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

Area of Science:

• Computational chemistry
• Bioinformatics
• Drug discovery

Background:

• Graph neural networks (GNNs) are effective for molecular topology but lack global chemical priors and struggle with few-shot learning and multi-task feature conflicts.
• Incorporating expert chemical knowledge into GNNs is crucial for improving molecular property prediction accuracy and interpretability.

Purpose of the Study:

• To develop FG-PromptNet, a novel cross-modal prompt learning framework for molecular property prediction.
• To infuse GNNs with chemical expert knowledge using a functional group structural-semantic dictionary.
• To enhance generalization in few-shot scenarios and resolve feature conflicts in multi-task learning.

Main Methods:

• Developed FG-PromptNet, a framework utilizing a functional group structural-semantic dictionary to enhance graph representations with chemical knowledge.
• Implemented a cross-attention module to focus on key molecular structures for accurate property inference.
• Introduced a task-aware gating mechanism for adaptive feature filtering to mitigate negative transfer in multi-task learning.
• Performed joint pre-training using graph-text contrastive learning and regression on a large molecular library.

Main Results:

• FG-PromptNet demonstrated superior performance over state-of-the-art methods on multiple MoleculeNet benchmark datasets.
• Achieved outstanding prediction accuracy, particularly in few-shot regression tasks.
• Interpretability experiments confirmed the model's ability to identify and locate functionally significant groups, providing chemical explanations for predictions.

Conclusions:

• FG-PromptNet effectively integrates chemical expert knowledge into GNNs for molecular property prediction.
• The framework shows significant improvements in accuracy, few-shot generalization, and interpretability for drug discovery applications.
• FG-PromptNet offers a promising approach for advancing computational drug screening and molecular property modeling.