AHA Scientific Sessions 2026 Abstract · Nov 2, 2026
From Proteins to Pathways: A Biologically Informed Neural Network for Risk Factors and Incident Coronary Artery Disease
By Tan Y, Li L, Liu Y et al.
Background: Coronary artery disease (CAD) remains the leading cause of death. Understanding the biological pathways that connect risk factors to disease onset may reveal mechanisms amenable to targeted prevention. Plasma proteomics offers a systems-level view of these pathways, yet conventional analyses test proteins individually, missing pathway architecture and protein-protein interactions; standard neural networks improve prediction but sacrifice interpretability. We developed a biologically informed neural network (BINN) wired to the Reactome pathway hierarchy so that every prediction is traceable to specific proteins and pathways by construction. Hypothesis: A pathway-constrained neural network applied to the plasma proteome can model both cross-sectional cardiometabolic risk factors and incident CAD with competitive discrimination while revealing interpretable, shared biological pathways between risk factors and disease. Methods: We built a BINN with four pathway-mapped layers, with increasing levels of abstraction, to plasma proteomics data from UK Biobank participants (N=39,455). Cross-sectional models predicted obesity and hyperlipidemia; a Cox proportional-hazards output modeled time to incident CAD. Protein- and pathway-level contributions were quantified by Shapley additive explanations. Results: Among 39,455 participants (55% female; median follow-up 13.5 years; 3,239 incident CAD cases), the BINN achieved an AUROC of 0.93 for obesity (vs 0.91, elastic net) and 0.87 for hyperlipidemia. Top-attributed obesity proteins (leptin, FABP4) were mapped to leptin signaling, white-adipocyte differentiation, and GLP-1 synthesis; top hyperlipidemia proteins (PLA2G7, APOA1, PCSK9) were partitioned to lipoprotein clearance, transport of small molecules, and metabolism (Fig 1). For incident CAD, the BINN achieved a Harrell C-index of 0.75: ; the leading proteins reflected established atherosclerosis biology (fibrinogen, von Willebrand factor, fibronectin), with hemostasis and extracellular-matrix organization as top pathways (Fig 2). Pathway attributions overlapped between risk factors and incident CAD. For example, hyperlipidemia and incident CAD shared 7 of their top 12 root Reactome pathways, such as hemostasis, extracellular-matrix organization, and metabolism (Fig 3), providing a pathway-level mechanistic link between a modifiable risk factor and incident disease. Conclusions: A single interpretable network predicted both cardiometabolic risk factors and incident CAD from the plasma proteome, tracing each prediction to specific proteins and biological pathways. Convergent pathway attributions between risk factor and CAD suggest shared mechanistic underpinnings that may inform targeted prevention strategies.