GPU accelerated chemical similarity calculation for compound library comparison.

Area of Science:

• Computational chemistry
• Bioinformatics
• High-performance computing

Background:

• Chemical similarity calculations are crucial for drug discovery processes like virtual screening and lead optimization.
• Existing methods for calculating Tanimoto similarity can be computationally intensive, especially for large compound libraries.

Purpose of the Study:

• To develop a novel, highly efficient GPU-accelerated algorithm for all-vs-all Tanimoto matrix calculation and nearest neighbor search.
• To demonstrate the performance advantages of the new algorithm over existing CPU-based and GPU-accelerated methods.

Main Methods:

• Implementation of a GPU-accelerated algorithm leveraging multicore GPU architecture and CUDA parallel programming.
• Utilizing intrinsic GPU instructions for optimized Tanimoto coefficient calculations with Unity fingerprints.
• Benchmarking the algorithm's performance against commercial CPU software and existing GPU-accelerated sparse vector algorithms.

Main Results:

• The novel GPU algorithm achieves up to 39 times superior performance compared to commercial CPU software.
• The approach is nearly 10 times faster than existing GPU-accelerated sparse vector algorithms for Tanimoto calculations using Unity fingerprints.
• Calculation of 324 billion Tanimoto coefficients between 32 million PubChem compounds and 10,000 Active Probes compounds was completed in approximately 20 minutes on a 128-CUDA-core GPU.

Conclusions:

• The developed GPU-accelerated algorithm offers a significant speedup for large-scale chemical similarity computations.
• This advancement can greatly benefit compound library design, virtual screening, and lead optimization in drug discovery.
• The method demonstrates the power of GPU computing for accelerating complex chemoinformatics tasks.