MolCL-SP: a multimodal contrastive learning framework with non-overlapping substructure perturbations for molecular...

Area of Science:

• Computational chemistry
• Machine learning
• Drug discovery

Background:

• Molecular property prediction is crucial for drug discovery but faces challenges in molecular representation.
• Existing methods suffer from representation bottlenecks (single-modal) or redundant information (multimodal).
• Current data augmentation neglects intrinsic atomic dependencies within molecular substructures.

Purpose of the Study:

• To develop a substructure-aware multimodal contrastive learning framework for molecular property prediction.
• To address limitations in representation and data augmentation for improved molecular machine learning.
• To enhance the interpretability and generalization capabilities of molecular property prediction models.

Main Methods:

• Proposed MolCL-SP, a multimodal contrastive learning framework integrating three modalities via a Transformer-based encoder.
• Implemented modality-specific reconstruction for organic alignment and fusion of cross-modal information.
• Introduced a novel substructure-based non-overlapping perturbation strategy for data augmentation.

Main Results:

• MolCL-SP achieved state-of-the-art performance on 2D and 3D molecular property prediction benchmarks.
• Demonstrated strong generalization capabilities on drug-drug interaction prediction tasks.
• Visualization revealed effective capture of discriminative molecular embeddings and emphasis on chemically meaningful substructures.

Conclusions:

• MolCL-SP offers a robust framework for molecular property prediction, outperforming existing methods.
• The substructure-aware approach enhances model interpretability and captures chemically relevant features.
• This work advances multimodal contrastive learning for molecular machine learning applications.