Small Molecule Optimization with Large Language Models

The researchers proposed a new algorithm called Mol-E, which combines the generative capabilities of large language models with evolutionary algorithms. Mol-E is trained based on molecules and their properties, aiming to solve small molecule optimization problems. On the Practical Molecular Optimization benchmark, this model achieved new state-of-the-art performance: the Top-10 AUC value was 17.500 in the task-independent mode and 20.551 in the task-informed mode. Additionally, Mol-E performed better than existing evaluation baseline models in multi-property optimization tasks as well as molecular docking tasks for DRD2, MK2, and AChE.