Consistency-regularized graph neural networks for molecular property prediction.

Area of Science:

• Computational Chemistry
• Machine Learning
• Cheminformatics

Background:

• Graph neural networks (GNNs) excel at molecular property prediction but struggle with small datasets.
• Standard data augmentation techniques often fail for molecular graphs, potentially altering inherent properties.

Purpose of the Study:

• To develop a novel method, consistency-regularized graph neural network (CRGNN), for effective molecular graph augmentation in GNN training.
• To enhance GNN performance on small molecular datasets by leveraging augmented graph representations.

Main Methods:

• Applied molecular graph augmentation to generate strongly and weakly augmented views of each molecular graph.
• Introduced a consistency regularization loss to encourage GNNs to map augmented views of the same graph closely in the representation space.
• Integrated this loss into the GNN learning objective.

Main Results:

• The CRGNN method effectively utilizes molecular graph augmentation, improving prediction performance.
• Demonstrated superior performance compared to existing methods on various molecular benchmark datasets.
• Performance gains were particularly significant on smaller training datasets.

Conclusions:

• Consistency regularization offers a viable strategy to mitigate the negative effects of molecular graph augmentation in GNNs.
• The proposed CRGNN method enhances the utility of data augmentation for molecular property prediction, especially in low-data regimes.
• This approach advances the application of GNNs in cheminformatics for small molecule datasets.