Generative AI in Drug Repurposing and Biomarker Discovery: A Multimodal Approach

EMERSON RAJA JOSEPH^1*Saranya K²Kalaiarasi Professor¹

• ¹Multimedia University, Cyberjaya, Malaysia
• ²Bannari Amman Institute of Technology Department of Computer Science and Engineering, Sathyamangalam, India

Computational drug repurposing has been widely explored using similarity-based methods, network diffusion, matrix factorization, deep learning, and graph neural networks (GNNs). Although recent heterogeneous GNN models such as TxGNN and GAT-based models have not only improved drug-disease association prediction, but also demonstrate serious limitations to use in real-world biomedical scenarios, including poor generalization to sparsely annotated diseases, limited disease-level adaptation, and inability to effectively combine heterogeneous evidence of curated databases, multi-omics profiles, and unstructured biomedical literature.This paper proposes a solution to these limitations: Heterogeneous Attention-based Meta-learning Graph Neural Network named HAMGNN, which uses three major innovations to handle these shortcomings: (i) relation-sensitive multi-head attention to prioritize biologically significant interactions among heterogeneous types of edges, (ii) a disease-focused meta-learning framework, that is able to adapt quickly to newly observed or under-informed diseases, and (iii) a literature-enhanced knowledge graph construction pipeline that encodes high-confidence, LLM-extracted therapeutic and biHAMGNN is tested on a large multimodal biomedical knowledge graph assembled on DrugBank, DisGeNET, and Hetionet which has more than 2.2 million edges. The proposed model has a ROC-AUC of 0.98 and a precision of 0.95 under a stringent disjoint disease-based (cold-start) evaluation protocol, which is an improvement of 10-15% over TxGNN and GAT-GNN on unseen disease generalization. Translational applicability of the framework is further demonstrated in Alzheimer disease and Long COVID case studies in which clinically plausible repurposing 2 candidates and disease-associated biomarker signature are identified using mechanistic pathways. On the whole, HAMGNN offers a generalized, biologically based, and unified model of evidence-based drug repurposing and biomarker discovery in complex and emerging diseases.