INTransformer: Data augmentation-based contrastive learning by injecting noise into transformer for molecular...

Area of Science:

• Computational chemistry
• Drug discovery
• Machine learning

Background:

• Molecular property prediction is crucial for identifying drug candidates.
• Deep learning models require substantial labeled data, which is often scarce in this field.
• Capturing global molecular information is essential for accurate property prediction.

Purpose of the Study:

• To develop a data augmentation method for molecular property prediction.
• To address the challenge of limited labeled molecular data.
• To enhance the ability of models to capture global molecular information.

Main Methods:

• Proposed INTransformer, a data augmentation technique utilizing contrastive learning.
• Employed two identical Transformer sub-encoders to process original and noisy SMILES data.
• Applied contrastive learning to ensure consistent molecular encoding between original and augmented data.

Main Results:

• INTransformer effectively augments limited labeled molecular data.
• The method enhances the extraction of global molecular representations.
• Achieved competitive performance on benchmark datasets for molecular property prediction.

Conclusions:

• INTransformer alleviates data limitations in deep learning for molecular property prediction.
• The approach improves the capture of critical global molecular information.
• Demonstrated competitive and effective performance against existing methods.