Self-supervised learning with chemistry-aware fragmentation for effective molecular property prediction.

Area of Science:

• Computational Chemistry
• Artificial Intelligence
• Drug Discovery

Background:

• Molecular property prediction (MPP) is vital for AI-aided drug discovery (AIDD).
• Self-supervised learning (SSL) shows promise for MPP but often overlooks critical molecular substructures.
• Existing SSL models struggle to effectively utilize substructure information for improved molecular representations.

Purpose of the Study:

• To develop a novel self-supervised learning framework, CAFE-MPP, that explicitly incorporates chemistry-aware substructures for enhanced molecular representations.
• To improve the performance of molecular property prediction by focusing on fragment-level representations.
• To address the data scarcity challenge in AIDD through effective representation learning.

Main Methods:

• Introduced a novel fragment-based molecular graph (FMG) to capture topological relationships between substructures.
• Employed a self-supervised contrastive learning framework at the fragment level with hard negative pairs for pre-training.
• Utilized a Graphormer model to generate final molecular representations from fragment embeddings for MPP.

Main Results:

• CAFE-MPP achieved state-of-the-art performance on 7 out of 11 benchmark datasets for MPP.
• Demonstrated superior performance compared to six other prominent self-supervised methods.
• Showcased that CAFE-MPP learns representations that implicitly encode crucial fragment information relevant to molecular properties.

Conclusions:

• CAFE-MPP effectively learns chemistry-aware molecular representations, significantly advancing MPP in AIDD.
• The fragment-level contrastive learning approach enhances predictive performance and addresses limitations of existing SSL methods.
• CAFE-MPP offers a promising direction for developing more accurate and data-efficient models in drug discovery.