Few-Shot Molecular Property Prediction via Prototype-Enhanced Contextual Embedding

Molecular property prediction presents significant challenges due to the limited availability of labeled data in many chemical and biological domains. To address this issue, we propose a few-shot molecular property prediction approach that leverages a Graph Isomorphism Network (GIN) as the molecular encoder and incorporates a self-attention mechanism to integrate prototype representations of both positive and negative samples into each molecular embedding, thereby obtaining enhanced molecular contextual embeddings. To enable rapid adaptation to novel tasks, we adopt a bi-level meta-learning framework in which the pretrained encoder is frozen, the outer loop learns transferable knowledge across tasks, and the inner loop selectively updates only the classifier parameters. This approach facilitates effective fewshot learning in molecular property prediction. Experimental results on benchmark few-shot molecular property datasets demonstrate the effectiveness of the proposed method.