ClusterX: a novel representation learning-based deep clustering framework for accurate visual inspection in virtual...

Area of Science:

• Computational chemistry
• Structural biology
• Bioinformatics

Background:

• Molecular clustering aids structure-based virtual screening.
• Traditional methods using fingerprints/descriptors lack receptor and interaction data, limiting accuracy.
• Accurate clustering is crucial for identifying potential drug candidates.

Purpose of the Study:

• To introduce ClusterX, a novel deep clustering framework for learning molecular representations of protein-ligand complexes.
• To improve the accuracy of ligand clustering in virtual screening by incorporating structural and interaction information.
• To provide a tool assisting computational chemists in visual decision-making during drug discovery.

Main Methods:

• Representing protein-ligand complexes as graphs.
• Utilizing joint optimization for learning cluster-friendly molecular features.
• Applying the framework to the KLIFs database and virtual screening datasets.

Main Results:

• ClusterX effectively distinguishes binding modes of kinase inhibitors.
• Clustering results on virtual screening data show ClusterX outperforms or matches traditional methods like SIFt and extended connectivity fingerprints.
• The framework learns representations that capture essential complex information for clustering.

Conclusions:

• ClusterX offers a novel approach to molecular clustering in virtual screening.
• The framework enhances the accuracy of identifying active hit compounds by considering protein-ligand interactions.
• ClusterX provides a valuable tool for computational medicinal chemists, aiding visual decision-making in drug discovery.