Improving Molecular Understanding of Large Language Model via Substructure-aware Instruction Tuning

ACL ARR 2026 January Submission4416 Authors

Keywords: NLP Applications, AI4Science, Molecule Understanding

Abstract:

Large Language Models (LLMs) have shown strong performance in molecular tasks, yet they often fail to capture fine-grained molecular information, particularly the presence of substructures and how they behave across diverse chemical contexts. Most existing approaches rely on surface-level cues, treating substructures as isolated markers rather than modeling their functional behaviors.
We introduce SubMol-Instructions, a substructure-aware instruction tuning dataset that explicitly links molecular substructures to their functional behaviors across reaction prediction, property prediction, and molecule translation tasks. Building on this dataset, we propose StructMol, a molecule LLM that learns substructure behaviors using multiple 1D molecular representations.
Experimental results on diverse chemical tasks show that our approach consistently outperforms state-of-the-art baselines, highlighting the importance of explicitly defining and learning substructural behaviors for improving fine-grained molecular understanding.

Paper Type: Long

Research Area: NLP Applications

Research Area Keywords: NLP Applications,data augmentation

Contribution Types: NLP engineering experiment

Languages Studied: English

Submission Number: 4416