AI-Driven Biomedical Knowledge Hypergraphs for Interpretable Drug Repurposing and Precision Therapy

About This Grant

Project summary/Abstract The traditional drug development pipeline is critically inefficient, slow, costly, and has high failure rates. These diff iculties are especially pronounced for patients with rare and understudied diseases, who face a significant unmet medical need. Computational drug repurposing is a promising way to address these public health issues. Existing computational drug repurposing has two key limitations in current methodologies: (1) the restricted representation of biomedical knowledge as pairwise graphs, and (2) the lack of generalizable, interpretable predictive models. To address the first limitation, we build upon the NCATS Biomedical Data Translator’s foundational knowledge graph to construct the first large-scale biomedical knowledge hypergraph (Aim 1), that explicitly captures higher-order relationships that involve more than two entities—such as how specific drug combinations benefit particular disease subtypes or patient subgroups defined by demographics or genetics. This enriched representation is critical for enabling precision medicine and the discovery of synergistic therapies, which cannot be effectively captured using traditional pairwise-only graph structures. To extract these higher-order relationships from literature, we will develop a novel reinforcement learning-enhanced large language model (RL-LLM) to extract higher-order relationships directly from biomedical literature. A key feature of our framework is a human-in-the-loop quality control process, in which domain experts curate and validate extracted relationships to ensure accuracy and biomedical relevance. To address the second limitation, we will develop HyperGAT, a hypergraph neural network framework that incorporates heterogeneous attention mechanisms, interpretability modules, and zero-shot learning capabilities to enable accurate and explainable predictions of drug repurposing opportunities and synergistic therapies (Aim 2). Our multidisciplinary team, combining expertise in graph machine learning, reinforcement learning, natural language processing, and translational science, is uniquely positioned to execute this vision. Together, these innovations will support NIH priorities in precision medicine and translational informatics. Deliverables include open-source tools, a next-generation biomedical hypergraph resource, and an interpretable predictive framework that enables scalable, context-aware therapeutic discovery beyond the capabilities of conventional KG-based approaches.