KGG: Knowledge-Guided Graph Self-Supervised Learning to Enhance Molecular Property Predictions.

Area of Science:

• Computational chemistry
• Machine learning in drug discovery
• Materials science informatics

Background:

• Graph Neural Networks (GNNs) show promise in molecular representation learning.
• Challenges include data scarcity and limited generalization due to expensive labeled data acquisition.
• Current GNNs often lack comprehensive chemical domain knowledge, such as orbital information, in their initial features.

Purpose of the Study:

• To address data scarcity and feature limitations in GNNs for molecular property prediction.
• To introduce a novel Knowledge-Guided Graph (KGG) framework.
• To improve the efficiency and accuracy of molecular property prediction using orbital-informed features.

Main Methods:

• Developed a Knowledge-Guided Graph (KGG) framework utilizing self-supervised learning.
• Pretrained models using orbital-level features to reduce reliance on extensive labeled datasets.
• Proposed novel representations for atomic hybridization and bond types incorporating orbital engagement.
• Utilized a cost-efficient pretraining strategy on approximately 250,000 molecules from the ZINC15 dataset.

Main Results:

• The KGG framework significantly outperforms state-of-the-art baselines on diverse downstream molecular property prediction tasks.
• Demonstrated data efficiency, requiring fewer molecules for pretraining compared to contemporary methods.
• t-SNE visualizations and comparisons with traditional molecular fingerprints validated the approach's effectiveness and robustness.

Conclusions:

• The KGG framework offers data efficiency and architectural versatility through orbital-informed representations.
• It effectively distills chemical knowledge from modest datasets, avoiding extensive pretraining.
• The approach excels in low-data fine-tuning, providing a robust foundation for various GNN architectures in drug discovery and materials science.