Improving Molecular Property Prediction in Small Language Models Using Graph-based Tools
For researchers using SLMs in molecular property prediction, this work provides a method to enhance zero-shot performance, but the gains are incremental as SLMs still lag behind specialized GNNs.
Small language models (SLMs) struggle with molecular property prediction due to structural blindness from SMILES strings. The proposed Context-Augmented Prompting framework, using GNN tools, improves accuracy by up to 74% on Tox21 and over 25% relative improvement, though a gap to specialized GNNs remains.
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.