Rationale-Guided LLM Improves Molecular Property Prediction

This research introduces MR-MoL, a novel multi-granular rationale-guided molecular Large Language Model (LLM) designed to improve molecular property prediction, a critical task in drug discovery. Existing molecular LLMs often represent molecules implicitly through SMILES sequences or molecular graphs, making the contribution of individual substructures opaque. While retrieval methods add external context, MR-MoL focuses on providing internal, chemist-like reasoning cues. The core innovation involves using a fine-tuned Graph Neural Network (GNN) to score individual substructures within a molecule. The most influential substructures are then serialized into a ranked, direction-tagged rationale, which the LLM reads alongside the traditional SMILES sequence and molecular graph. This rationale is provided at three levels of granularity: Murcko scaffolds, BRICS fragments, and functional groups, offering a comprehensive view of substructure contributions. MR-MoL is the first method to directly expose GNN-derived attributions to an LLM as evidence for property prediction. Experimental results on eight MoleculeNet tasks show that MR-MoL achieves the best overall performance among generalist models and significantly narrows the performance gap with specialist models. Diagnostic tests confirm that the model actively uses the rationale's direction, rank, and substructure information, reproducing known structure-property relationships.