Token-Mol: Accelerating 3D Drug Design with Language Models

Token-Mol 1.0, introduced by Jike Wang and collaborators, marks a breakthrough in token-based 3D drug design using large language models (LLMs). By encoding both 2D and 3D molecular structures, as well as chemical properties, into discrete tokens, Token-Mol bridges the gap between language models and spatial molecular understanding. Built on a transformer decoder with a novel Gaussian cross-entropy loss for regression tasks, it achieves over 10–20% improvements in molecular conformation generation and outperforms prior token-based models by 30% in property prediction. In real-world molecular docking scenarios, Token-Mol boosts drug-likeness and synthetic accessibility, while running up to 35 times faster than traditional diffusion-based models. Its integration with reinforcement learning further enhances drug affinity optimization, positioning Token-Mol as a powerful engine for next-generation AI-driven drug discovery.