BitBIRCH: efficient clustering of large molecular libraries.

Area of Science:

• Computational chemistry
• Cheminformatics
• Machine learning applications

Background:

• Machine learning (ML) is crucial for analyzing large chemical datasets.
• Clustering is a key technique for exploring chemical space.
• Existing clustering methods struggle with the scale of modern molecular libraries.

Purpose of the Study:

• To introduce BitBIRCH, a novel, efficient clustering algorithm.
• To address the time and memory limitations of current methods for large-scale chemical data.
• To enable the analysis of billion-molecule datasets.

Main Methods:

• BitBIRCH employs a tree structure for O(N) time scaling, similar to BIRCH.
• It utilizes the instant similarity (iSIM) formalism for binary fingerprints and Tanimoto similarity.
• Parallel and iterative approximations are used for handling extremely large datasets.

Main Results:

• BitBIRCH demonstrates >1000x speed improvement over Taylor-Butina for 1.5M molecules.
• The algorithm achieves high efficiency without sacrificing cluster quality.
• Clustering of one billion molecules was completed in under 5 hours.

Conclusions:

• BitBIRCH offers a scalable and efficient solution for clustering massive molecular libraries.
• The method overcomes computational bottlenecks in chemical space analysis.
• It paves the way for analyzing unprecedentedly large chemical datasets using ML.