UniDrug-LLM: A Reliability-Aware Multimodal LLM-based Framework for Drug Discovery

Area of Science:

• Biomedical informatics
• Computational drug discovery
• Artificial intelligence in medicine

Background:

• Current drug discovery methods struggle with isolated data modalities, incomplete information, and unreliable predictions.
• Existing models often fail to generalize to new drugs or targets due to these limitations.

Purpose of the Study:

• To develop a unified, reliability-aware multi-modal fusion framework for biomedical drug discovery.
• To address modality isolation, incomplete relational supervision, and uneven per-sample reliability.

Main Methods:

• Proposed UniDrug-LLM, a framework integrating molecular structures, knowledge graphs, structure-aware features, and text.
• Introduced Cross-Modal Subspace Alignment (CMSA) to synthesize missing relational embeddings from incomplete knowledge graphs.
• Developed Cross-Modal Orthogonal Decomposition Network (CMON) for estimating per-sample modality reliability and adaptive fusion via digest tokens.
• Utilized a LoRA-tuned LLM backbone for end-to-end prediction.

Main Results:

• UniDrug-LLM achieved state-of-the-art performance, exceeding baselines by up to 8.8% in key tasks.
• Demonstrated strong generalization capabilities, particularly in drug-target interaction prediction for cold-start scenarios.
• Successfully integrated diverse data modalities for enhanced predictive power.

Conclusions:

• UniDrug-LLM offers a principled paradigm shift towards reliability-aware multi-modal learning in drug discovery.
• The framework effectively breaks down single-modal isolation, leading to more robust and generalizable predictions.
• This approach holds significant promise for accelerating the discovery of novel therapeutics.