Large-scale proteomics projects often generate massive and highly redundant tandem mass spectra. Spectral clustering algorithms can reduce the redundancy in these data sets and thus speed up database searching for peptide identification, a major bottleneck for proteomic data analysis. The key challenge of spectral clustering is to reduce the redundancy in the MS/MS spectra data while retaining sufficient sensitivity to identify peptides from the clustered spectra. We present the software msCRUSH, which implements a novel spectral clustering algorithm based on the locality sensitive hashing technique. When tested on a large-scale proteomic data set consisting of 23.6 million spectra (including 14.4 million spectra of charge 2+), msCRUSH runs 6.9-11.3 times faster than the state-of-the-art spectral clustering software, PRIDE Cluster, while achieving higher clustering sensitivity and comparable accuracy. Using the consensus spectra reported by msCRUSH, commonly used spectra search engines MSGF+ and Mascot can identify 3 and 1% more unique peptides, respectively, compared with the identification results from the raw MS/MS spectra at the same false discovery rate (1% FDR) of peptide level. msCRUSH is implemented in C++ and is released as open-source software.
Keywords: MS/MS spectra; algorithm; identification; locality sensitive hashing; mass spectrometry; optimization; peptide; proteomics; similarity; spectral clustering.