DruGagent: Multi-Agent Large Language Model-Based Reasoning for Drug-Target Interaction Prediction.

Area of Science:

• Computational biology
• Artificial intelligence in drug discovery

Background:

• Large language models (LLMs) offer human-like interfaces but struggle with accuracy in complex scenarios requiring multiple perspectives.
• Existing drug-target interaction (DTI) prediction methods face challenges in interpretability and handling biological system complexity.

Purpose of the Study:

• To develop DrugAgent, a multi-agent LLM system for enhanced DTI prediction with transparent reasoning.
• To improve the consistency, reliability, and interpretability of DTI predictions for drug discovery.

Main Methods:

• Implemented a coordinator-based multi-agent architecture tailored for the DTI domain.
• Integrated domain-specific data sources: machine learning predictions, knowledge graphs, and literature evidence.
• Incorporated Chain-of-Thought (CoT) and ReAct (Reason+Act) frameworks for transparent DTI reasoning.

Main Results:

• DrugAgent achieved a 45% higher F1 score compared to a non-reasoning multi-agent model (GPT-4o mini) on a kinase inhibitor dataset (0.514 vs 0.355).
• Ablation studies identified the AI agent as most impactful, followed by KG and search agents.
• The system provides detailed, human-interpretable reasoning for DTI predictions.

Conclusions:

• DrugAgent offers a robust multi-agent LLM framework for accurate and interpretable DTI prediction.
• The integration of diverse data sources and reasoning frameworks is crucial for advancing drug discovery tools.
• Transparent reasoning is essential for clinical decision-making and regulatory compliance in biomedical applications.