Identifying Semantic Component for Robust Molecular Property Prediction.

Area of Science:

• Machine Learning
• Computational Chemistry
• Artificial Intelligence

Background:

• Graph neural networks excel at molecular property prediction but struggle with out-of-distribution generalization.
• Current methods often rely on discriminative representations, which can lead to misidentifications and reduced model robustness.

Purpose of the Study:

• To propose a generative model, Semantic-Components Identifiability (SCI), to enhance the generalization ability of graph neural networks for molecular property prediction.
• To improve model robustness by explicitly identifying semantic-relevant (SR) and semantic-irrelevant (SI) latent variables.

Main Methods:

• Formulated a data generation process from atom to molecular level, splitting the latent space into SI substructures, SR substructures, and SR atom variables.
• Restricted minimal changes of SR atom variables and applied semantic latent substructure regularization to mitigate variance.
• Proved block-wise identifiability of SR substructures and comment-wise identifiability of SR atom variables.

Main Results:

• Achieved state-of-the-art performance on molecular property prediction tasks.
• Demonstrated general improvement across 21 datasets in 3 mainstream benchmarks.
• Visualization results provided insightful explanations for prediction outcomes.

Conclusions:

• The proposed SCI method enhances out-of-distribution generalization in graph neural networks for molecular property prediction.
• SCI's generative approach improves model robustness by ensuring semantic-components identifiability.
• The method offers interpretable insights into prediction mechanisms.