Graph Tools Enhance Small Language Models for Molecular Prediction.

Small language models (SLMs) show promise for predicting molecular properties directly from SMILES strings, but they often struggle with "structural blindness" because sequence representations don't fully capture crucial graph-topological information. To overcome this, a new "Context-Augmented Prompting" framework has been developed. This framework enables SLMs to use agentic tools during inference. Specifically, a trained Graph Neural Network (GNN) expert provides a predictive hint along with its confidence, and another GNN extracts an instance-specific explanatory subgraph (e.g., a subgraph SMILES and a descriptive paragraph). By enriching prompts with this graph-derived context, the framework achieved substantial accuracy gains across multiple datasets, sometimes exceeding 74% relative improvement. While significant, a gap still remains between these enhanced SLMs and specialized GNN models, indicating both the value and current limits of text-conditioned reasoning for molecular structures.