Multi agent large language models for biomedical hypothesis generation in drug combination discovery.

Area of Science:

• Artificial Intelligence
• Biomedical Research
• Pharmacology

Background:

• Large language models (LLMs) show promise in scientific reasoning but struggle with hypothesis generation in data-scarce fields.
• Predicting combinatorial therapies for complex diseases like Alzheimer's (AD) is challenging due to limited data.

Purpose of the Study:

• To introduce Coated-LLM, an AI framework for predicting efficacious combinatorial therapies in data-scarce domains, using AD as a case study.
• To leverage AI-driven scientific collaboration to overcome limitations in traditional data-driven prediction methods.

Main Methods:

• Coated-LLM utilizes specialized LLM agents (Researcher, Reviewers, Moderator) for systematic hypothesis generation and evaluation.
• In-context learning techniques are employed to enhance the AI's reasoning capabilities.
• The framework was tested using Alzheimer's disease data, comparing its performance against traditional knowledge-based approaches.

Main Results:

• Coated-LLM achieved higher accuracy (0.74) than traditional methods (0.52) in predicting therapeutic efficacy for Alzheimer's disease.
• External validation further confirmed the framework's predictive power with an accuracy of 0.82.
• A novel drug combination identified by Coated-LLM was experimentally validated to significantly reduce amyloid aggregation in vitro.

Conclusions:

• Coated-LLM demonstrates a scalable approach for hypothesis generation in biomedical research, augmenting human scientific reasoning.
• The AI framework shows significant potential for accelerating the discovery of novel therapeutic strategies for complex diseases like Alzheimer's.
• This work highlights the capability of AI in addressing data scarcity challenges in scientific discovery.