Predicting molecular properties based on the interpretable graph neural network with multistep focus mechanism.

Area of Science:

• Computational Chemistry
• Machine Learning
• Drug Discovery

Background:

• Deep learning, particularly graph neural networks (GNNs), excels at molecular property prediction.
• Current GNN models often function as black boxes, hindering trust and understanding of their predictions.

Purpose of the Study:

• To develop a novel GNN model that provides interpretable predictions for molecular properties.
• To enhance user trust in deep learning models for chemical applications.

Main Methods:

• Proposed the Iteratively Focused Graph Network (IFGN), a novel GNN architecture.
• Implemented a multistep focus mechanism to identify key atoms/groups influencing predictions.
• Integrated visualization techniques for generating multistep interpretations of model behavior.

Main Results:

• The IFGN model demonstrated strong prediction performance across eight diverse datasets.
• The multistep focus mechanism significantly improved model interpretability.
• The model's ability to identify key molecular features was validated.

Conclusions:

• The IFGN model offers a powerful and interpretable approach to molecular property prediction.
• The proposed mechanism enhances both prediction accuracy and model transparency.
• A publicly accessible website for the IFGN model has been developed for researcher convenience.