Molecular property prediction by contrastive learning with attention-guided positive sample selection.

Area of Science:

• Computational chemistry
• Machine learning in drug discovery

Background:

• Molecular property prediction (MPP) is crucial for drug design.
• Self-supervised learning (SSL), particularly contrastive learning (CL), shows promise for feature learning.
• Effective positive sample selection is a key challenge in CL.

Purpose of the Study:

• To develop an improved CL method for MPP.
• To address the challenge of positive sample selection in CL.

Main Methods:

• Propose Contrastive Learning with Attention-guided Positive-sample Selection (CLAPS).
• Utilize an attention-guided scheme for positive sample generation.
• Employ a Transformer encoder for feature extraction and contrastive loss computation.

Main Results:

• CLAPS demonstrates superior performance in molecular property prediction.
• The proposed method outperforms current state-of-the-art (SOTA) approaches on benchmark datasets.
• Attention-guided positive sample selection significantly impacts CL performance.

Conclusions:

• CLAPS offers a novel and effective approach for molecular property prediction.
• The method advances the application of SSL in cheminformatics.
• Publicly available code facilitates further research and application.