A Novel Multi-Perspective Framework for Molecule Pretraining: From Atom to Motif Views.

Area of Science:

• Computational chemistry
• Drug discovery
• Machine learning

Background:

• Predicting molecular properties is crucial for drug discovery but hindered by costly experiments and limited labeled data.
• Self-supervised pretraining on unlabeled data can reduce annotation dependency.
• Existing methods often fail to retain domain-specific chemical knowledge, particularly critical substructures like motifs, and struggle with multiscale information integration.

Purpose of the Study:

• To introduce A2M-Mol, a multi-perspective molecular pretraining framework designed to overcome the limitations of current methods.
• To effectively integrate atom-level and motif-level chemical information for improved molecular property prediction.
• To enhance the preservation of chemical knowledge and interpretability in self-supervised molecular pretraining.

Main Methods:

• Developed A2M-Mol, a framework utilizing four parallel graph constructions for atom-level and motif-level views.
• Implemented cross-view alignment and multiscale fusion to explicitly encode chemical knowledge.
• Employed tailored contrastive learning coordinating self-supervised tasks: cross-view correspondence, atomic reconstruction, global topology modeling, and property constraint enforcement.

Main Results:

• A2M-Mol demonstrated consistent improvements over state-of-the-art methods across various benchmarks and backbone architectures.
• Ablation studies confirmed significant synergistic effects among the proposed self-supervised tasks.
• The framework maintained robust predictive accuracy across different data scales, validating its effectiveness.

Conclusions:

• A2M-Mol offers a robust and effective approach to molecular property prediction by integrating multiscale chemical information.
• The framework successfully preserves critical chemical knowledge and enhances interpretability.
• A2M-Mol shows significant potential to accelerate drug discovery by improving the efficiency and accuracy of molecular property prediction.