Improving Molecular Property Prediction in Small Language Models Using Graph-based Tools

Researchers propose a Context-Augmented Prompting framework that uses graph neural network tools to improve molecular property prediction in small language models. The method yields accuracy gains of up to 74% on Tox21 and over 25% on MUTAG, though a gap remains compared to specialized GNN models.

arXiv:2607.13115v1 Announce Type: new Abstract: Small language models SLMs have shown promise for zero-shot molecular property prediction from SMILES strings, yet they often suffer from structural blindness because sequence representations under-specify key graph-topological cues. We propose a modular Context-Augmented Prompting framework that enables agentic tool use at inference time: a trained GNN expert model provides a predictive hint with confidence, and a GNN extracts an instance-specific explanatory subgraph e.g., a subgraph SMILES and an accompanying explanatory paragraph . We evaluate three commonly used SLMs on MUTAG and Tox21 under five prompting configurations ranging from SMILES-only to using all available tools at hand. Across two datasets, enriching prompts with graph-derived context yields substantial accuracy gains, often exceeding 25% relative improvement and up to 74% on Tox21. We further validate the functional relevance of the extracted motifs via a necessity-based edge-drop intervention. Despite the observed gains, a persistent gap remains to specialized GNN models, highlighting both the value and limits of text-conditioned reasoning for molecular structure.