MR-MoL: Multi-Granular Rationale-Guided Molecular LLM for Property Prediction
A recent preprint on arXiv (2608.10480) presents MR-MoL, a multi-granular, rationale-guided molecular large language model (LLM) aimed at enhancing the prediction of molecular properties, crucial for drug discovery. Unlike traditional molecular LLMs that utilize 1D SMILES sequences or 2D molecular graphs—encoding information in a way that obscures the roles of specific substructures—MR-MoL offers internal evidence through a fine-tuned graph neural network (GNN) that scores each substructure via masking. The model generates a ranked, direction-tagged rationale of the most significant substructures, which the LLM processes alongside the SMILES sequence and molecular graph. This rationale includes three levels of granularity: Murcko scaffolds with side chains and two additional unspecified levels. This methodology aims to enhance the model's transparency and align its reasoning with chemists' evaluations, potentially improving predictive accuracy and interpretability, particularly in the realm of AI and computational drug discovery.
Key facts
- MR-MoL is a multi-granular rationale-guided molecular LLM for property prediction.
- It is described in arXiv preprint 2608.10480, announced as a new paper.
- The model uses a fine-tuned GNN to score substructures via masking.
- The most influential substructures are serialized as a ranked, direction-tagged rationale.
- The rationale is read alongside the SMILES sequence and molecular graph.
- The rationale spans three levels of granularity, including Murcko scaffolds with side chains.
- The goal is to provide internal evidence for predictions, unlike external retrieval methods.
- The work targets drug discovery applications.
Entities
Institutions
- arXiv