BitBIRCH-Lean

BitBIRCH-Lean implements a fast, memory-efficient version of the Bit-BIRCH algorithm for high-throughput clustering of molecular libraries on typical workstations.

Key Features:

• Dynamic types and bit-packed fingerprints: These are incorporated inside the clustering tree to improve upon the original BitBIRCH implementation.
• Compressed data operations: Most operations are efficiently performed on compressed data.
• Optional C++ extensions: These accelerate bottleneck calculations, providing up to 2X speedup.
• Parallel multi-round variant: A parallel, multi-round variant of the Bit-BIRCH algorithm exploits efficiency gains to cluster hundreds of millions of molecules in minutes with no loss in cluster quality.

Scientific Applications:

• High-throughput clustering of molecular libraries: Enables clustering of up to billions of drug-like molecules on typical workstations.
• Large-scale chemical space exploration: Benchmark tests show it serves as an efficient alternative to GPU-accelerated methods for processing vast amounts of molecules.

Methodology:

BitBIRCH-Lean performs clustering using the Bit-BIRCH algorithm with dynamic types and bit-packed fingerprints stored inside the clustering tree; most operations are performed on compressed data, with optional C++ extensions accelerating bottleneck calculations.