Triple Generative Self-Supervised Learning Method for Molecular Property Prediction.

Area of Science:

• Computational chemistry
• Machine learning in drug discovery

Background:

• Molecular property prediction is crucial for drug discovery.
• Self-supervised learning can leverage unlabeled data to improve prediction performance.

Purpose of the Study:

• To propose a novel triple generative self-supervised learning (TGSS) method for molecular property prediction.
• To enhance feature extraction and model interpretability in molecular property prediction.

Main Methods:

• Employed three encoders: BiLSTM, Transformer, and GAT for pre-training.
• Utilized a variational autoencoder (VAE) for feature reconstruction.
• Incorporated a feature fusion module and atomic similarity heat maps for interpretability.

Main Results:

• The TGSS method demonstrated accuracy on chemical and biological benchmark datasets.
• Comparative experiments validated the effectiveness of the proposed approach.
• Atomic similarity heat maps provided insights into feature extraction rationality.

Conclusions:

• The proposed TGSS method offers an effective approach for molecular property prediction.
• The integration of multiple encoders and feature fusion enhances predictive performance.
• The method improves model interpretability, aiding in understanding molecular features.