FG-PromptNet: Cross-modal prompt learning for explainable molecular property prediction.

Molecular property prediction is a critical task in the field of drug discovery and bioinformatics. In recent years, while graph neural networks excel at capturing molecular topology, they struggle with incorporating global chemical priors, generalizing to few-shot scenarios, and resolving feature conflicts in multi-task learning. To address these issues, we propose FG-PromptNet, a cross-modal prompt learning framework that leverages a functional group structural-semantic dictionary to infuse graph representations with chemical expert knowledge. Firstly, a cross-attention module is designed to guide the model to accurately focus on the key structures that determine molecular properties during inference. Secondly, a task-aware gating mechanism is introduced to achieve adaptive feature filtering for different prediction tasks, effectively alleviating the negative transfer phenomenon. In addition, the model performs joint pre-training of graph-text contrastive learning and regression tasks on a large-scale unlabeled molecular library. Experimental results on multiple benchmark datasets of MoleculeNet show that FG-PromptNet outperforms existing state-of-the-art methods in most tasks, and exhibits outstanding prediction accuracy especially in few-shot regression tasks. Further interpretability experiments confirm that the model can identify and locate functional group structures with key biochemical significance, providing reliable chemical explanations for drug screening.