The field of molecular property prediction and generation is rapidly advancing, driven by the integration of large language models (LLMs) and innovative frameworks. Recent developments have focused on improving the interpretability and performance of LLMs in predicting molecular properties and generating molecular structures. Notably, the use of chain-of-thought reasoning, attribute-guided reinforcement learning, and causality-aware transformers has shown promise in enhancing the accuracy and diversity of molecular generation. Additionally, the application of LLMs in organic synthesis has transformed the field, enabling the automation of reaction prediction and execution. Overall, these advances have the potential to revolutionize the field of molecular property prediction and generation, enabling faster and more accurate discovery of new molecules and reactions.

Noteworthy papers include: CoTox, which proposes a novel framework for multi-toxicity prediction using LLMs and chain-of-thought reasoning, achieving state-of-the-art performance and improved interpretability. AttriLens-Mol, which introduces an attribute-guided reinforcement learning framework for molecular property prediction with LLMs, significantly boosting performance and enabling more effective prediction of molecular properties. MolSnap, which proposes a causality-aware framework for molecular generation conditioned on textual descriptions, achieving high-quality and diverse generation while maintaining fast inference. Large Language Models Transform Organic Synthesis, which surveys the milestones that have turned LLMs into practical lab partners, enabling the automation of reaction prediction and execution and supporting greener, data-driven chemistry.