Self-Supervised Pre-Training via Multi-View Graph Information Bottleneck for Molecular Property Prediction.

Area of Science:

• Computational chemistry
• Machine learning in drug discovery
• Bioinformatics

Background:

• Molecular representation learning accelerates drug discovery using machine learning embeddings.
• Self-supervised pre-training is crucial due to limited labeled molecular data.
• Existing graph pre-training methods face limitations in preserving molecular properties during augmentation.

Purpose of the Study:

• To propose a novel self-supervised molecular pre-training model, the Molecular Graph Information Bottleneck (MGIB).
• To address limitations in graph augmentation and subgraph identification for molecular learning.
• To enhance the effectiveness of molecular representation learning for drug discovery tasks.

Main Methods:

• Developed the self-supervised Molecular Graph Information Bottleneck (MGIB) model.
• Employed atom and motif views for observing molecular graphs.
• Utilized a learnable graph compression process to extract core subgraphs.
• Extended the graph information bottleneck concept to self-supervised pre-training.

Main Results:

• MGIB effectively extracts informative molecular subgraphs, demonstrating interpretability.
• The self-supervised graph information bottleneck significantly contributes to representation learning.
• MGIB achieved superior performance in molecular property prediction across 7 binary classification and 6 regression tasks.

Conclusions:

• MGIB offers an effective approach to self-supervised molecular pre-training.
• The model enhances molecular representation learning by preserving intrinsic properties and identifying key subgraphs.
• MGIB demonstrates significant potential for advancing drug analysis and discovery.