DrugAssist: a large language model for molecule optimization.

Area of Science:

• Artificial Intelligence in Drug Discovery
• Computational Chemistry
• Machine Learning for Molecular Optimization

Background:

• Large language models (LLMs) demonstrate strong performance across various tasks, prompting their application in drug discovery.
• Molecule optimization, a crucial step in drug development, remains underexplored by current LLM applications.
• Existing methods often neglect expert feedback and iterative refinement, key components of successful drug discovery.

Purpose of the Study:

• To introduce DrugAssist, an interactive LLM-based model for molecule optimization.
• To address the limitations of non-interactive LLM approaches in drug discovery.
• To leverage LLM interactivity and generalizability for enhanced molecule optimization.

Main Methods:

• Developed DrugAssist, an interactive molecule optimization model utilizing human-machine dialogue.
• Employed LLM's interactivity and generalizability for iterative refinement of molecular properties.
• Leveraged expert feedback within the optimization loop.

Main Results:

• DrugAssist achieved state-of-the-art results in both single and multiple property optimization.
• Demonstrated significant potential in transferability and iterative optimization capabilities.
• Publicly released the 'MolOpt-Instructions' dataset for LLM fine-tuning in molecule optimization.

Conclusions:

• Interactive LLM approaches like DrugAssist can significantly advance molecule optimization in drug discovery.
• The integration of expert knowledge with LLMs offers a promising direction for future research.
• Publicly available code and data will facilitate further advancements in LLM-driven drug discovery.