Virtual Bonding Enhanced Graph Self-Supervised Learning for Molecular Property Prediction.

Area of Science:

• Computational chemistry
• Machine learning in drug discovery

Background:

• Accurate molecular property prediction is vital for drug design.
• Graph Neural Networks (GNNs) and Self-supervised Learning (SSL) are common but often miss long-range interactions.
• Weak, long-range interatomic interactions significantly influence molecular properties.

Purpose of the Study:

• To develop a novel SSL framework, VIBE-MPP, that incorporates weak interactions and 3D spatial information.
• To improve the accuracy of molecular property prediction for drug discovery.
• To address limitations of existing graph-based deep learning methods.

Main Methods:

• Introduced Virtual Bonding Enhanced Molecular Property Prediction (VIBE-MPP) framework.
• Utilized a Virtual Bonding Graph Neural Network (VBGNN) to create enhanced molecular graphs.
• Employed a Dual-level Self-supervised Boosted Pretraining (DSBP) with four pretext tasks.
• Incorporated virtual bonds to represent interactions within a 10 Å radius.

Main Results:

• VIBE-MPP demonstrated superior performance on 10 benchmark datasets for classification and regression tasks.
• Achieved an average improvement of 3.20% over state-of-the-art baseline models.
• Attained optimal performance on four regression datasets.
• Visualizations confirmed effective capture of molecular properties and semantic information.

Conclusions:

• VIBE-MPP effectively integrates weak interactions and 3D spatial data for enhanced molecular representation.
• The framework significantly advances molecular property prediction accuracy in drug design.
• This approach offers a promising direction for future deep learning applications in computational chemistry.