induced gene expression profiles.
• Utilized drug-gene interaction data for further refinement.

Main Results:

• The framework accurately distinguished active enhancers across seven cell lines.
• GWAS heritability was significantly enriched in MCF7 enhancers, identifying a relevant cellular context for breast cancer.
• Identified 63 candidate compounds for breast cancer treatment, including 18 approved drugs like fulvestrant.
• Prioritized compounds showed significantly higher anti-correlated transcriptional effects across core breast cancer pathways compared to non-prioritized ones.
• Refined candidates to eight compounds with supporting experimental or clinical evidence.

Conclusions:

• Established a regulatory variant-guided drug repurposing framework connecting noncoding genetic variation to therapeutic candidates.
• Demonstrated a generalizable strategy for translating noncoding genome insights into pharmacologically relevant hypotheses.
• Highlighted the potential of deep learning in advancing drug discovery through noncoding genetic variation analysis.