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

arXiv cs.AI · 18 hours ago

Researchers developed a Context-Augmented Prompting framework that enhances small language models' ability to predict molecular properties by incorporating graph neural network tools that provide structural hints and explanatory subgraphs during inference. The framework achieved accuracy improvements exceeding 25% relative gain on the MUTAG dataset and up to 74% on the Tox21 dataset compared to SMILES-only prompts. Despite these gains, a performance gap remains between the augmented language models and specialized graph neural network models for molecular property prediction.