Toward an improved clustering of large data sets using maximum common substructures and topological fingerprints

Area of Science:

• Computational chemistry
• Cheminformatics
• Data mining

Background:

• Analyzing large chemical datasets is crucial for drug discovery.
• Existing methods struggle with scalability and accurate chemotype identification.
• Understanding structure-activity relationships (SAR) requires robust molecular grouping.

Purpose of the Study:

• To develop a scalable algorithm for clustering large datasets based on chemotypes.
• To enable improved analysis of high-throughput screening (HTS) and virtual screening (VS) data.
• To facilitate insights into structure-activity relationships and chemotype hopping.

Main Methods:

• Hierarchical k-means algorithm applied to molecular fingerprints for preclustering.
• Maximum common substructure (MCS) approach for chemotype extraction from terminal clusters.
• Iterative fusion of similar chemotypes and singletons, followed by overlap-based grouping.
• Second round of hierarchical k-means using chemotype representatives for final hierarchical grouping.

Main Results:

• Successful grouping of large molecular datasets (>100,000 molecules) by chemotype.
• Identification of chemotypes based on shared structural features (rings, atoms, heavy atoms).
• Demonstration of utility in analyzing reverse transcriptase inhibitors and evaluating similarity searching routines.
• Interactive graphical user interface for visualizing SAR and chemotype hopping potential.

Conclusions:

• The developed algorithm provides a high-quality, scalable solution for analyzing large chemical datasets.
• It enhances the analysis of HTS and VS results, aiding in drug discovery and lead optimization.
• The method effectively identifies chemotypes and supports the exploration of chemotype hopping strategies.