AutoLead: An LLM-Guided Bayesian Optimization Framework for Multi-Objective Lead Optimization.

Area of Science:

• Medicinal Chemistry
• Computational Drug Discovery
• Artificial Intelligence in Pharmacology

Background:

• Drug discovery lead optimization is a complex, multiobjective challenge hindering new therapeutic development.
• Traditional methods face difficulties in exploring chemical space and optimizing conflicting molecular properties simultaneously.

Purpose of the Study:

• To present AutoLead, a novel framework integrating large language models (LLMs) and multiobjective Bayesian optimization for efficient lead optimization.
• To demonstrate AutoLead's capability in guiding the search for novel drug-like molecules satisfying multiple objectives.

Main Methods:

• Integration of LLMs for chemical reasoning with multiobjective Bayesian optimization.
• Evaluation on two molecular optimization tasks and a new benchmark dataset simulating realistic lead optimization scenarios.
• Development of a dataset focused on modifying compounds violating Lipinski's rule of five to meet criteria and improve QED score.

Main Results:

• AutoLead achieved state-of-the-art results on molecular optimization tasks.
• The framework demonstrated effectiveness in guiding the search for molecules with multiple optimized properties.
• A case study highlighted the potential of LLMs combined with black-box optimization for practical drug discovery.

Conclusions:

• Combining LLMs with black-box optimization offers a more efficient and practical approach to drug discovery.
• AutoLead represents a significant advancement in tackling the multiobjective challenges of lead optimization.
• The developed benchmark dataset facilitates further research in realistic drug discovery scenarios.