ACGCN: Graph Convolutional Networks for Activity Cliff Prediction between Matched Molecular Pairs.

Area of Science:

• Medicinal Chemistry
• Computational Chemistry
• Drug Discovery

Background:

• Activity cliffs (AC) represent a critical challenge in drug-target interaction studies, defined by significant activity differences in structurally similar compounds.
• Understanding ACs is crucial for deciphering complex target protein properties and advancing drug discovery.
• ACs offer valuable insights for developing more potent therapeutic agents.

Purpose of the Study:

• To propose and evaluate graph convolutional networks for predicting activity cliffs.
• To introduce the Activity Cliff prediction using Graph Convolutional Networks (ACGCNs) models.
• To enhance the understanding of structure-activity relationships in drug discovery.

Main Methods:

• Development of graph convolutional network models (ACGCNs) for activity cliff prediction.
• Application of ACGCNs to predict activity cliffs across three popular target datasets: thrombin, Mu opioid receptor, and melanocortin receptor.
• Utilizing gradient-weighted class activation mapping for visualizing molecular graph activation weights.

Main Results:

• ACGCNs demonstrated superior performance compared to several existing methods in predicting activity cliffs.
• The models successfully predicted ACs for thrombin, Mu opioid receptor, and melanocortin receptor datasets.
• Gradient-weighted class activation mapping effectively visualized important substructures within molecular graphs.

Conclusions:

• Graph convolutional networks provide a powerful approach for predicting activity cliffs.
• ACGCNs offer a promising tool for drug discovery by identifying key structural features influencing drug activity.
• The visualization technique aids in understanding molecular docking and substructure importance.