Prototype-based contrastive substructure identification for molecular property prediction.

Area of Science:

• Computational chemistry
• Machine learning
• Graph representation learning

Background:

• Substructure-based methods are crucial for molecular property prediction (MPP).
• Current MPP approaches often rely on predefined substructure rules, limiting adaptability.
• There is a need for methods that can autonomously identify relevant substructures for MPP tasks.

Purpose of the Study:

• To propose Prototype-based cOntrastive Substructure IdentificaTion (POSIT), a self-supervised framework for adaptive substructure discovery.
• To guide end-to-end molecular fragmentation by learning substructural prototypes.
• To enhance molecular representations for MPP tasks.

Main Methods:

• POSIT employs a self-supervised pre-training strategy with a soft connectivity constraint for topological meaningfulness.
• A prototype-substructure contrastive clustering objective aligns substructures with prototypes based on attribute similarity.
• A cross-scale attention mechanism integrates substructure information during fine-tuning.

Main Results:

• Experiments on diverse datasets demonstrate POSIT's effectiveness in both classification and regression MPP tasks.
• Visualization analysis confirms that identified substructures align with chemical priors.
• The framework successfully guides end-to-end molecular fragmentation.

Conclusions:

• POSIT offers an adaptive and autonomous approach to substructure identification for molecular property prediction.
• The framework improves molecular representations by effectively integrating substructure-level information.
• POSIT represents a significant advancement in self-supervised learning for cheminformatics.