From Gene Networks to Therapeutics: A Causal Inference and Deep Learning Approach for Drug Discovery

Background/Objectives: Drug discovery is a lengthy and expensive process, taking an average of 10 years and more than $2 billion from target discovery to drug approval. It is even more challenging in complex diseases due to the disease heterogeneity and limited knowledge about the underlying mechanisms. We present a novel computational framework that integrates network analysis, statistical mediation, and deep learning to identify causal target genes and repurposable small-molecule candidates. Methods: We applied weighted gene co-expression network analysis (WGCNA) and bidirectional mediation analysis (causal WGCNA) to transcriptomic data from idiopathic pulmonary fibrosis (IPF) patients to identify genes causally linked to the disease phenotype. These genes were used as a phenotypic signature for deep learning-based compound screening using the DeepCE model. Results: Using RNA-seq data from 103 IPF patients and 103 controls, we identified seven significantly correlated modules and 145 causal genes. Five of these genes (ITM2C, PRTFDC1, CRABP2, CPNE7, and NMNAT2) were predictive of disease severity in IPF. Our compound screening identified several promising candidates, such as Telaglenastat (GLS1 inhibitor), Merestinib (MET kinase inhibitor), and Cilostazol (PDE3 inhibitor), with significant inverse correlation with the IPF-specific gene signature. Conclusions: This study demonstrates the utility of combining causal inference and deep learning for drug discovery. Our framework identified novel gene targets and therapeutic candidates for IPF, offering a scalable strategy for phenotype-driven drug discovery and repurposing.