Interpretable neural network for risk stratification and drug target discovery based on PBMC transcriptomes

Peripheral blood mononuclear cells (PBMCs) play a central role in immune surveillance and disease pathogenesis and provide a minimally invasive source for molecular screening. Here, we developed ME-NET, an intrinsically interpretable deep learning framework for multi-disease screening and risk stratification from PBMC transcriptomic profiles. Using harmonized datasets from heterogeneous clinical backgrounds, ME-NET showed consistent performance and outperformed standard machine-learning baselines. By incorporating curated pathway priors in a pathway-informed layer, ME-NET connects predictions to pathway-level signals and contributing genes, enabling biologically grounded interpretation of disease-associated immune programs. Finally, integrating ME-NET-prioritized targets with drug-gene interaction resources identified candidate compounds for potential therapeutic repurposing. Collectively, these results support the use of interpretable deep learning on blood-derived transcriptomes for scalable multi-disease screening and for prioritizing biomarkers and drug targets.