A Hierarchical MS²/MS³ Database Search Algorithm for Automated Analysis of Phosphopeptide Tandem Mass Spectra

Abstract

A novel hierarchical MS²/MS³ database search algorithm has been developed to analyze MS²/MS³ phosphopeptides proteomic data. The algorithm is incorporated in an automated database search program, MassMatrix. The algorithm matches experimental MS² spectra against a supplied protein database to determine candidate peptide matches. It then matches the corresponding experimental MS³ spectra against those candidate peptide matches. The MS² and MS³ spectra are used in concert to arrive at peptide matches with overall higher confidence rather than combining MS² and MS³ data searched separately. Receiver operating characteristic analysis showed that hierarchical MS²/MS³ database searches with MassMatrix had better sensitivity and specificity than the two-stage MS²/MS³ database searches obtained with MassMatrix, Mascot and X!Tandem. A greater number of true peptide matches at a given false rate were identified by use of this new algorithm for data collected on both LCQ and LTQ-FTICR mass spectrometers. The additional MS³ spectral data also improved the overall reliability and the number of true positives due to the fact that the true positives of the MS²/MS³ search results had higher scores than those of the MS².

1 INTRODUCTION

Tandem mass spectrometry has been widely used in protein identification and characterization. In tandem mass spectrometry, the MS/MS or MS² spectra produced by fragmentation of peptide ions contain product ion signatures that can be used to sequence the peptides and characterize their post-translational modifications (PTMs).[1] However, phosphopeptide MS² spectra often do not contain sufficient sequencing information to identify the peptide. Poor sequencing of phosphopeptide ions is due to the labile nature of the phosphate group resulting in MS² spectra that are dominated by the neutral loss of the phosphate moiety. As a result, phosphopeptides are not identified as reliably as non-phosphorylated peptides in LC-MS/MS experiments. To overcome the shortcomings of MS² experiments for phosphopeptides, a third stage of mass spectrometry (MS³) can be performed on ions in the MS² scans resulting from the neutral loss of phosphate. These fragment ions produce MS³ spectra with sufficient fragment ions to not only identify the peptide but often determine the site of phosphorylation.[2]

There are several de novo sequencing-based algorithms that have been developed for analysis of MS² and MS³ spectral data.[3, 4] However, these algorithms can not be used in high-throughput data analysis due to their high computational expenses. Therefore, the analysis of MS²/MS³ experimental data for phosphopeptides relies primarily on database search algorithms, such as Mascot, SEQUEST, X!Tandem and OMSSA.[5–9] Often data analysis by database search programs is performed in two stages (Figure 1). In the first stage, MS² spectral data are searched against a supplied protein database to obtain a set of peptide and protein identifications for the MS² data. In the second stage, MS³ spectral data are also searched against the same protein database to obtain an additional set of peptide and protein identifications. Integration of these two sets of results is necessary to give the overall protein and peptide identifications. Ulintz et al has recently published an algorithm to integrate scores for the two matches of a set of MS² and MS³ spectra.[2]

Figure 1.

Diagram of two-stage MS²/MS³ database search process. The MS² and MS³ data are searched in two parallel independent database searches. In the first database search stage, MS² spectral data are searched against a protein database to give a set of peptide and protein identifications for the MS² data. In the second stage, MS³ spectral data are searched against the same protein database to give the other set of peptide and protein identifications for the MS³ data. Peptide matches for the MS² and MS³ data are then merged to give the final list of peptide matches for the data set by use of software tools.

Overall the integrated two-stage approach results in better results than analysis of MS² and MS³ data separately. However, this approach does not take advantage of the inherent hierarchical nature of MS² and MS³ data. In the two-stage search process, the fact that MS² and MS³ spectra are created from the same peptide precursor ion following two consecutive fragmentations is ignored by the database search algorithm. Therefore, the MS² and MS³ spectra for the same precursor may result in different peptide matches. Furthermore, for typical data sets collected on high mass accuracy capable mass spectrometers, the MS² precursor ions are measured at high mass accuracy and the MS³ precursor ions are often measured at a much lower mass accuracy. In this case, MS³ spectral data do not fully exploit the benefit from the high mass accuracy of the original precursor ions.

In this manuscript we describe a novel algorithm for performing hierarchical MS²/MS³ database searches. This algorithm performs MS²/MS³ pattern matching analysis and returns peptide/protein identifications for each set of MS²/MS³ spectra. This approach does not use post-search merging of results from the MS² and MS³ data that can lead to confounding peptide and protein matches. The algorithm first searches MS² spectral data against a supplied protein database, and then searches the associated MS³ spectral data against candidate peptide matches obtained in the prior MS² search. In this manner, MS² and MS³ data are used in concert to arrive at peptide identifications. The MS²/MS³ search algorithm described herein takes full advantage of the hierarchical nature of the MS²/MS³ data. The hierarchical search process eliminates the discrepancy between the MS² peptide matches and MS³ peptide matches that may occur in the two-stage search process. Furthermore, the high mass accuracy of the precursor ion for the MS² experiment can be inherited by the MS³ data analysis in the hierarchical search algorithm resulted in overall improved confidence in peptide and protein identifications.

2 MATERIALS AND METHODS

2.1 Sample Preparation and Mass Spectrometry

α-Casein from bovine milk was purchased from Sigma-Aldrich (St. Louis, MO). The α-Casein was digested by trypsin in 25 mM ammonium biocarbonate buffer (pH = 8.0) at 37 °C for 1 hour. Enzymes were used in 50:1 ratio (substrate:enzyme). The tryptic digests were then dried and dissolved in HPLC water with 0.1% formic acid (pH = 3.0) to a final concentration of 1.0 μg/μl. The phosphopeptides in the solution were then enriched by use of a zirconium dioxide coated NuTips (Glugen Corp., Columbia, MD) as described by Kweon and Hakansson.[10] The resulted peptides were identified by use of data-dependent LC-MS³ on a LCQ Deca XP ion trap and a LTQ-FTICR mass spectrometer (Thermo Fisher, San Jose, CA, USA). 2.0 μL of enriched peptides with a total concentration of 1.0 μg/μL before enrichment was injected into the LC-MS system and eluted off the capillary HPLC column into the mass spectrometer with a linear gradient of 5% – 50% of mobile phase B over 28 minutes at a overall flow rate of ~250 nL/min. Solvent A was water with 0.1% formic acid and solvent B was acetenitrile with 0.1% formic acid. Ions were fragmented by use of collision induced dissociation (CID). The MS³ scan was targeted at phosphorylation neutral loss ions in the MS² scan with a mass difference of 98.0 Da, 80.0 Da, 49.0 Da, 40 Da, 32.7 Da or 26.7 Da from the precursor mass.

2.2 Database Search and Search Parameters

The .RAW data files obtained from the LCQ Deca XP ion trap and LTQ-FTICR mass spectrometers were converted to mzXML files by use of ReAdW (http://tools.proteomecenter.org/ReAdW.php). For low mass accuracy data collected on the LCQ Deca XP ion trap mass spectrometer, LC-MS/MS spectra that were not derived from singly charged precursor ions were extracted as both doubly and triply charged precursors. For high mass accuracy data collected on the LTQ-FTICR mass spectrometer, isotope distributions for the precursor ions of the MS² spectra were deconvoluted to obtain the charges and monoisotopic m/z values of the precursor ions by use of ReAdW. However, we found that the precursor m/z values for some MS³ spectra in the mzXML files created by ReAdW in this way were incorrect. Therefore, a Perl script, ReAdW_patch (www.massmatrix.net), was developed to address this problem with mzXML files. The ReAdW_patch uses an associated mzData file, which is created from the .RAW data file by use of Xcalibur program (Thermo Fisher, San Jose, CA) and contains correct precursor m/z values for all MS³ spectra, to fix the incorrect MS³ precursor m/z values in the mzXML file.

The hierarchical MS²/MS³ database searches of the mzXML files were performed by use of the online version of MassMatrix (www.massmatrix.net) with the following options: i) Variable modifications: Sodium adduct of Aspartic acid and Glutamic acid, Phosphorylation of Serine, Theronine and Tyrosine; ii) Enzyme: trypsin; iii) Missed Cleavages: 2; iv) Peptide Length: 6 to 42 amino acid residues; v) Mass tolerances of 2.0 Da and 10 ppm for the precursor ions on LCQ Deca XP ion trap and LTQ-FTICR mass spectrometers respectively; and vi) Mass tolerances of 0.8 Da for the product ions. The standard data sets were searched against a protein database containing both the target protein database (α-Casein) and a decoy reversed National Center for Biotechnology Information non-redundant (NCBInr) human database (96,997 decoy protein sequences). The hierarchical search algorithm was automatically enabled in the searches by MassMatrix when MS³ spectral data were detected in the input mzXML files.

The two data sets were also evaluated by use of two-stage MS²/MS³ database searches in MassMatrix, Mascot (www.matrixscience.com) and X!Tandem (www.thegpm.org/TANDEM/), in which the MS² and MS³ data for a data set were searched in two parallel and separate database search processes against the same database. The MS² and MS³ spectral data of an mzXML were extracted to two separate MGF files by use of the tools available at www.massmatrix.net. For MS² data, precursor ion m/z values and charges in the mzXML were preserved during the extraction. For MS³ data, precursor ion m/z values in the mzXML were preserved and precursor ion charges for the MS³ were assumed to be the same as the precursor ion charges of their precursor MS² spectra. This assumption is valid here because the MS³ experiments were targeted at phosphorylation neutral loss ions and those ions had the same charge state as their precursor ions. It was also found that best search results were obtained by use of this extraction approach. The MGF files containing MS² and MS³ data for each experiment were then searched separately in MassMatrix, Mascot and X!Tandem against the same protein database as the one used in the hierarchical MS²/MS³ database searches. The search parameters for searches of the MS² data were identical to those in the hierarchical MS²/MS³ data searches. For the MS³ data, the mass tolerances of precursor ions for both LCQ and LTQ-FTICR data sets were set to be 0.8 Da due to fact that the precursor ions of MS³ spectra were measured with the same mass accuracy as those product ions of MS² spectra. In other words, both precursor and product ions of MS³spectra were measured with low mass accuracy even on the LTQ-FTICR mass spectrometer. Additional variable modifications, water loss of Serine and Theronine, were added to MS³ searches due to the facts that MS³ spectra were created from phosphorylation neutral loss ions and those ions have a mass difference of −18.0 Da from the original peptide. All other parameters in the MS³ searches were identical to those in the hierarchical MS²/MS³ data searches.

Results from MassMatrix and Mascot were output as html files. Results from X!Tandem were output as pepXML format. The scan number, charge, calculated mass, observed mass, mass difference, missed cleavages, score(s) and peptide sequence for each match from the three programs were extracted from the original output files into tab delimited TXT files by use of Perl scripts. All outputted peptide matches without any filtering from the three search programs were considered. The hierarchical MS²/MS³ searches in MassMatrix were performed in a single stage and thus merging of results was not required. For the two-stage MS²/MS³ searches in the three programs, the lists of peptide matches from the MS² and MS³searches were merged by use of a Perl script. In brief, for each set of MS² and MS³ spectra, a MS³ spectral peptide match was considered consistent to a MS² spectral peptide match if the MS³ peptide sequence was the same as or a subsequence of the MS² peptide sequence. Consistent MS²and MS³ peptide matches for the same set of MS²/MS³ spectra were combined as one peptide match with a score equal to the sum of scores of the MS² and MS³matches for MassMatrix and Mascot or the product of expectation values of the MS² and MS³ matches for X!Tandem. A MS² or MS³ spectral peptide match without a consistent counterpart was also considered as a match for the set of MS²/MS³ spectra with its original score. Under the circumstances that there were multiple peptide matches for a set of MS²/MS³ from a search program after merging, the one with the highest score was considered as the best match. All the Perl scripts used for result extraction and merging have been made available at www.massmatrix.net.

Receiver operating characteristic (ROC) analysis was used to evaluate the search algorithms. The pp score in MassMatrix, score in Mascot, expectation value in X!Tandem were used for ROC analysis. The decoy reversed human database creates ~10,000 times more theoretical peptides as the target α-Casein protein sequences. Therefore, false positive matches from the α-Casein proteins were assumed to be negligible. Thus, the peptide matches returned from the target α-Casein proteins were considered as true positives (TPs) while those from the decoy reversed human proteins were considered as false positives (FPs). [11]

3 RESULTS AND DISCUSSION

3.1 Hierarchical MS²/MS³ Database Search Algorithm in MassMatrix

Figure 2 shows the diagram of the hierarchical MS²/MS³ database search algorithm in MassMatrix. Theoretical peptide ions are created from the protein database by in silico digestion, addition of posttranslational modifications with specified PTMs and fragmentation. MassMatrix then matches the experimental MS² spectra to the theoretical peptide spectra to obtain candidate peptide matches. In this step, all peptide matches with theoretical precursor m/z values that match to the experimental MS² precursor ions are considered as valid candidate peptide matches even if their MS² product ion spectral quality is extremely poor and their scores are as low as 0. In the next step of the search process, MassMatrix matches the corresponding MS³ experimental spectra to the candidate peptide matches returned from the MS² search. During this step, the precursor m/z value of a MS³ spectrum is matched against all product ions in the theoretical MS² spectrum of a candidate peptide sequence. All potential MS³ precursor ions are then fragmented in silico and matched to the experimental MS³ spectrum. In some cases, multiple product ions in the theoretical MS² spectrum, such as b/y product ions and b/y product ions with neutral losses, may match a particular MS³ precursor. The best match between the theoretical MS³ spectra and the experimental one is determined to be the one(s) with the highest statistical score(s).[12]

Figure 2.

Diagram of hierarchical MS²/MS³ database search algorithm in MassMatrix. Theoretical peptides are created from the protein database by in silico digestion and modification with specified PTMs. MassMatrix first matches experimental MS² spectra to the theoretical peptides to obtain candidate peptide matches. MassMatrix then matches the corresponding MS³ spectra against each candidate peptide matches from the MS² search results. In this manner, the MS² and MS³ spectral data are used in concert to obtain the peptide matches.

The pp score for a MS² or MS³ spectrum match is defined as rhe negative common logarithm of the probability that the match is random for either the MS² or MS³ spectrum.[12] The pp score for a match from a hierarchical MS²/MS³ search is defined as the negative common logarithm of the probability that the match is random with regard to both MS² and MS³ spectrum, and is calculated by

$$\begin{array}{cc}\hfill {\mathit{pp}}_{{\mathrm{MS}}^2∕{\mathrm{MS}}^3}& =-\log (\mathrm{p}-{\text{value}}_{{\mathrm{MS}}^2∕{\mathrm{MS}}^3})\hfill \\ \hfill & =-\log (\mathrm{p}-{\text{value}}_{{\mathrm{MS}}^2}\times \mathrm{p}-{\text{value}}_{{\mathrm{MS}}^3})\hfill \\ \hfill & =-{\mathit{pp}}_{{\mathrm{MS}}^2}+{\mathit{pp}}_{{\mathrm{MS}}^3}\hfill \end{array}$$ (2)

The quality of a peptide match for a set of MS²/MS³ spectra is measured based on its final pp_MS²/MS³ score instead of either pp_MS² or pp_MS³ scores independently. Therefore, a set of MS²/MS³ spectra with either a MS² or MS³ spectrum of good quality will still return a significant peptied.

The algorithm also accounts for the modification site localizations for peptides with PTMs. Under the circumstances where there are several potential peptide matches with the same sequence but different modification site localizations for a set of MS²/MS³ spectra, all peptides are considered as potential peptide matches in the MS² step due to the fact that they have the same theoretical precursor MS² m/z value. These peptide matches are then searched in the MS³ search step. During the MS³ search step, those matches may or may not generate the same sets of theoretical MS³ spectra. For those MS³ spectra from phosphorylation neutral loss ions, those matches generate different sets of theoretical MS³ spectra due to their different phosphate neutral loss site locations. For those MS³ spectra from a subsequence that does not contain the potential modification sites, those peptide matches will have the same sets of theoretical MS³ spectra. The theoretical MS³ spectra are matched to the experimental MS³ and scores between the matches and the MS³ spectrum are calculated. MassMatrix infers the specific modification site locations based on the final pp_MS²/MS³ scores if the best match has much higher pp scores (Δpp > 6.0) than all other candidates. An example is shown in Figure 3 where a phosphopeptide containing multiple potential phosphorylation sites was identified with the specific phosphorylation site by the hierarchical MS²/MS³ search in MassMatrix. However, there are circumstances where several peptide matches may have very similar scores due to close proximity of modification sites or spectra resulting from a mixture of phosphopeptides. The user can specify whether the software will return all peptide matches or just the match with the highest pp_MS²/MS_³ score.

Figure 3.

Identification of a phosphorylation site on α-Casein-S2 by hierarchical MS²/MS³ database search. The figure demonstrates the capability of hierarchical search to localize the phosphorylation on a phosphopeptide containing multiple possible phosphorylation sites. The MS² (top) and MS³ (bottom) spectra are shown to highlight the effect of neutral loss on the MS² spectra. The detected phosphorylation site was previously reported by Kweon and Hakansson [10]. Neutral losses in the figure are labeled as follows: ^~ (loss of phosphate moiety),* (loss of ammonia), ' (loss of water).

We must draw a careful distinction between our hierarchical MS²/MS³ approach and other two-stage approaches. At present there are no database search programs that directly handle MS³ spectral data. Rather researches have to treat MS³ spectral as an additional set of MS² data, search MS³ spectral data separately in the same way as that for MS², and then merge the results from MS² and MS³ searches (a two-stage approach). However, in the two-stage analysis, there are two lists of candidate peptide matches for each pair of MS²/MS³ spectra. Each list is from the MS² search or the MS³ search. The two lists of candidates output from the database search program can be different even when multiple candidate peptide matches are allowed for each spectrum due to either the low quality of the MS² or the low quality of the MS³ spectrum.

Hierarchical MS²/MS³ is a natural and naive approach for analyzing DDNL MS³ spectra. The hierarchical approach does not simply use the MS² precursor to narrow down the list of candidate peptides for MS³ spectra. In this algorithm, MS² and MS³ spectra are searched in concert to obtain peptide matches with overall higher confidence instead of searched separately against the original protein database. Therefore, each pair of MS²/MS³ spectra has a single list of candidate peptide matches. The type of the ion in the MS² spectrum that undergo fragmentation to create the MS³ spectrum can also be determined in addition to the peptide sequence. For peptides with MS² spectra of poor quality due to the predominant neutral losses, such as phosphopeptides, their MS³ data may contain the necessary sequencing information to sequence the peptides and differentiate true and false positive matches.

The hierarchical MS²/MS³ search algorithm described herein is automated in MassMatrix and triggered by the program when MS³ spectra are detected in the input data file. This database search process is performed in a single stage and the match information for both MS² and MS³ are reported in a single result file. The algorithm is generic and can be used to search the data from all types of MS²/MS³ experiments. The performance of the algorithm on MS²/MS³ proteomic data for phosphopeptides were evaluated and discussed in details in the following sections.

3.2 Evaluation of Hierarchical MS²/MS³ Database Search

The hierarchical MS²/MS³ database search algorithm was evaluated by searching two data sets for tryptic digests of α-Casein from an LCQ Deca XP mass spectrometer and a LTQ-FTICR mass spectrometer. The data sets were searched against a database with target α-Casein protein sequences and the reversed human database appended as decoy sequences. From the LCQ data set, 480 out of 1310 spectra were scored with potential peptide matches. From the LTQ-FTICR data set, 1382 out of 4747 spectra were scored with potential peptide matches. The complete lists of peptide matches for the two data sets are provided in supplementary tables 1 & 2.

The search results were evaluated and compared with those from the two-stage MS²/MS³ searches in MassMatrix, Mascot and X!Tandem by use of receiver operating characteristic (ROC) analysis.[8, 13, 14] Because false positive peptide matches from the target α-Casein proteins were considered negligible due to the large decoy database, peptide matches from α-Casein proteins were considered as TPs and those from the decoy database were considered as FPs.[11] ROC curves were created by plotting TP against FP as the score threshold decreased in the search results. The ROC curves for the hierarchical MS²/MS³ search results and those from the two-stage searches in MassMatrix, Mascot and X!Tandem are displayed in Figure 4.

Figure 4.

ROC analysis of search results from the hierarchical MS²/MS³ searches in MassMatrix and two stage MS²/MS³ searches in MassMatrix, Mascot and X!Tandem for a tryptic of αCasein: (a) all the peptide matches for the LCQ data set, (b) all the peptide matches for the LTQ-FTICR data set, (c) phosphopeptide matches for the LCQ data set, and (d) phosphopeptide matches for the LTQ-FTICR data set.

An ideal database search results should return all true positives with scores higher than all false positives and a ROC curve with a right angle. A ROC curve toward the left indicates higher specificity and a curve toward the top indicates higher sensitivity. It can be seen from Figure 4a that the hierarchical MS²/MS³ search had better overall sensitivity than the two-stage MS²/MS³ searches in MassMatrix, Mascot and X!Tandem for the LCQ proteomic data. Figure 4c shows the ROC results for the phosphopeptides only. After enrichment, phosphopeptides were of higher abundance than non-phosphorylated peptides. Due to the fact that only a small portion of the peptide matches from the LCQ data set were non-phosphorylated peptides, The ROC results for phosphopeptides in the LCQ data set (Figure 4c) were very similar to those for all the peptides (Figure 4a). For the LTQ-FTICR data set, there are many non-phosphorylated peptide matches of relatively poorer quality than phosphopeptides. The sensitivity and specificity of phosphopeptides in the data set were much higher than those of all the peptides as indicated by the ROC analysis (Figure 4d vs. Figure 4b). Since the MS³ experiments were targeted at phosphorylation neutral loss ions in the MS² spectra, all the MS³ spectral data were presumably due to phosphopeptides. The improvement of the hierarchical MS²/MS³ search over the two-stage MS²/MS³ search in MassMatrix was not significant in the ROC analysis of all the peptides as shown in Figure 4b. However, the hierarchical MS²/MS³ search gained higher sensitivity and specificity than the two-stage MS²/MS³ searches in all three programs for the phosphopeptides as shown in Figure 4d.

Overall, the hierarchical MS²/MS³ searches performed in MassMatrix had improved sensitivities and specificities for the phosphopeptides than the two-stage MS²/MS³ searches in MassMatrix, Mascot and X!Tandem. At a false rate $\left(\frac{\mathrm{FP}}{\mathrm{TP}+\mathrm{FP}}\right)$ of 5%, the hierarchical MS²/MS³ search in MassMatrix returned 119 true positive phosphopeptides and the two stage MS²/MS³ searches in MassMatrix, Mascot and X!Tandem returned 112, 46 and 20 true phosphopeptides respectively for the LCQ data set. At the same false rate for the LTQ-FTICR data set, the hierarchical MS²/MS³ search in MassMatrix returned 394 true positive phosphopeptides and the two-stage MS²/MS³ searches in MassMatrix, Mascot and X!Tandem returned 305, 320, and 310 true phosphopeptides respectively. These results suggest the advantage of hierarchical MS²/MS³ database search algorithm over the two-stage MS²/MS³ search processes, especially for the data collect on LTQ-FTICR mass spectrometers.

There are two main factors that contribute to the higher sensitivities and specificities of the hierarchical MS²/MS³ search algorithm than the two-stage MS²/MS³ search approach. The first is that the hierarchical search process eliminates the discrepancy between the MS² peptide matches and MS³ peptide matches that may occur in the two-stage search process. In the hierarchical search process, a set of MS² and MS³ spectra is always used in concert to arrive at a peptide match (Figure 2). However, in the two-stage search process, the fact that a set of MS² and MS³ spectra is created from the same peptide by use of two consecutive fragmentations is ignored during database searching (Figure 1). Therefore, the MS² and MS³ spectra from the same original precursor ion may return two different peptide matches in the final merge of results. The other main advantage of the hierarchical search process is that the high mass accuracy of the precursor ion for the MS² experiment is inherited by the MS³ data analysis. In a typical MS²/MS³ experiment performed on high mass accuracy capable mass spectrometers (LTQ-FTICR and LTQ-Orbitrap mass spectrometers), precursor ions for the MS² spectra are measured with high mass accuracy (< 10 ppm), whereas the product ions for the MS² spectra (including the MS³ precursor) and product ions for the MS³ spectra are all measured with a relative lower mass accuracy (0.5 ~ 1.0 Da). In the two-stage process as shown in Figure 1, the MS³ data have to be searched with low mass accuracies for both precursor and product ions and the advantage of high mass accuracy is lost for MS³ data analysis. However, the high mass accuracy for the MS² precursor ions is inherited during the search process of MS³ spectral data in the hierarchical MS²/MS³ search process due to its hierarchical nature. In the hierarchical MS²/MS³ search process, MS³ spectra are only matched against the peptide candidates for their precursor MS² spectra. In other words, those peptides must have masses that matched the higher mass accuracy MS² precursor ion. In this way, the high accuracy of the MS² precursor ion is naturally inherited during the MS³ database searching (Figure 2). Due to the second factor, the improvement of the hierarchical MS²/MS³ search over the two-stage MS²/MS³ search was more significant for the high mass accuracy LTQ-FTICR data than that for the low mass accuracy LCQ data (Figure 4c vs. Figure 4d).

3.3 Effect of MS³ on Score Distribution

Figures 5a & 5b show the pp score distributions of TPs and FPs for the hierarchical MS²/MS³ searches and the searches without considering any MS³ data in MassMatrix. The additional MS³ data had little effect on the score distributions of FPs due to the randomness of FPs. However, the score distributions of TPs shifted to higher values after including the MS³ data in the hierarchical search mode. This separation of score resulted in improved sensitivities and specificities of the search results. It also improved the overall reliability and the number of the true positive peptide matches due to the fact that they had higher statistical pp scores.

Figure 5.

Score distributions of TPs and FPs in the hierarchical MS²/MS³ searches and the searches of MS² data only: (a) all the peptide matches returned for the LCQ data set, (b) all the peptide matches returned for the LTQ-FTICR data set, (c) phospopeptide matches for the LCQ data set, and (d) phosphopeptide matches for the LTQ-FTICR data set. The score distributions for TPs were split into two groups as labeled “1” and “2”.

The pp score distributions for TPs were split into two groups in the hierarchical MS²/MS³ search compared with those for TPs in the search without considering MS³ data (Figure 5). Group 1 represents peptides with good quality MS³ spectral matches (pp score ≥ 6.0) where their pp scores were improved significantly by including the MS³ data. Peptide matches in this group are well separated from FPs and can be identified with high sensitivity, specificity and reliability. Group 2 represents those peptides with moderate or poor quality MS³ spectral matches (pp score < 6.0) and those without any MS³ spectral matches. Their pp scores were only slightly improved or not improved at all.

Figures 5c & 5d show the pp score distributions for the phosphopeptides only. The pp score distributions for phosphopeptides from the LCQ data set (Figure 5a) were similar to those for all the peptides (Figure 5c). For the LTQ-FTICR data set, there were many non-phosphorylated peptide matches of relatively poorer quality than phosphopeptides. Because the MS³ targeted phosphopeptide ions, the score distribution of TPs for phosphopeptides was different from that for all the peptides as shown in Figures 5b & 5d. A great portion of the true positive phosphopeptide matches of the hierarchical MS²/MS³ search fell in group 1. These results suggest that the hierarchical MS³ experiments were effective for targeted phosphopeptide identification on both LCQ and LTQ-FTICR mass spectrometers.

4 CONCLUDING REMARKS

A novel algorithm to analyze hierarchical MS²/MS³ experiments was developed and automated in a database search program, MassMatrix. Due to the fact that the MS² and MS³ spectral data are collected sequentially from the same peptide precursor ion, the hierarchical MS²/MS³ database search algorithm has advantages over the two-stage search algorithms in which MS² and MS³ spectral data are searched independently. The hierarchical MS²/MS³ search algorithm takes full advantage of the hierarchical nature of the MS²/MS³ data. The hierarchical search process eliminates the discrepancy between the MS² peptide matches and MS³ peptide matches that may occur in the two-stage search process. Furthermore, the high mass accuracy of the precursor ion for the MS² experiment can be inherited by the MS³ data analysis in the hierarchical search algorithm.

The algorithm was evaluated and compared with the two-stage search approaches using the search programs MassMatrix, Mascot and X!Tandem. Receiver operating characteristic analysis showed that the hierarchical MS²/MS³ database search improved sensitivitiesties and specificities for phosphopeptides, especially for the data collect on an LTQ-FTICR mass spectrometer. At a false rate of 5%, the hierarchical MS²/MS³ search in MassMatrix returned 118 true positive phosphopeptides and the two-stage MS²/MS³ searches in MassMatrix, Mascot and X!Tandem returned 112, 39 and 13 true phosphopeptides respectively for the LCQ data set. At the same false rate for the LTQ-FTICR data set, the hierarchical MS²/MS³ search in MassMatrix returned 418 true positive phosphopeptides and the two-stage MS²/MS³ searches in MassMatrix, Mascot and X!Tandem returned 350, 370, and 321 true phosphopeptides respectively. Score distributions indicated that the additional MS³ data improved the overall reliability and the number of true positives due to the fact that the true positives of the MS²/MS³ search results had higher scores than those of the MS² results.

Glossary

Abbreviations

MS/MS or MS²

second stage of tandem mass spectrometry

MS³

third stage of tandem mass spectrometry

PTM

post-translational modification

CID

collision induced dissociation

ROC

receiver operating characteristic

TP

true positive

FP

false positive