A proteomics search algorithm specifically designed for high-resolution tandem mass spectra

Abstract

The acquisition of high-resolution tandem mass spectra (MS/MS) is becoming more prevalent in proteomics, but most researchers employ peptide identification algorithms that were designed prior to this development. Here, we demonstrate new software, Morpheus, designed specifically for high-mass accuracy data, based on a simple score that is little more than the number of matching products. For a diverse collection of data sets from a variety of organisms (E. coli, yeast, human) acquired on a variety of instruments (quadrupole-time-of-flight, ion trap-orbitrap, and quadrupole-orbitrap) in different laboratories, Morpheus gives more spectrum, peptide, and protein identifications at a 1% false discovery rate (FDR) than Mascot, Open Mass Spectrometry Search Algorithm (OMSSA), and Sequest. Additionally, Morpheus is 1.5 to 4.6 times faster, depending on the data set, than the next fastest algorithm, OMSSA. Morpheus was developed in C# .NET and is available free and open source under a permissive license.

Figure 1

An example spectrum, scan number 12,864 of the second replicate of the Q–OT human dataset. The Morpheus score is merely the sum of the number of matching products (10) and the fraction of abundance matched (30.2%), for a total of 10.302. This score, for the peptide of sequence NGVPAVGLK, was the best among all the peptides within the precursor mass tolerance.

Figure 2

Comparison of PSM, distinct peptide, and protein group identifications at 1% FDR with Mascot, OMSSA, Sequest, ZCore, and Morpheus for (a) Q–TOF E. coli, (b) dcLIT–OT yeast, (c) dcLIT–OT human, (d) Q–OT human, (e) LIT–OT ETD human, and (f) dcLIT–OT/IT yeast datasets. For all five high–mass accuracy datasets, Morpheus is the highest in all three quantities.

Figure 3

Examination of the impact of abundance filtering (AF), charge state assignment (CSA), and de-isotoping (DI) of MS/MS spectra on the performance of Morpheus for (a) Q–TOF E. coli, (b) dcLIT–OT yeast, (c) dcLIT–OT human, (d) Q–OT human, (e) LIT–OT ETD human, and (f) dcLIT–OT/IT yeast datasets. Surprisingly, aside from abundance filtering to retain the top 400 peaks–the most basic preprocessing–these features do not prove essential, except in the Q–OT and dcLIT–OT/IT datasets. This leaves scoring as the primary explanation for the exceptional performance of Morpheus.

Figure 4

Venn diagrams for Mascot, OMSSA, Sequest, and Morpheus for the second replicate of the Q–OT human dataset at the (a) PSM, (b) distinct peptide, and (c) protein group levels. The massive overlap of all four algorithms underscores the reliability of the results, although Morpheus had the most identifications unique to it for all three metrics.

Figure 5

Pseudo-ROC curves (identifications versus FDR) for Mascot, OMSSA, Sequest, and Morpheus for the second replicate of the Q–OT human dataset at the (a) PSM, (b) distinct peptide, and (c) protein group levels. The dashed line indicates the 1% FDR threshold, at which Morpheus had the most identifications for all three metrics. The oscillatory behavior of the Morpheus curves in (a) and (b) is due to the simple score, but does not seem to have any practical negative impact.

Figure 6

Speed comparison of OMSSA and Morpheus. Both algorithms were specified to use 8 threads, the maximum practical for the computer used which had dual quad-core processors. Note the y-axis is base-10 logarithmic. Morpheus ranges from 1.5 to 4.6 times faster than OMSSA, with the largest increases for large databases and large datasets (i.e. high number of MS/MS spectra).