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

Area of Science:

• Computational biology
• Bioinformatics
• Artificial intelligence in medicine

Background:

• Existing graph neural network (GNN) models for drug repurposing struggle with new diseases and integrating diverse data sources.
• Limitations include poor generalization to sparsely annotated diseases and inability to combine curated databases, multi-omics, and literature.

Purpose of the Study:

• To introduce HAMGNN, a heterogeneous attention-based meta-learning GNN for improved drug repurposing.
• To address limitations of current GNNs in handling sparse data and integrating heterogeneous biomedical information.

Main Methods:

• Developed HAMGNN with relation-sensitive multi-head attention and a disease-focused meta-learning framework.
• Constructed a literature-enhanced knowledge graph using LLM-extracted therapeutic information from DrugBank, DisGeNET, and Hetionet.
• Evaluated on a large multimodal biomedical knowledge graph using a stringent cold-start protocol.

Main Results:

• HAMGNN achieved 0.98 ROC-AUC and 0.95 precision, outperforming existing models by 10%-15% on unseen diseases.
• Demonstrated translational applicability in Alzheimer's disease and Long COVID case studies.
• Identified plausible repurposing candidates and biomarker signatures through mechanistic pathways.

Conclusions:

• HAMGNN provides a generalized, biologically grounded framework for evidence-based drug repurposing.
• Enables effective biomarker discovery for complex and emerging diseases.
• Represents a unified approach for integrating heterogeneous biomedical evidence.