MR-MoL Uses Multi-Granular Rationales for Molecular Property Prediction
August 12, 2026
MR-MoL integrates fine-tuned GNN substructure scores as ranked, direction-tagged rationales for LLMs. By serializing these rationales alongside SMILES and molecular graphs, the model provides interpretable evidence for molecular property predictions.
HOW THIS AFFECTS YOU
●
researcherYou can improve the interpretability of molecular LLMs by providing explicit substructure rationale.
●
healthThis adds a layer of chemical reasoning transparency to AI-driven drug discovery workflows.