Technique of Augmenting Molecular Graph Data by Perturbating Hidden Features.

Area of Science:

• Computational chemistry
• Machine learning
• Drug discovery

Background:

• Quantitative structure-property relationship (QSPR) models are crucial for identifying molecules with specific properties in drug discovery and materials science.
• Graph neural networks (GNNs) have demonstrated strong predictive performance in QSPR modeling.
• Training GNNs typically requires large datasets, posing a challenge for studies with limited samples.

Purpose of the Study:

• To develop an effective data augmentation method for GNNs applicable to small datasets.
• To improve the feature extraction capabilities of GNNs when training data is scarce.
• To enhance the prediction performance of GNNs in quantitative structure-property relationship tasks.

Main Methods:

• A novel graph data augmentation technique was designed by introducing random perturbations to vertex features during the message passing phase of GNNs.
• The proposed method was evaluated on both regression and classification tasks.
• The effectiveness of data augmentation was analyzed based on the timing of perturbation addition within the GNN architecture.

Main Results:

• The proposed data augmentation method proved effective in enhancing GNN performance for QSPR modeling.
• Perturbations were most impactful when applied immediately before the readout layer of the GNN.
• The benefits of this data augmentation approach were most pronounced on small datasets, particularly those with around 1000 samples.

Conclusions:

• The developed graph data augmentation method offers a viable solution for improving GNN performance with limited data.
• This technique is particularly valuable for accelerating drug discovery and materials development processes where large datasets may not be available.
• The findings highlight the potential of targeted data augmentation strategies in advancing machine learning applications in chemistry.