Generative AI in drug repurposing and biomarker discovery: a multimodal approach

Computational drug repurposing has been widely explored using similarity-based methods, network diffusion, matrix factorization, deep learning, and graph neural networks (GNNs). However, recent heterogeneous GNN models, such as TxGNN and GAT-based models, demonstrate serious limitations for real-world biomedical applications, including poor generalization to sparsely annotated diseases, limited disease-level adaptation, and inability to effectively combine heterogeneous evidence from curated databases, multi-omics profiles, and unstructured biomedical literature.

This article proposes a heterogeneous attention-based meta-learning graph neural network named HAMGNN, which employs three major innovations: (i) relation-sensitive multi-head attention to prioritize biologically significant interactions among heterogeneous edge types, (ii) a disease-focused meta-learning framework enabling rapid adaptation to newly observed or under-informed diseases, and (iii) a literature-enhanced knowledge graph construction pipeline encoding high-confidence, LLM-extracted therapeutic information. The model was tested on a large multimodal biomedical knowledge graph assembled from DrugBank, DisGeNET, and Hetionet, comprising more than 2.2 million edges, using a stringent disjoint disease-based (cold-start) evaluation protocol.

HAMGNN achieved a receiver operating characteristic-area under the curve (ROC-AUC) of 0.98 and precision of 0.95, representing a 10%-15% improvement over TxGNN and GAT-GNN on unseen disease generalization. Translational applicability was demonstrated through Alzheimer's disease and Long COVID case studies, identifying clinically plausible repurposing candidates and disease-associated biomarker signatures via mechanistic pathways.

HAMGNN offers a generalized, biologically grounded, and unified framework for evidence-based drug repurposing and biomarker discovery in complex and emerging diseases.