Periodicity-aware deep learning for polymers.

Area of Science:

• Polymer Chemistry
• Computational Chemistry
• Materials Science

Background:

• Deep learning accelerates chemical research but lacks unified frameworks for complex polymer structures.
• Current methods oversimplify polymers, neglecting periodicity and limiting model generalization.
• A unified deep learning approach is needed to capture the inherent periodicity of polymers.

Purpose of the Study:

• Introduce PerioGT, a novel periodicity-aware deep learning framework for polymer chemistry.
• Enhance the generalization capabilities of deep learning models in polymer science.
• Leverage polymer periodicity to improve downstream task performance and discover new materials.

Main Methods:

• Developed a periodicity-aware deep learning framework, PerioGT, for polymers.
• Incorporated a chemical knowledge-driven periodicity prior using contrastive learning during pre-training.
• Employed a graph augmentation strategy with virtual nodes to model complex chemical interactions.

Main Results:

• PerioGT achieved state-of-the-art performance on 16 diverse downstream tasks.
• Wet-lab experiments validated PerioGT's real-world potential by identifying two novel antimicrobial polymers.
• Introducing the periodicity prior significantly enhanced model performance across tasks.

Conclusions:

• PerioGT offers a unified deep learning framework for polymer chemistry by incorporating periodicity.
• The periodicity prior is crucial for improving model generalization and performance in polymer research.
• PerioGT demonstrates significant potential for accelerating materials discovery, including the identification of functional polymers.