Few-shot molecular property optimization via a domain-specialized large language model

Area of Science:

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

Background:

• Large language models (LLMs) show promise in various AI fields but struggle with molecular structure-property relationships in chemistry.
• Existing LLMs face limitations in few-shot learning for small-molecule generation and optimization, hindering drug discovery.
• Accurately capturing nuanced molecular structures is crucial for predicting pharmacochemical properties.

Purpose of the Study:

• To introduce DrugLLM, a novel LLM specifically designed for molecular optimization in drug discovery.
• To address the limitations of current LLMs in handling complex molecular data for drug development.
• To accelerate the process of identifying and optimizing bioactive compounds.

Main Methods:

• Development of DrugLLM, a specialized LLM for molecular optimization.
• Implementation of Functional Group Tokenization (FGT) for efficient molecular representation, achieving over 53% token compression compared to SMILES.
• Introduction of a novel pre-training strategy enabling iterative molecular structure prediction and modification for property optimization.

Main Results:

• DrugLLM demonstrated state-of-the-art performance in few-shot molecular generation, outperforming existing LLMs like GPT-4.
• The model successfully optimized HCN2 inhibitors, leading to the identification of two validated bioactive compounds.
• DrugLLM achieved significant token compression, improving efficiency in molecular data processing.

Conclusions:

• DrugLLM represents a significant advancement in applying LLMs to molecular optimization and drug discovery.
• The developed FGT method and pre-training strategy enhance LLM capabilities for chemical and biological applications.
• DrugLLM shows strong potential to accelerate AI-driven drug discovery through efficient molecular optimization.