Large Language Model Agent for Modular Task Execution in Drug Discovery.

Area of Science:

• Computational chemistry
• Artificial intelligence in drug discovery
• Bioinformatics

Background:

• Early-stage drug discovery involves complex, multi-step computational processes.
• Automating these tasks can significantly accelerate the identification of novel therapeutics.
• Integrating diverse computational tools remains a challenge.

Purpose of the Study:

• To develop and validate a modular AI-powered framework for automating early-stage computational drug discovery.
• To enhance the efficiency and accuracy of key tasks including data retrieval, molecular generation, and property prediction.
• To demonstrate the framework's capability in iterative molecular refinement and 3D structure generation.

Main Methods:

• A modular framework integrating large language models (LLMs) with domain-specific computational tools.
• Utilizing retrieval-augmented generation for literature-grounded question answering.
• Employing generative models for de novo molecular design and multiproperty prediction.
• Implementing iterative refinement guided by predicted properties and drug-likeness filters.
• Generating 3D protein-ligand complexes using specialized models (e.g., Boltz-2).

Main Results:

• The framework successfully automated biomedical data retrieval, literature Q&A, molecular generation, and property prediction.
• Iterative refinement increased molecules with QED >0.6 from 34 to 55 and Ghose filter compliance from 32 to 55 (out of 100).
• 3D protein-ligand complexes were generated with rapid binding affinity estimates, aiding structure evaluation.

Conclusions:

• The AI framework effectively streamlines and automates critical early-stage drug discovery tasks.
• The modular design allows for flexible integration of new tools, supporting scalable AI-assisted therapeutic discovery.
• This approach shows significant potential for accelerating the screening, prioritization, and evaluation of drug candidates.