Molecular property prediction by semantic-invariant contrastive learning.

Area of Science:

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

Background:

• Contrastive learning is a key technique for self-supervised molecular representation learning.
• Existing methods using noise-adding for view generation can cause semantic inconsistency, leading to poor model performance.

Purpose of the Study:

• To develop a novel semantic-invariant view generation method for contrastive learning.
• To introduce the Fragment-based Semantic-Invariant Contrastive Learning (FraSICL) model for enhanced molecular property prediction.

Main Methods:

• Proposed a semantic-invariant view generation by decomposing molecular graphs into fragment pairs.
• Developed the FraSICL model with two branches for view representation generation.
• Incorporated multi-view fusion and an auxiliary similarity loss for improved information utilization.

Main Results:

• FraSICL achieves state-of-the-art performance on benchmark datasets for molecular property prediction.
• The model demonstrates superior results compared to existing methods, especially with limited pre-training data.
• The proposed semantic-invariant view generation addresses the limitations of noise-adding techniques.

Conclusions:

• FraSICL offers a robust and efficient approach for self-supervised molecular representation learning.
• The method effectively overcomes semantic inconsistency issues in contrastive learning for drug discovery.
• The FraSICL model shows significant potential for advancing AI-aided drug design and discovery.