Mix-Key: graph mixup with key structures for molecular property prediction

Area of Science:

• Computational chemistry
• Machine learning for drug discovery

Background:

• Molecular property prediction is crucial but hindered by limited labeled data requiring extensive experiments.
• Data augmentation techniques like Mixup show promise but face challenges with molecular graph data.
• The non-Euclidean nature of molecular graphs and sensitivity to structural changes limit traditional augmentation methods.

Purpose of the Study:

• To develop a novel data augmentation method, Mix-Key, specifically designed for molecular property prediction.
• To address the limitations of existing methods in handling molecular graph structures and data scarcity.
• To improve the accuracy and robustness of molecular property prediction models.

Main Methods:

• Mix-Key focuses on molecular scaffolds and functional groups, generating invariant isomers to preserve core molecular information.
• Incorporates molecular fingerprint similarity and node similarity to capture scaffold-functional group interactions and ensure graph correlation.
• Determines an optimal mixup ratio between original graphs and generated isomers for enhanced data augmentation.

Main Results:

• Extensive validation on diverse molecular datasets using various Graph Neural Network architectures.
• Mix-Key consistently demonstrated superior performance compared to existing data augmentation methods.
• Significant improvements in molecular property prediction accuracy were observed across multiple datasets.

Conclusions:

• Mix-Key is an effective data augmentation strategy for molecular property prediction, overcoming challenges associated with graph data.
• The method enhances model performance by generating informative augmented molecular graphs that preserve essential chemical information.
• Mix-Key offers a promising solution for data scarcity in computational chemistry and drug discovery applications.