A fast deisotoping algorithm and its implementation in the MSFragger search engine

Datasets

Publicly available data was downloaded from Proteome Xchange²² and PRIDE²³ repositories (Table S1). “Closed” and “Open” searches used PXD001468 (Chick et al.³). Data is high pH-fractionated HEK293 digest analyzed on an Orbitrap Q Exactive, previously used for open search development. “Fusion Lumos” search used PXD004940 (Espadas et al.²⁴), which includes HeLa digest data acquired with a range of settings on an Orbitrap Fusion Lumos instrument. “Sciex” search used PXD005111 (Ammar et al.²⁵), containing Jurkat cell digest data collected on a Sciex TripleToF 6600+ instrument . “timsTOF” search used PXD010012 (Meier et al.²⁶) DDA-PASEF data acquired on a Bruker timsTOF Pro . “Glyco” search used PXD011533 (Riley et al.²⁷), which contains enriched N-glycoproteomics data from mouse brain tissue analyzed on an Orbitrap Fusion Lumos instrument. Finally, “peptiomics” search used neuropeptidomics data from mouse trigeminal ganglia acquired on an Orbitrap Fusion from PXD013662 (Anapindi et al.⁷).

MSFragger Searches

Deisotoping has been incorporated as an option in MSFragger starting with version 2.3. All searches described here were performed in a beta version of MSFragger (2.4-RC6-glyco) that included glycoproteomics search capabilities to enable searches of glyco data, but all beta capabilities have since been incorporated into the released version of MSFragger (starting with v3.0)²⁸. Full parameters for all MSFragger searches and Philosopher validation and filtering can be found in the supporting information. Parameter files are named to match the labels provided in figures and tables. Briefly, searches were performed as follows. The reviewed Human proteome from UniProt (downloaded 8/22/2019, 20464 sequences) was used with reversed decoys appended in Philosopher²⁹ for all searches except the mouse N-glycopeptide data of Riley et al. and mouse neuropeptide data of Anapindi et al.. For searching data from Anapindi et al., the reviewed mouse (Mus musculus) proteome from Uniprot (downloaded 9/24/2019, 17019 sequences) was used with decoys appended in Philosopher. For Riley et al., the glycoprotein-focused database with 3,574 sequences used by Riley et al. was used with reversed decoys appended in Philosopher. For all searches except the peptidomics (nonspecific) search of neuropeptide data, trypsin digestion with 1 or 2 missed cleavages was allowed with resulting peptide lengths of 7 to 50 amino acids. Variable modifications of Met oxidation (up to 2 per peptide), protein N-terminal acetylation and peptide N-terminal pyroglutamate formation (Gln, Cys) were allowed with no more than 3 total modifications per peptide. For closed searches, the precursor mass tolerance was 20 ppm and the fragment mass tolerance was 10 ppm (Orbitrap) or 20 ppm (TOF) with mass calibration enabled (without parameter optimization) and isotope off-by-X errors of 0/1/2 allowed. Spectra were square root transformed and had precursor peaks removed prior to analysis. For open searches, precursor mass tolerance was −150 to 500 Da, with the same fragment tolerances as closed searches, calibration enabled, no isotope errors, and localization of delta masses enabled³⁰. Each search was performed with and without deisotoping, with all other parameters identical.

For nonspecific searches of endogenous neuropeptides, no enzyme was specified and variable modifications of oxidation (Met), deamidation (peptide C-terminus), acetylation (protein N-terminus, Lys), pyroglutamate formation (N-terminal Gln), and water loss (N-terminal Glu) were allowed, with a maximum of 1 each and 3 total modifications per peptide. Peptide lengths of 7 to 70 amino acids were allowed with masses of 500 to 12,000 Da. Three matched ions were required for modeling and five matched ions were required for reporting to pepXML output file. Precursor and fragment mass tolerances were 20 and 25 ppm, no calibration was performed, and isotope errors of 0/1/2 were allowed. Since the number of candidate peptides exceeded the maximum possible within MSFragger (~ 2 billion), it was run using the split database method available within FragPipe.

For N-glycopeptide searches in the data of Riley et al., raw files were split into separate mzML files by activation type (HCD or AI-ETD). Up to 3 missed cleavages were allowed and a list of 182 glycan masses (identical to those used in Riley et al.) were set as N-glycan mass offsets. For HCD spectra, b, y, Y, b~, and y~ ions were considered (~ refers to the addition of one HexNAc, 203.0794 Da), without localization of delta masses. AI-ETD spectra considered b, y, c, z, and Y ions, used delta mass localization, and removed precursor and M + e⁻ peaks. Default Y and oxonium ion lists were used.

FDR filtering was performed with the Philosopher²⁹ pipeline, including PeptideProphet³¹ modeling of peptide probabilities, ProteinProphet³² protein inference, and Philosopher’s internal FDR filter. In PeptideProphet, closed searches used the accmass, ppm, nonparam, expectscore and decoyprobs options, while open searches used nonparam, expectscore, and decoyprobs options along with the extended mass model¹ (masswidth 1000) and clevel −2. All searches used ProteinProphet with default options except maxppmdiff increased to 20,000,000. All searches were filtered to 1% PSM and protein level FDR using the razor peptides method with a sequential filtering step to remove any PSMs from proteins that did not pass FDR.

Comparison with other deisotoping tools

Deisotoping speed was compared by analyzing the Chick et al. dataset with MSFragger, MS2Deisotoper¹⁰, and HardKlor^{18, 33}. Default parameters were used for MS2Deisotoper and MSFragger. HardKlor parameters were default except only MS2 spectra were deisotoped, data was pre-centroided, and instrument was set to Orbitrap, resolution 17500, to match the input data. All comparisons were performed on a linux server with 28 cores (56 threads) and 500 GB RAM. As MSFragger is multithreaded and the other tools are not, tests were also performed with the server restricted to 8 and 1 thread(s) for comparison to performance on typical computer systems and with the same resources as the other tools, respectively.

Quality of deisotoping results was compared by performing closed and open database searches on the output of MSFragger, MS2Deisotoper, HardKlor, and MSConvert deisotoping methods on the Chick et al. data. MSFragger search parameters were as described above. MSFragger and MS2Deisotoper deisotoping parameters were left at defaults. Numerous parameters were tested for HardKlor, with the results from the best parameters found reported. Correlation was reduced to 0.1, depth increased to 5, and max features increased to 25. MSConvert deisotoping was performed with the Poisson method with a maximum charge of 6. MS2Deisotoper generates MGF files as output, which were searched directly in MSFragger with deisotoping disabled and all other parameters the same as in searching the MSFragger deisotoped data. MSConvert was set to output mzML files, which were searched directly with MSFragger. HardKlor reports results in a non-standard text format, which was converted to MGF with a Python script. Reported peaks were decharged and converted to M-H⁺ using the charge information provided, and MSFragger was set to search only 1⁺ fragment peaks (as is done for MSFragger-deisotoped data).

Development of the deisotoping algorithm

Our algorithm began as a reimplementation of the ProteoWizard MSConvert^{34, 35} deisotoping algorithm¹⁹, after we noticed some improvement when using MSConvert to do deisotoping prior to performing standard searches with MSFragger. When using MSConvert as an external tool, however, the charge of the deisotoped MS2 peaks cannot be communicated to MSFragger, making an internal deisotoping algorithm preferable. We developed a parallelized modified algorithm within MSFragger that consists of the following steps applied to centroided spectra (Fig. 1):

Evaluating the impact of deisotoping across database search methods

Deisotoping removes uninformative signals from the experimental spectrum, reducing the change of random matches to fragment ions from incorrect peptide(s) and improving the quality of search results. However, deisotoping methods do not always determine the correct monoisotopic peak, e.g. due to overlapping clusters of signals or deviations from the expected averagine distribution, which could result in the incorrect removal of relevant peaks. Our algorithm is capable of disentangling moderately overlapped distributions (see Fig. S1 for examples). Furthermore, peaks with no neighbors (“singleton” peaks) cannot have their charge state and monoisotopic peak determined due to insufficient information. For peaks of unknown isotope/charge, we tested several methods: removal of all unknown peaks, assuming all unknowns are singly charged, and searching unknowns as both singly and doubly charged. These peaks can still represent useful peptide fragmentation signals if the remainder of the isotopic peaks are filtered out, for example in data from Orbitrap mass spectrometers as a result of the noise suppression and other signal processing performed during the conversion from raw transients to Fourier transformed mass spectra. As a result, removal of all unknowns proved too stringent of a filter, reducing the output PSM count in all conditions tested, so it was removed as an option and not included in the final data analysis. We designed our deisotoping method to be tolerant of these caveats, and only remove peaks when a confident match is made to an isotopic cluster.

We performed a variety of searches on data from several instrument types and raw formats to characterize the performance of our deisotoping method. In initial tests, we compared three conditions: no deisotoping (“d0”), deisotoping with peaks of unknown charge assumed to be singly-charged (“d1”), and deisotoping with unknown peaks considered as both singly and double charged (“d2”). For the d0 (no deisotoping) case, searches were performed with a max fragment charge of 2, meaning that all observed peaks were searched as both singly and doubly charged. Effectively, this means that the d0 and d2 methods function identically for unknown peaks, but d2 removes isotopic peaks from any confidently identified isotopic clusters and only searches them at the identified charge state (including charge states higher than 2⁺). The d2 method performed worse than d1 in all cases tested, and thus we chose to present the results of d1 and d0 only, simplifying the comparison to with and without deisotoping, respectively.

We compared the speed and degree of improvement afforded to proteomic database searches of our method with several freely available deisotoping tools. The key distinguishing feature of our method is its speed, deisotoping the complete Chick et al. dataset in a few seconds, 100 to over 1000 times faster than the other tools tested (Table S2). This speed comes from reducing the computational complexity of the deisotoping problem by using a simple Poisson approximation for the averagine distribution (as previously described¹⁶), optimizing the implementation of the cluster evaluation for speed, and parallelizing the implementation to take advantage of the multiple processors available on modern computers. In evaluating the accuracy of the deisotoping, we optimized specifically for improving the quality of proteomics search results (including leaving singleton peaks in spectra). Comparing against the tools used in speed testing as well as MSConvert, searches performed on data deisotoped with our method yield slightly more PSMs than when deisotoping with other tools (Table S3). Obtaining quality search results from extremely fast deisotoping made it possible to seamlessly integrate deisotoping into MSFragger searches, allowing us to evaluate utility of deisotoping in a variety of search and data types.

For “closed” searches, i.e. fully enzymatically digested peptides with few modifications, deisotoping resulted in minimal change to PSM counts (< 1% improvement) with a modest increase in speed (~10%) (Fig. 2). This was not unexpected, as the density of theoretical fragment masses is sufficiently low in a closed search for the extra spectral peaks included when forgoing deisotoping to be unlikely to match to incorrect peptides frequently enough to generate high-scoring false positives and reduce the sensitivity of the search. This is in contrast to results from MS2Deisotoper¹⁰, which increased PSMs found by 3.5% for a typical closed search. MSFragger performs several spectral processing steps, including mass calibration to enable the use of a narrow fragment mass tolerance and considering only the top N peaks in each spectrum (150 by default). Removing all of these processing options increased the gain in PSMs when performing deisotoping to 3.1% (Table S4), close to that obtained by MS2Deisotoper, indicating that the difference is largely a result of different spectral processing. . An open search of the Chick et al. data with MSFragger, however, obtained nearly 25,000 more PSMs with deisotoping than without deisotoping, while reducing the search time by approx. 20% (Table 1). The search space for open searches is very large, requiring a significant amount of time to search each spectral peak and greatly increasing the odds of random matches. By removing peaks from consideration through deisotoping and decharging, we increase the speed and reduce the number of high scoring decoy matches, ultimately improving the sensitivity of the search. The process of deisotoping itself adds a negligible amount of time to the search, taking just fractions of a second per raw file, or a few seconds total, for each search with deisotoping (Table 1).

This trend holds for other complex search types, including a glycoproteomics search considering hundreds of possible glycan compositions and a nonspecific search of endogenous neuropeptide data. The glycopeptide search conducted on the data of Riley et al. resulted in annotation of 12.7% more PSMs with deisotoping (Fig. 2), and a speed improvement of over 60%. The increase in glycoPSMs translated to the identification of several hundred additional glycopeptides and just over one hundred additional glycoproteins (Table 1). The data in question includes HCD and AI-ETD scans of glycopeptides, which often contain more fragment ions and more high mass fragments with many isotopic peaks than is typical for fragmentation of unmodified tryptic peptides due to the large size and facile fragmentation of the glycan groups. As a result, deisotoping is able to simplify the spectra to a greater degree, resulting in greatly improved speed. Deisotoping is also particularly promising for searches of peptidomic and immunopeptidomic data given both the large size of undigested endogenous peptides and the massive search space involved in nonspecific database search. Our nonspecific search of endogenous neuropeptide data resulted in the largest gain in PSM counts compared to no deisotoping, obtaining 37% more PSMs than an identical search without deisotoping (Table 1). There was minimal improvement in search speed for this data with deisotoping (Fig. 2), largely because the vast majority of the search time is spent preparing the fragment indices (due to the large database size and resulting requirement to split into several smaller searches), rather than searching peaks in each spectrum. Each raw file was deisotoped for each sub-search, resulting in each being deisotoped (redundantly) many times, but the total deisotoping time required was still only 0.5% of the total search time due to the speed of the deisotoping algorithm.

In comparing across search types, we found that deisotoping resulted in improved speed and search performance in all cases, with the degree of improvement scaling with the complexity of the search space. In the glycoproteomics and peptidomics data, which contain much larger precursors on average than typical tryptic digests (Table S5), we observed greatly improved search performance, which we attribute to the increased presence of larger fragment ions with more isotopic peaks resulting from larger precursors, in addition to very large search spaces. All the data compared in these analyses was collected on Orbitrap mass spectrometers, which, given the effects of spectral preprocessing (and thus complexity) on performance, made us curious about the impact of deisotoping on data from other instruments. We compared results from identical parameter open searches with and without deisotoping on datasets from a Sciex 6600+ QToF instrument and a Bruker timsTOF Pro IM-QToF instrument acquired in PASEF³⁶ mode to data from Orbitrap Fusion Lumos and Q Exactive instruments (Table 2). In data from all instruments, open search time was improved between 20-40% with deisotoping, despite large differences in the complexity of the various datasets and total search time required (Fig. 3). Interestingly, the improvements to the resulting PSM count varied considerably across the datasets, from 0.5% in the Espadas et al. Orbitrap Fusion Lumos data to 13.5% in the Sciex 6600+ data (Fig. 3). The 6600+ dataset had many more peaks per spectrum than other data analyzed, particularly including many low-intensity peaks that appeared to be a result of noise (Figs. S2–S4), meaning that decreasing spectral complexity through deisotoping tended to have a larger impact.

These results demonstrate that while deisotoping nearly always resulted in improved PSM output and reduced search time, the degree of those improvements ranges considerably with the complexity and quality of the input data as well as the size of the search space. Raw data with more noise, such as the Imhof et al. dataset, was generally improved more than the data with fewer peaks per spectrum, but there was considerable variation within each instrument type. It is likely that the degree of improvement from deisotoping will vary as much or more as a function of individual data sets and instrument parameters than as a result of instrument type, underscoring the importance of optimal tuning and operation of any MS instrument for proteomics.