Graph Attention-Based Feature Selection for Multi-Omics Drug Target Prediction in Cardiovascular Diseases

Abstract

Drug target identification in cardiovascular diseases faces computational challenges due to the high dimensionality of multi-omics datasets. Here, we present a graph attention framework that integrates genomic variants, proteomic expression profiles, and metabolomic signatures using hierarchical attention mechanisms applied to molecular interaction networks. In these networks, nodes represent molecular entities and edges capture experimentally validated functional associations, thereby encoding key biological relationships. The attention mechanism assigns adaptive importance weights αᵢⱼ to neighboring nodes, facilitating selective feature propagation while maintaining the integrity of biological signals. Validation across three cardiovascular cohorts-encompassing 12,226 patients with whole-genome sequencing, proteomics, and metabolomics data-achieves 87.3% target identification accuracy alongside a 72.0% reduction in feature dimensionality. Analysis of attention weights highlights differential pathway contributions, with MAPK signaling (0.342), calcium homeostasis (0.298), and PI3K-AKT cascades (0.276) identified as principal therapeutic nodes. The framework successfully recovers 23 FDA-approved cardiovascular drugs and predicts 17 investigational compounds currently in clinical trials. Computational complexity decreases from O(n²d) to O(nkd), where k denotes the selected features (k << n), resulting in a 4.2-fold speedup in execution. Gradient-based attribution methods further provide mechanistic interpretability, linking molecular features to pathway-level biological processes. This approach bridges the computational and biological gap in precision cardiovascular medicine by offering mathematically grounded feature selection with preserved mechanistic transparency.