SciGroveBeta
Cheminformatics

Multi-Granular Rationale-Guided Molecular LLM for Property Prediction

Junwoo Park, Minyoung Shin, Cheol Soon Lee, Sujee Lee

Featured August 16, 2026

Original Paper

AI-generated analysis — This is SciGrove's AI interpretation of the paper, not peer-reviewed content. Always refer to the original paper.

Simply

A new AI model for chemistry learns to predict molecule properties better by not just looking at the whole molecule, but also by getting simple, ranked clues about which parts of the molecule make a property go up or down.

In depth
The paper introduces MR-MoL, a molecular LLM that enhances property prediction by providing explicit, multi-granular rationale derived from a GNN. This rationale, which includes ranked and direction-tagged substructures, is fed to the LLM alongside SMILES and molecular graph representations, allowing the model to ground its predictions in chemically intuitive evidence. This approach significantly improves performance and demonstrates the LLM's ability to interpret and utilize structural cues.

Key Takeaways

  • 1
    MR-MoL integrates GNN-derived attributions as explicit, ranked, and direction-tagged textual rationales into a molecular LLM's input.
  • 2
    The rationale provides multi-granular structural evidence (Murcko scaffolds, BRICS fragments, functional groups) to guide property predictions.
  • 3
    The model significantly outperforms generalist LLMs and narrows the performance gap to specialist models on MoleculeNet tasks, demonstrating effective use of the rationale.

Conceptual Flow

HIGH LEVEL
1
Methodology (The "Logic")

The model takes molecule pictures and text, then adds special clues about important parts to help a big AI brain predict properties.

Molecule Picture
Molecule Text
Find Key Parts
Key Part Clues
Combined Info
2
Results (The "Impact")

By using these special clues, the new model predicts molecule properties much better than other general AI models.

Old AI Prediction
New AI Prediction
Compare Accuracy
New AI is Better

This breakdown was generated by SciGrove. Get the same analysis — intuition, storyboard, peer review, a runnable prototype and a glossary — on any paper you upload or paste a DOI for.