Unsupervised clustering is an essential tool for understanding the structure of large molecular libraries of drug-like molecules. Established pipelines to cluster these datasets rely on generating binary molecular fingerprints and separating them using the Taylor-Butina sphere exclusion algorithm based on Tanimoto similarity. Applying this workflow to libraries of hundreds of millions of molecules is infeasible, since it

scales quadratically (O(N²)) with dataset size. In this context, the BitBIRCH algorithm, a modification of the O(N) BIRCH clustering technique that extends it to binary features and exploits the iSIM framework for fast calculation of average Tanimoto similarities, offers a practical solution for fast clustering of large libraries. Here we present BitBIRCH-Lean, a memory- and compute-efficient implementation of the BitBIRCH algorithm. BitBIRCH-Lean is also parallelizable over groups of fingerprints, drawing on recently developed clustering refinement techniques to preserve representative cluster structure. We demonstrate how BitBIRCH-Lean enables clustering hundreds of millions of molecules on a single workstation, a task previously infeasible even on HPC clusters, while maintaining high cluster quality and control.

Kenneth Lopez Perez

University of Florida

Appearances