A knowledge-guided pre-training framework for improving molecular representation learning.

Area of Science:

• Computational chemistry
• Artificial intelligence in drug discovery
• Machine learning for molecular property prediction

Background:

• Effective molecular feature representation is crucial for drug discovery.
• Pre-training graph neural networks (GNNs) using self-supervised learning addresses data scarcity in molecular property prediction.
• Existing self-supervised methods face challenges due to ill-defined strategies and limited GNN capacity.

Purpose of the Study:

• To propose Knowledge-guided Pre-training of Graph Transformer (KPGT), a self-supervised learning framework.
• To generate generalizable and robust molecular representations.
• To overcome limitations of current self-supervised learning approaches for molecular property prediction.

Main Methods:

• Developed KPGT, integrating a graph transformer tailored for molecular graphs.
• Employed a knowledge-guided pre-training strategy to capture molecular structural and semantic knowledge.
• Utilized extensive computational testing across 63 datasets.

Main Results:

• KPGT demonstrated superior performance in predicting molecular properties across diverse domains.
• The framework successfully identified potential inhibitors for antitumor targets (HPK1 and FGFR1).
• KPGT provides enhanced molecular representations compared to existing methods.

Conclusions:

• KPGT offers a powerful tool for advancing AI-aided drug discovery.
• The framework effectively captures both structural and semantic information in molecules.
• KPGT shows practical applicability and validates its utility in identifying drug candidates.