[AML] A Multi-Agent Framework for Reliable and Consistent Drug-Target Interaction Prediction Using Large Language Models
Keywords: Drug-Target Interaction (DTI) Prediction, Large Language Models (LLMs), Multi-Agent Framework
Abstract: Drug-target interaction (DTI) prediction plays a vital role in accelerating drug discovery by identifying potential drug candidates. Despite significant advances in large language models, challenges such as prediction inconsistencies and AI hallucinations hinder the effective use of (LLMs in this domain. This paper proposes a domain-knowledge-guided, multi-agent framework to address these issues and enhance the reliability of DTI predictions. Inspired by collaborative human problem-solving, the framework decomposes the prediction process into specialized sub-tasks—protein sequence analysis, drug molecule analysis, and binding affinity prediction—each managed by distinct LLM agents. These agents collaborate within a mixture-of-experts model, with a debate-based ensemble method resolving discrepancies and ensuring robust final predictions. Using the BindingDB dataset, the framework is tested against baseline LLMs and fine-tuned machine learning models, demonstrating improved accuracy and prediction consistency. The findings suggest that this multi-agent approach not only advances DTI prediction but also holds promise for broader scientific discovery applications where naive LLM deployment is insufficient.
Submission Number: 20
Loading