A new molecular large language model (LLM) has been proposed for molecular property prediction. The model, called MR-MoL, uses a multi-granular rationale-guided approach to provide direct evidence of the influence of internal substructures on molecular properties. MR-MoL uses a fine-tuned graph neural network (GNN) to score each substructure and serializes the most influential ones as a ranked, direction-tagged rationale. The rationale spans three levels of granularity: Murcko scaffolds with their side chains, BRICS fragments, and functional groups. The results show that MR-MoL achieves the best overall results on eight MoleculeNet tasks and further confirms that the model reads and utilizes the rationale in five diagnostic experiments. Blogger's Review: The proposal of MR-MoL marks a significant advance in the field of molecular property prediction, and its multi-granular rationale-guided approach enables the model to better understand the influence of internal substructures, thereby improving the accuracy of predictions.