Token-Mol 1.0: tokenized drug design with large language models.

Area of Science:

• Computational chemistry
• Artificial intelligence in drug discovery
• Molecular modeling

Background:

• Large language models (LLMs) show promise in drug design but often fail to incorporate 3D molecular structures effectively.
• Existing methods face challenges in representing complex molecular information for AI models.

Purpose of the Study:

• To introduce Token-Mol, a token-only model for 3D drug design that encodes 2D/3D structures and properties into discrete tokens.
• To enhance molecular conformation generation, property prediction, and pocket-based molecular generation using AI.

Main Methods:

• Developed Token-Mol, a transformer decoder-based model utilizing causal masking.
• Introduced a Gaussian cross-entropy loss function for regression tasks.
• Encoded 2D/3D molecular structures and properties into discrete tokens.

Main Results:

• Token-Mol improved molecular conformation generation by over 10% and 20% on two datasets.
• Achieved 30% better property prediction compared to existing token-only models.
• Enhanced drug-likeness and synthetic accessibility by approximately 11% and 14% in pocket-based generation.
• Demonstrated 35x speed improvement over diffusion models and improved real-world success rates.

Conclusions:

• Token-Mol offers a powerful and efficient solution for 3D drug design by integrating diverse molecular information.
• The model significantly advances AI-driven drug discovery through improved performance and speed.
• Combining Token-Mol with reinforcement learning further optimizes key drug-likeness and affinity parameters.