Exploiting Thread-Level and Instruction-Level Parallelism to Cluster Mass Spectrometry Data using Multicore Architectures.

Modern mass spectrometers can produce large numbers of peptide spectra from complex biological samples in a short time. A substantial amount of redundancy is observed in these data sets from peptides that may get selected multiple times in Liquid Chromatography Tandem Mass Spectrometry (LC-MS/MS) experiments. A large number of spectra do not get mapped to specific peptide sequences due to low signal-to-noise (S/N) ratio of the spectra from these machines. Clustering is one way to mitigate the problems of these complex mass spectrometry data sets. Recently we presented a graph theoretic framework, known as CAMS, for clustering of large-scale mass spectrometry data. CAMS utilized a novel metric to exploit the spatial patterns in the mass spectrometry peaks which allowed highly accurate clustering results. However, comparison of each spectrum with every other spectrum makes the clustering problem computationally inefficient. In this paper we present a parallel algorithm, called P-CAMS, that uses thread-level and instruction-level parallelism on multicore architectures to substantially decrease running times. P-CAMS relies on intelligent matrix completion to reduce the number of comparisons, threads to run on each core and Single Instruction Multiple Data (SIMD) paradigm inside each thread to exploit massive parallelism on multicore architectures. A carefully crafted load-balanced scheme that uses spatial locations of the mass spectrometry peaks mapped to nearest level cache and core allows super-linear speedups. We study the scalability of the algorithm with a wide variety of mass spectrometry data and variation in architecture specific parameters. The results show that SIMD style data parallelism combined with thread-level parallelism for multicore architectures is a powerful combination that allows substantial reduction in runtimes even for all-to-all comparison algorithms. The quality assessment is performed using real-world data set and is shown to be consistent with the serial version of the same algorithm.