GDMol: Generative Double-Masking Self-Supervised Learning for Molecular Property Prediction.

Area of Science:

• Computational Chemistry
• Drug Discovery
• Machine Learning

Background:

• Effective molecular feature representation is vital for predicting drug properties.
• Graph neural networks (GNNs) pre-trained with self-supervised learning address limited labeled data in molecular property prediction.
• Traditional GNNs use single masking, limiting self-supervised training to local molecular information.

Purpose of the Study:

• To develop a novel generative double-masking self-supervised learning model for molecular property prediction.
• To improve molecular representation by capturing global information and semantic knowledge.
• To achieve more accurate and robust predictions in drug property analysis.

Main Methods:

• Proposing GDMol, a generative double-masking self-supervised learning framework.
• Integrating generative learning into self-supervised learning for latent representations.
• Applying a second masking round to latent representations to capture global molecular features.

Main Results:

• GDMol demonstrated superior performance in molecular property prediction across five diverse datasets.
• Analysis of gradient changes revealed the contribution of local structures to prediction outcomes, enhancing interpretability.
• The model provides targeted insights for optimizing drug molecules.

Conclusions:

• The research introduces novel insights for improving molecular property prediction.
• This work highlights the potential of generative and self-supervised learning in chemistry.
• It paves the way for future research in applying these methods to chemical tasks.