Activity Cliff-Informed Contrastive Learning for Molecular Property Prediction.

Area of Science:

• Medicinal Chemistry
• Computational Chemistry
• Drug Discovery

Background:

• Quantitative structure-activity relationship (QSAR) modeling is vital for drug design.
• Graph neural networks (GNNs) excel at molecular activity prediction but often miss activity cliffs (ACs).
• ACs, where similar molecules have different bioactivities, pose challenges for current GNNs.

Purpose of the Study:

• To introduce AC-awareness (ACA) to enhance molecular representation learning for activity modeling.
• To develop ACANet, an AC-informed contrastive learning approach.
• To improve the sensitivity of GNNs to ACs in chemical compounds.

Main Methods:

• Developed AC-awareness (ACA) as an inductive bias for molecular representation learning.
• Implemented ACA by jointly optimizing latent space metric learning and target space task performance.
• Integrated ACANet, an AC-informed contrastive learning method, with existing GNN architectures.

Main Results:

• AC-informed molecular representations consistently outperformed standard models on 39 benchmark datasets.
• Demonstrated superior performance in both regression and classification tasks for bioactivity prediction.
• Showcased strong predictive capabilities for pharmacokinetic and safety-related molecular properties.

Conclusions:

• ACA significantly enhances molecular representation learning for activity prediction.
• ACANet provides a valuable tool for identifying and refining lead compounds in early drug discovery.
• Activity-informed molecular representations are crucial for effective virtual screening and drug development.