Simulate Scientific Reasoning with Multiple Large Language Models: An Application to Alzheimer's Disease...

Area of Science:

• Artificial Intelligence in Medicine
• Computational Biology
• Drug Discovery

Background:

• Data-driven prediction of combinatorial therapies is often infeasible due to limited experimental data.
• Traditional knowledge-based methods have limitations in predicting complex therapeutic interactions.

Purpose of the Study:

• To develop an AI-driven framework utilizing large language models (LLMs) to simulate scientific reasoning and peer review.
• To predict efficacious combinatorial therapies in scenarios where data-driven prediction is not viable.

Main Methods:

• The framework integrates LLMs to mimic scientific reasoning and peer review processes.
• It incorporates few-shot learning, external knowledge, self-consistency checks, and internal review mechanisms.
• The approach focuses on knowledge-driven hypothesis generation.

Main Results:

• The proposed framework achieved a prediction accuracy of 0.74, outperforming traditional knowledge-based prediction (0.52).
• External validation using private experimental data demonstrated a high accuracy of 0.82.
• Ablation studies confirmed the critical role of high-quality examples, external knowledge, and self-consistency.

Conclusions:

• The AI framework provides an automated and effective method for generating high-quality hypotheses in biological inference.
• This approach offers a viable alternative for combinatorial therapy prediction when data is scarce.
• The developed framework has the potential to accelerate drug discovery and therapeutic development.