Initial GPU Optimization of Template Modeling Score

Abstract

Accuracy in the prediction of protein structures is key in understanding the biological functions of different proteins. Numerous measures of similarity tools for protein structures have been developed over the years, and these include Root Mean Square Deviation (RMSD), as well as Template Modeling Score (TM-score). While RMSD is influenced by the length of the protein and therefore the similarity between superimposed models can be affected by divergent loops in the models, TM-score is rather a robust and a more accurate method. TM-score, however, is much slower than RMSD in terms of calculations for the optimal superimposed model. Here, we present initial optimization work on GPU-TM-score, a GPU accelerated Template Modeling Score for fast and accurate measuring of similarity between protein structures. Our optimization is based on OpenACC parallelization and performance analysis of bottleneck regions and the KABSCH algorithm for the calculation of optimal superimposition within parallel architectures. Our initial results indicate an average 3.14× speedup compared to original TM-score on a benchmark set of 20 protein structures. This speedup is recorded on an Nvidia Volta V100 GPU compared to an AMD EPYC 7742 64-core processor.