Skip to main content
Download PDF

Fast alignment of mass spectra in large proteomics datasets, capturing dissimilarities arising from multiple complex modifications of peptides

BMC Bioinformatics volume 24, Article number: 421 (2023) Cite this article

Abstract

Background

In proteomics, the interpretation of mass spectra representing peptides carrying multiple complex modifications remains challenging, as it is difficult to strike a balance between reasonable execution time, a limited number of false positives, and a huge search space allowing any number of modifications without a priori. The scientific community needs new developments in this area to aid in the discovery of novel post-translational modifications that may play important roles in disease.

Results

To make progress on this issue, we implemented SpecGlobX (SpecGlob eXTended to eXperimental spectra), a standalone Java application that quickly determines the best spectral alignments of a (possibly very large) list of Peptide-to-Spectrum Matches (PSMs) provided by any open modification search method, or generated by the user. As input, SpecGlobX reads a file containing spectra in MGF or mzML format and a semicolon-delimited spreadsheet describing the PSMs. SpecGlobX returns the best alignment for each PSM as output, splitting the mass difference between the spectrum and the peptide into one or more shifts while considering the possibility of non-aligned masses (a phenomenon resulting from many situations including neutral losses). SpecGlobX is fast, able to align one million PSMs in about 1.5 min on a standard desktop. Firstly, we remind the foundations of the algorithm and detail how we adapted SpecGlob (the method we previously developed following the same aim, but limited to the interpretation of perfect simulated spectra) to the interpretation of imperfect experimental spectra. Then, we highlight the interest of SpecGlobX as a complementary tool downstream to three open modification search methods on a large simulated spectra dataset. Finally, we ran SpecGlobX on a proteome-wide dataset downloaded from PRIDE to demonstrate that SpecGlobX functions just as well on simulated and experimental spectra. We then carefully analyzed a limited set of interpretations.

Conclusions

SpecGlobX is helpful as a decision support tool, providing keys to interpret peptides carrying complex modifications still poorly considered by current open modification search software. Better alignment of PSMs enhances confidence in the identification of spectra provided by open modification search methods and should improve the interpretation rate of spectra.

Peer Review reports

Background

Interpreting fragmentation mass spectra and their assignment to a peptide sequence is an important and challenging issue in proteomics, especially when peptides carry one or several modifications. These modifications could explain the low rate of interpreted spectra after a standard bottom-up mass spectrometry analysis [1, 2]. While the modifications on proteins may be the consequence of co- or post-translational events in the cell, they may also be due to sample preparation (use of chemicals during the preparation, e.g., surfactants) or to the presence of allelic variants not described in the protein databases. Although reference databases of modifications exist and can draw up a list of more than a thousand modifications such as in Unimod [3], the variety of putative modifications is so large that existing databases cannot be exhaustive. Besides, it is also difficult, if not impossible, to predict and account for all possible modifications occurring in a sample.

In recent years, several open modification search (OMS) methods have been proposed to enhance the identification of mass spectra corresponding to modified peptides. These identifications are achieved through the comparison of experimental spectra to ‘reference spectra’, either simulated in silico from peptide candidates—referred to as theoretical spectra—or collected in a spectral database [4]. Whereas conventional methods (also called restricted or closed-search methods) try to identify experimental spectra in comparison to reference spectra in a narrow search mass window, OMS methods evaluate all or at least a large part of the reference spectra without any (or only a limited restriction) on their masses. Several advantages distinguish OMS methods from their conventional methods counterparts: (1) they enable the interpretation of spectra corresponding to peptides with unanticipated modifications; (2) the large number of modifications they can underpin does not alter identification results confidence as it occurs when an extremely large search space is generated by the introduction of many variable modifications on simulated spectra in the reference spectra database [5, 6].

The result of an OMS method is a list of PSMs (St, Se, ∆M) where St denotes the theoretical fragmentation spectrum of a candidate peptide, Se denotes an experimental spectrum, and ∆M is the mass difference between St and Se. This mass difference ∆M, if above the mass accuracy of the instrument, should be explained by one or several modifications carried by the peptide in the experience. Therefore, one or more mass shift(s) (resulting from the modification(s)) should be applied to St to align it with Se. When Se displays the fragmentation pattern of a peptide carrying only one modification, identification and localization (more or less precisely according to the software) of this modification are already resolved by several methods [7,8,9,10,11], mainly by testing successively the location of the modification on each amino acid. The detection of pairs of modified and unmodified peptides that may coexist in the same sample can also boost the sensitivity of the protein modification mapping [12]. When Se corresponds to a peptide carrying more than one modification, the interpretation is much more complicated because ∆M has to be split into several mass shifts. Several software [13,14,15,16,17,18] attempt to explain ∆M by a combination of masses related to modifications stored in a predetermined list, or deduced from the most frequent ones observed in the sample. However, to our knowledge, there is currently no open-source software adapted to routine laboratory use allowing the ∆M interpretation without any a priori on the modifications, including labile ones. To fill this gap, we implemented SpecGlobX, which aligns very quickly pairs of spectra (St, Se).

SpecGlobX (SpecGlob eXtended version adapted to eXperimental spectra) is a standalone software based on a dynamic programming algorithm. It derives from SpecGlob [19], which we first developed using perfect simulated spectra. This initial step was essential to evaluate the SpecGlob algorithm’s ability to consider complex and variable modifications and to optimize its execution time. To reach a fast execution speed, we introduced—among other things—a simplification in the alignment of spectra compared to previous methods also based on dynamic programming [20,21,22]: each peptide is modeled by a series of peaks corresponding to only the b-ions generated by an ideal fragmentation. However, on their side, experimental spectra are not ideal; they lack some fragmentation peaks and are noisy. In this article, we present the main adaptations developed in SpecGlobX towards the applicability of SpecGlob on experimental spectra, considering the imperfections of these spectra. Then, we highlight the benefits that SpecGlobX delivers downstream to three OMS methods: SpecOMS [23], MODPlus [14], and MSFragger [9] on a simulated dataset that mimics experimental spectra. Indeed, even if the complexity of an experimental spectrum is difficult to reproduce, simulated spectra are helpful to understand the behavior of an algorithm when there is no ground truth available for large-scale interpretation of PSM lists. Finally, we challenged SpecGlobX on a set of experimental spectra already interpreted by different OMS software, namely the HEK293 dataset [24]. Overall, we demonstrate that SpecGlobX highlights modifications that other methods missed while maintaining a good execution time.

SpecGlobX implementation

The algorithm of SpecGlobX follows three processing steps described in detail below. First, the completion of experimental spectra compensates for missing peaks; second, the alignment between completed experimental spectra and theoretical spectra including possible mass shifts; third, the optimization of the location of the suggested mass shifts (post-processing step).

Completion process of the experimental spectra

Usually, a peptide p with an amino acid sequence a1a2…ai…an is represented by a spectrum containing peaks that correspond to the b-ions {b1,…, bi, …, bn} and the y-ions {y1, …, yi,…, yn} generated by a perfect fragmentation. Then, a mass modification applied to amino acid ai results in a mass shift of the peaks bi to bn and y1 to yn−i+1, all the remaining peaks being unchanged. The shift of only a part of the peaks creates a complex situation to manage when aligning spectra with a dynamic programming approach. To overcome this difficulty, in SpecGlobX, a peptide p is only represented by its b-ion peaks, so spectra alignment only requires aligning pairs of b-ion peaks from St (representing amino acids) with pairs of b-ion peaks from Se. However, distinguishing b- from y-ions in an experimental spectrum is not easy a priori (i.e., before interpretation). Moreover, it is well known an experimental spectrum is likely to have several missing b-ion peaks. However, SpecGlobX needs as many b-ion peaks as possible to adjust its alignment. So, as we did in [25], SpecGlobX transforms Se into a new spectrum called Sec (for completed Se) by adding new peaks according to the following rule: each original peak in Se is hypothesized at first as a b-ion peak (assumption 1), so it can be directly compared to the b-ion peaks of the theoretical model (St); then each original peak in Se is hypothesized as corresponding to a y-ion (assumption 2); in that latter situation, SpecGlobX adds a new peak in the spectrum Sec, called a “complementary peak”, whose mass is M + m — mass(peak), where M is the total mass of the experimental spectrum Se and m the mass of the proton. We underline that a complementary peak in Sec can replace a missing b-ion peak provided the y-ion peak is observed in Se. When a peptide fragmentation produces both b- and y-ions simultaneously in Se, except in particular situations we will focus on later in this manuscript, the "completion process" does not add any peak. Conversely, if the fragmentation process has generated only one of the two expected ions, an additional peak is added in Sec compared to Se.

Representing peptides by their b-ions rather than their y-ions was initially arbitrary. It just allows an alignment in the direction where peptides are read – from left to right. Although y-ions are generally dominant in spectra, we assume that the introduction of complementary peaks in both solutions leads to similar results. Moreover, given the accuracy of fragment masses in current mass spectrometers (of the order of 0.005 Da), we hypothesize that the noise added by complementary peaks should not significantly interfere with alignments. The generation of a series of complementary peaks is illustrated in Fig. 1a.

Fig. 1
figure 1

Alignment of peptide HINATESVR on a simulated spectrum carrying two modifications. a On the left, the theoretical spectrum St of peptide HINATESVR is modeled only by its b-ion peaks, and below, the simulated spectrum Se is displayed with dummy intensities to mimic an experimental spectrum. However, note that intensities have no effect during the alignment with SpecGlobX. Se contains two modifications compared to St and only two b-ion peaks (represented in blue) are present in the spectrum. On the right, complementary peaks – below the x-axis—are added to Se, to generate the spectrum Sec; For example, one ‘original peak’ and its complementary peak are connected by a dotted line for better readability. b The score matrix was computed by SpecGlobX to align both spectra by dynamic programming. Masses have been rounded for simplification of the graphical representation; masses corresponding to complementary peaks are in green; transitions are numbered on the last row. The green (when SpecGlobX found an alignment) or orange arrows -when SpecGlobX found a realignment (with a mass shift)- delineates the path followed by the traceback step interpreting Se. c The interpretation of Se as HI[N][0.98]ATE[S][27.99]VR in terms of peptide sequence and mass shifts suggests a deamidation on “N” (mass increment of 0.98 Da) and a formylation on “S” (mass increment of 27.99 Da); non-aligned amino acids and mass shift values are in brackets. Original b-ion peaks are in blue in the interpreted Se and y-ions are in red

Full size image

Alignment between experimental and theoretical spectra

SpecGlobX like SpecGlob [19] relies on a dynamic programming technique to find the best alignment between the series of b-ion peaks in a theoretical spectrum St representing the amino acids of a peptide p and pairs of peaks of the complemented spectrum Sec. Briefly, each cell of a matrix D of size N x M (where N is the number of amino acids in p and M the number of peaks in Sec) is filled according to a scoring system based on (1) a boolean, aafound, which is set to true only if the mass of the considered amino acid of St is found between two peaks of Se; (2) three elementary scores SA, SR, and SN (where A stands for Align and refers to an extension of an alignment in progress, R for Re-align and refers to the introduction of a shift to “re-align” the i-th amino acid, and N for Non-aligned when aafound is false).

More precisely, each cell D[i][j] with 0 < i ≤ N – 1 and 0 < j ≤ M – 1 is computed according to the following rules:

  • If aafound = true, D[i][j] = max (D[i-1][k] + SA; max D[i-1][m] + SR, 0 ≤ m < k)

  • If aafound = false, D[i][j] = D[i-1][j] + SN

SpecGlob looks for an alignment that maximizes the score on the last row of D and possibly splits the mass difference ∆M between St and Se into several mass shifts. SpecGlob is described in detail in [19].

We now explain the adaptations made on SpecGlob to implement SpecGlobX on experimental spectra and interpret them successfully. In SpecGlobX, the scoring system globally takes (positively) into account the number of aligned amino acids and (negatively) the number of inserted mass shifts and the number of amino acids that are not aligned. Each time SpecGlobX introduces a mass shift in an alignment (an elementary decision we call a realignment), this increases the risk to align a b-ion peak of St on an inappropriate ion peak of Sec. To mitigate this risk of misalignment, the score penalty is defined such that a realignment is preferred over non-aligned amino acids only if this realignment is subsequently compensated by the alignment of at least two amino acids. Thus, in the SpecGlobX scoring system, a realignment is more penalized than the non-alignment of one amino acid.

Considering the above rule, SpecGlobX uses the following scoring system –empirically chosen—during the dynamic programming step: (1) when a pair of peaks is aligned without any mass shift, the score SA is increased by 10 if the fragmentation generated both the b- and y-ion (illustrated in transitions 2 and 9 in Fig. 1b) and by 7 otherwise (i.e., if only the b-ion or the y-ion peak is present in Se, as illustrated in transitions 1, 5 and 6 in Fig. 1b); (2) when a pair of peaks corresponding to an amino acid is aligned, but this alignment requires a mass shift to displace the pair of peaks from an amino acid on a pair of peaks of Sec, then the score SR is decreased by 8 or 6 depending on whether the fragmentation generated both b- and y-ions or only one of the two (illustrated in transitions 4 and 8 Fig. 1b); (3) when the amino acid is non-aligned, then the score SN is decreased by 4 (illustrated in transitions 3 and 7 in Fig. 1b). The whole filling of the dynamic programming matrix corresponding to the alignment of the St and Sec spectra corresponding to the peptide HINATESVR is illustrated in Fig. 1b. Lastly, SpecGlobX generates the best alignment under the form of a string as was done in [19] by a traceback step (Fig. 1c).

For interested readers who want to dig deeper into the algorithm, we also mention a small subtlety that must be taken into account while the dynamic programming matrix is filled in SpecGlobX, compared to its previous version SpecGlob. A key observation in experimental spectra is that when a peak between two amino acids is missing in Sec, not only those two amino acids cannot be aligned (because their masses cannot be found in the spectrum), but this non-alignment generates a subsequent realignment whose mass shift is equal to a null value. To deal with this issue, this particular realignment (only due to one or several missing peaks) must be scored as an alignment (positive contribution to the score) rather than as a realignment (negative contribution to the score).

Optimization of the location of the suggested mass shifts

In this last processing step, SpecGlobX adjusts the locations of the suggested mass shifts (if any) to maximize the number of shared peaks between St and Se. One of the most important goals of this step is to highlight the presence of a non-aligned mass if any, and potentially to relocate modifications on the C-terminal side of the peptide.

The “completion process” compensates for a missing b-ion in Sec when the y-ion is displayed in Se, as long as the mass of the complementary peak can be directly inferred from the mass of the peptide. Regrettably, this relation is not always applicable, for instance when the precursor ion loses a neutral chemical group (such as water) in an early stage of the fragmentation process. Similar situations happen when Se merges the fragmentation of two peptides (possibly of identical sequences) linked by a chemical bond (such as a disulfide bridge) leading to homo or heterodimers. In these cases, all the complementary peaks are shifted from their expected value by a constant mass, which we denote as Mloss. Since the mass of the fragmented peptide observed in Se does not fit the measured mass of the experimental peptide, Mloss can be considered a non-aligned mass. To exemplify the generation of Sec in this instance, in Fig. 2a, we display the spectrum we obtain after the completion process when a neutral loss of 301.99 Da (i.e. the mass of a frequent modification referenced in Unimod, although still annotated as ‘unknown’) has been added to the spectrum Se shown Fig. 1a. In this example, ∆M is the sum of two mass shifts observed in Se (0.98 Da and 27.99 Da, i.e., the same mass shifts as in Fig. 1) and one Mloss (301.99 Da). The dynamic programming process can then align some of the complementary peaks that do not correspond to any b-ion peaks in the original spectrum Se. The number of shared peaks between Sec and St is increased by this alignment, but the number of shared peaks between Se and St is not. This situation is illustrated by the simulated spectrum displayed in Fig. 2. Since the b1-ion of peptide HINATESVR is not present in the spectrum, the mass of the first amino acid H is only given by the presence of the y8-ion. Then, the first amino acid H can be only aligned on the complementary peak shifted by 301.99 Da, a shift only due to Mloss in Sec. If we compute directly the number of shared peaks between Se and the suggested alignment [301.99]HI[N][0.98]ATE[S][27.99]VR, none of the y-ion peaks would align (due to Mloss). To circumvent this issue, the post-processing step evaluates whether each shift increases the number of peaks shared between Se and St. If not, the irrelevant shift is removed from the alignment and its value is accumulated to form the non-aligned mass indicated after the symbol ‘_’ in the result. We illustrate this transformation in Fig. 2c.

Fig. 2
figure 2

Alignment of peptide HINATESVR on a simulated spectrum carrying two modifications and a neutral loss. a The simulated spectrum Se is the same as the one presented in Fig. 1a, except that the mass of the peptide has been increased by a neutral loss of 301.99 Da. As a result, complementary peaks in Sec –represented below the x-axis – are shifted by 301.99 Da compared to the complementary peaks displayed in Fig. 1c; For example, the peak 690.34 is connected to its complementary peak 668.16 (366.17 + 301.99) by a dotted line b The score matrix computed by SpecGlobX to align the peptide HINATESVR with Sec by dynamic programming; Masses have been rounded for simplification of the graphical representation; c After the traceback step, Sec is interpreted as [301.99]HI[N][0.98]ATE[S][27.99]VR and after the post-processing step, Se is interpreted as HI[N][0.98]ATE[S][27.99]VR_301.99, identifying peptide HINATESVR with two modifications (a deamidation on “N” and a formylation on “S”) and a non-aligned mass of 301.99 Da (neutral loss), which is the correct identification

Full size image

The fact that the SpecGlobX scoring system has been designed to limit the number of shifts, as was already discussed, is another problem that may prevent a proper alignment. Indeed, if a modification is located on the C-terminal residue of a peptide, a non-aligned mass will provide a better score than a shift in the last rows of D (because a shift lowers the score without giving the chance to compensate for this decrease at the end of the peptide). To solve this problem, SpecGlobX tests systematically whether a modification on the C-terminal residues should be preferred to a non-aligned mass and chooses this option if the number of aligned peaks is higher. This principle works well if the non-aligned mass found at the end of the alignment does not accumulate several modification masses (for instance, a neutral loss and a modification). Note, however, that a neutral loss could have been evidenced earlier in the alignment by a switch between a series of aligned b-peaks and a series of aligned y-peaks, a switch observed frequently.

Finally, the post-processing step includes other transformations to facilitate the interpretation of alignments. For instance, non-aligned amino acids associated with a negative shift are removed if this deletion does not degrade the number of shared peaks (the case for a semi-tryptic peptide).

As a result of SpecGlobX, both alignments (before and after post-processing) are returned to the user.

Results and discussion

Application features

We implemented SpecGlobX as a standalone Java application. As input, SpecGlobX reads a file containing spectra in MGF or mzML format (parsed by the JMZReader library [26] and a comma-delimited spreadsheet that describes the list of PSMs to be aligned. This list of PSMs may be the result of any OMS method or a list of PSMs generated by a user.

The user has immediate access to the most frequent parameters from the interface (Fig. 3), while he/she can also edit an additional file containing all the parameters and scores used in the implementation. A click on the ‘Launch Alignments’ button starts SpecGlobX as a mono or multi-thread execution and produces a comma-delimited CSV file as output. For each PSM, two alignments are returned: the ‘preAlignedPeptide’ column gives the alignment obtained just after the traceback step, while the ‘alignedPeptide’ column displays the alignment once the post-processing step is done. SpecGlobX returns all alignments as strings: amino acids not explained in the alignment and shifts are in square brackets. Therefore, the user visualizes which amino acids have corroborating peaks in the aligned spectrum. Moreover, the value of the non-aligned mass (if any) is written after the symbol ‘_’ at the end of the string. SpecGlobX also returns the number of peaks shared between Se and St before and after the alignments, and the percentage of the overall signal intensity explained by the alignment.

Fig. 3
figure 3

SpecGlobX graphic interface

Full size image

SpecGlobX is fast and can align one million PSMs in about 1.5 min on a desktop with 8 allocated threads and in about 10 min in a mono-thread configuration (around 100 times faster than InsPecT).

Evaluation of SpecGlobX with a simulated spectra dataset

Firstly, we challenged SpecGlobX on a simulated spectra dataset Dsim generated from 50,000 tryptic unique peptides randomly selected from the human proteome—downloaded from Ensembl99, release GrCh38 [27]—and whose lengths span from 12 to 25 amino acids. We transformed those peptides into doubly charged simulated spectra with the following modifications: we introduced a deamidation on each asparagine (N+ 0.984016 Da) and a sodium adduct on each aspartic acid (D+ 21.981943 Da); next, we randomly removed 20% of the peaks to simulate missing peaks. Two-thirds of the removed peaks were b-ion peaks and one-third were y-ion peaks, as we observed that missing peaks are mainly b-ion peaks in experimental fragmentation spectra acquired in an HCD cell (a common instrument configuration of high-resolution mass spectrometers used in proteomics). Then, we simulated noise by adding up to 60 peaks of randomly selected masses. Lastly, we simulated a neutral loss of 17.03 Da on the peptide mass of each spectrum. We provided the complete list of ‘correct PSMs’ as input to SpecGlobX i.e., each simulated spectrum was associated with the peptide it derived from. Depending on the peptide’s composition, simulated spectra in Dsim contain a range of one to five modifications (among which at least one neutral loss of 17.03 Da). Given that every spectrum in Dsim is simulated, we can unambiguously determine whether a modification suggested by SpecGlobX is correct. Our assessment shows that SpecGlobX performs well: it detected the neutral loss on a large proportion of spectra (nearly 70%) and correctly identified 53% of all the expected modifications (Table 1, first row).

Table 1 Performance of SpecGlobX measured on the simulated dataset Dsim. At first, we provided the full list of ‘correct PSMs’ as input to SpecGlobX: each simulated spectrum is associated with the peptide sequence it derives from (row 1). Then, we executed SpecGlobX on the PSMs returned by three different OMS methods (rows 2 to 4). Several percentages of correct identifications are summarized with and without (W/o) SpecGlobX. As a first criterion (column 2), a PSM is counted as correct if the spectrum is associated with the peptide sequence it derives from; the second criterion measures the percentage of PSMs that exhibits the neutral loss among the 50,000 PSMs (columns 3 and 4); the third criterion counts the PSMs in which all detected modifications are correctly identified and placed related to all PSMs (columns 5 et 6); the last criterion refers to the percentage of modifications that are correctly identified relative to the number of modifications (about 103,000) incorporated in the simulated spectra (columns 7 and 8)
Full size table

Given its underlying algorithm, it appears that SpecGlobX's performance depends on the recognition of the peptide's amino acids in the spectrum, which can be more or less effective depending on the percentage of missing peaks, the absence of fragmentations representing certain amino acids or the presence of co-fragmented peptides in the spectrum. We varied several spectral simulation parameters to assess the effects of these potential sources of reduced quality spectra (Additional file 1: Tables S1, S2 and S3). Achieving excellent results when spectra display complete fragmentation information, SpecGlobX performance unsurprisingly declines when spectra quality deteriorates. However, SpecGlobX proved resilience to loss of informative peaks or co-fragmentation of peptides. Whether a substantial part of the fragmentation information is lost (up to 50% missing peaks or up to six amino acids missing in the fragmentation traces) or whether spectra represent co-fragmented peptides, SpecGlobX still return correct interpretations on highly modified peptides without any a priori.

Complementarity between SpecGlobX and existing OMS methods

Secondly, we evaluated whether SpecGlobX improves the results provided by three existing OMS software: MODPlus [14], MSFragger [9], and SpecOMS [23]. To this end, we first challenged each of these tools with Dsim spectra (see Additional file 1 for detailed parameters) and we examined each software’s ability to interpret the mass modifications in the correctly identified PSMs. On one hand, MODPlus explains ∆M observed between Se and St by combining modifications stored in a predefined list (consisting of all or only a selection of modifications referenced in Unimod, or a list provided by the user). We have previously shown that this strategy is particularly efficient if the objective is to identify the maximum number of peptides carrying common or known in-advance modifications. However, as soon as a peptide carries a modification not reported in the list, MODPlus fails in recovering a correct interpretation. Since MODPlus does not consider mass losses, it was unable to provide a good interpretation of any of the correctly identified PSMs, even though it interpreted and located 18% of the modifications (Table 1). On the other hand, MSFragger and SpecOMS try to explain each modification by a single shift. Therefore, both of them yielded good interpretations for only 27% of the correctly identified PSMs, which corresponded to peptides carrying a single modification (Table 1). Then, we applied SpecGlobX on the PSMs correctly identified by SpecOMS, MODPlus, and MSFragger. Regardless of the search engine used to produce the PSMs, a large part of the neutral losses misinterpreted by the three search engines can be recovered by SpecGlobX. Similarly, the percentage of correct modifications relative to the number of expected modifications (based on the number of N and D in Dsim) is multiplied by two or three depending on the OMS software used. The limitation of SpecGlobX, therefore, is that even when the PSM is correct, only half of the modifications (53%) are perfectly identified and localized. This is not a surprise since some modifications cannot be corroborated without any a priori due to missing peaks. In addition, if two modifications are present on two adjacent amino acids, SpecGlobX is not able to distinguish them. Consequently, although a manual curation by an expert is still needed to remove ambiguities in the proposed alignments, SpecGlobX is helpful as a decision-support tool. SpecGlobX provides keys to interpret peptides carrying complex modifications, while this issue is still poorly considered by current open modification search software.

Alignments highlighted on the HEK293 dataset

As a final evaluation, we ran SpecGlobX as a post-processing tool of SpecOMS on 24 spectra datasets generated from HEK293 cells [24] downloaded from PRIDE (PXD001468). This collection of spectra datasets has been used several times to evaluate open modification search methods. We converted the raw spectra files into the MGF format with msConvert release 3.0 [28] and obtained a list of PSMs using SpecOMS (configuration described in Additional file 1). Next, we divided the PSMs into three groups according to ∆M (∆M < 0, ∆M = 0, ∆M > 0) and filtered the PSMs to comply with an FDR < 1% computed separately for each group using a target-decoy approach. Overall, 429,703 spectra were validated, among which 149,794 had a positive ∆M and 16,545 a negative ∆M. Finally, we ran SpecGlobX and clustered the mass shifts (resp. the non-aligned masses) into 1 Da wide mass bins between − 500 and 500 Da, and counted their number below − 500 Da and above 500 Da. This results in the histograms shown in Fig. 4. Non-aligned masses are in the majority (Fig. 4a), with many isotopic errors, but there are also a large number of ∆M > 500 Daltons. This phenomenon is explained below, using an example based on a small subset of spectra.

Fig. 4
figure 4

Distribution of mass shifts and non-aligned masses on the HEK293 dataset interpretation with SpecGlobX downstream to SpecOMS (FDR < 1%). a A general histogram overview shows the major peaks. b In this detailed histogram above, non-aligned major peaks around 1 Da have been removed and the mass delta scale is reduced from − 200 to 400 Da so that minor peaks can be observed. The number of mass shifts in each 1 Da wide bin is represented in orange. c In the detailed histogram below, the number of non-aligned mass shifts in each 1 Da wide bin is represented in blue

Full size image

Our results are consistent with those published in [24], highlighting the most common modifications that have already been observed (Fig. 4b). However, we note that there is a large proportion of non-aligned masses.

Next, to emphasize the value of SpecGlobX on experimental spectra, we focused on the interpretation of a subset of 77 spectra identifying peptide pep = DATNVGDEGGFAPNIIENK according to SpecOMS (results are in Additional file 2). This subset of spectra seems to be a good representation of a set of spectra identifying the same peptide in the dataset. Indeed, as shown in Fig. 5, among the 77 spectra identifying pep, only 15 exhibit ∆M values around 0 or 1 Da (referring to an exact match to the expected mass or an incorrect picking of the good isotope, respectively), while the others are distributed over 44 different values of ∆M between − 500 and 2065 Da.

Fig. 5
figure 5

Distribution of mass deltas carried by a set of spectra identifying peptide DATNVGDEGGFAPNIIENK. These spectra are interpreted from the file b1922_293T_proteinID_02A_QE3_122212

Full size image

When examining the alignments suggested by SpecGlobX, we note a large number of non-aligned masses we did not expect (36 out of the 77 alignments returned by SpecGlobX contain a non-aligned mass), an observation we can extend to the alignments of all PSMs obtained on the complete dataset. Situations combining non-aligned masses including neutral losses and one or several modifications on the same peptide -as previously evaluated with Dsim- seems to be a reality in this dataset. The exploration of the spectra relations as done in [29] helps to decipher whether these neutral loss identifications are correct.

All the number of shared peaks between pairs of spectra are easily and rapidly completed by running SpecOMS (with the search_mode parameter set to “spectra”). Based on these data, we transformed the relations between the 77 spectra representing pep into a spectral network using GePhi [30], in which each node corresponds to a spectrum and two vertices are connected by a link when the two spectra share at least 26 out of the 50 most intense peaks. The resulting graph displayed in Fig. 6 includes 74 connected nodes and 998 links. Links between pairs of spectra are weighted by their number of shared peaks. The force-directed layout algorithm Force Atlas [31] generated the detailed arrangement of nodes on the drawing. As a consequence, groups of nodes that are densely connected (i.e., that share the same set of peaks) are immediately highlighted as clusters of nodes.

Fig. 6
figure 6

The closeness of spectra identifying peptide DATNVGDEGGFAPNIIENK (HEK293 dataset). In this picture, two vertices representing spectra are linked if the two spectra share at least 26 out of the 50 most intense peaks. The size of each node is related to its degree, the label on each node displays the ∆M value, and colors on the nodes group similar masses and their isotopes for readability. Each link between a pair of spectra is weighted by the number of shared peaks. The arrangement of nodes is done by the directed Force Atlas algorithm so that the topology of the graph represents the closeness of spectra in terms of number of shared peaks

Full size image

We remind that two spectra representing pep with different neutral losses should share most of their peaks –and should share most of their peaks with the spectra identifying pep without modification (∆M = 0 or ∆M = 1 corresponding to incorrect use of isotopic peaks). On the contrary, the number of peaks shared between two spectra representing pep drops by half for each modification differentiating the two spectra. The presence of many non-aligned masses is confirmed by several manual interpretations and can be generalized by the topology of the graph with a large central cluster For instance, ∆M = {301.99; 302.99; 303.99} Da represents a labile modification in three isotopic forms (an annotated spectrum plot (Spectrum #54,484) highlighting the labile modification is available in Additional file 1). Consistent with that, all the spectra carrying these modifications and identifying pep with a strong confidence share most of their peaks with most other spectra, particularly with spectra associated with PSMs with ∆M = 0 or 1.

Next, we interpreted the 77 spectra identifying pep with MODPlus and MSFragger to evaluate if the behavior on experimental spectra is similar to what we observed on simulated spectra. The answer is clearly yes (the interpretation provided by each software is given in Additional file 2). Even though MODPlus provides the same PSMs as SpecOMS with only two exceptions, the interpretations in terms of modifications are the same for only 19 of these 75 PSMs. On the contrary, MSFragger differs more frequently on the PSM (5 spectra are interpreted as different peptides), but the interpretations of the modifications carried by 40 among 72 PSMs are the same.

As already highlighted by simulated spectra, well-known modifications are well-located by MODPlus (D+ 21.98 Da, N-Ter+ 43.01 Da, D+ 53.92 Da). However, one surprise is some of the nodes associated with the latter ∆M (Fig. 5, ∆M = +53.92 Da for instance) are densely connected to nodes corresponding to mass losses, whereas some other nodes of the same ∆M are close to each other but disconnected from the cluster of nodes. This suggests that some spectra contain a clear signature of the modification with a shift of peaks carrying the modification, while other spectra have partly or totally lost this signature. These findings –which may be related to the significant proportion of non-aligned masses seen in Fig. 4a—imply that non-aligned peaks are more frequent in spectra than previously thought, leading to a lower number of shared peaks if those peaks are not taken into account by search engines.

Our small experimental dataset also illustrates that the counterpart of a priori knowledge and ∆M restrictions sometimes have bad effects on the interpretations of spectra (annotated spectrum plots discussed below are available in Additional file 1). On one hand, spectrum #54,025 interpreted without any doubt by MODPlus as the O-linked glycosylated peptide YGKDATNVGDEGGFAPNILENK (probability = 1, glycan 1914.69 Da) is more reliably interpreted as a homodimer of pep (an artifact that can be frequently observed [32] added to the frequent neutral loss 301.98 Da with SpecGlobX and MSFragger. On the other hand, MSFragger suggests another peptide than pep for spectrum #53,811, leading to an incorrect PSM. Indeed, some readers could be intrigued by the presence of spectrum #53,811 associated with ∆M = − 451.18 Da at the center of the cluster of nodes. The spectrum position in the graph suggests that #53,811 shares most of its peaks with the entire peptide pep (nodes with ∆M = 0 or 1), which excludes the possibility that #53,811 arises from a semi-tryptic form of pep (the MODPlus suggestion, but in this case, the number of shared peaks with other spectra representing pep would have been much lower). At the same time, the ∆M value is negative, suggesting a loss of amino acids. Even though none of the software has found this most probable interpretation, the usage of SpecGlobX as a decision support tool combined with an expert analysis of the spectrum graph has rather easily led us to the following explanation: the charge state evaluation of the spectrum is wrong (3+ rather than 2+) and the spectrum carries the labile modification of 302.99 Da (present on a large proportion of the spectra in this sample).

Integration of SpecGlobX with other OMS methods

Because SpecGlobX has a command line interface, it is simple to integrate in a workflow. On the other hand, rather than being used as a post-processing tool for OMS methods, it could also be integrated into OMS engines to identify the best candidates, taking into account spectra with multiple modifications.

Before implementing this integration, minor changes in the algorithm should be explored. First, SpecGlobX performs a kind of “global alignment”, in which the score used to start the traceback comes from a cell in the last row of the dynamic programming matrix D. This latter score may not be the highest in D. For instance, if Se corresponds to a semi-tryptic peptide that has been truncated from its C-Terminal end, the best alignment score in D drops after a maximum because the last amino acids are not found. Thus, while the “global alignment score” determines the optimal alignment path for PSMs, the highest score in the matrix -so-called the “local alignment score”- may be better to inter-classify PSMs. Note that this “local alignment score” computed without any traceback or post-processing step would be particularly fast to compute. In the same vein, if Se aligns with the C-terminal part of the peptide but not with its N-terminal, the initial imperfect alignment should not prevent further partial alignment. To avoid this pitfall, a solution might be to limit the lower score in D–if the score gets lower, it is reset to the minimum value. Second, each method has its final score philosophy, in which the SpecGlobX alignment score may play a role. Some fine-tuning may be required to achieve satisfactory results.

Conclusions

SpecGlobX is an easy-to-use software developed to quickly align PSMs. SpecGlobX can be seen as an additional tool to existing OMS software, providing a good decision support tool to highlight complex modifications carried by peptides. We have shown on a large set of simulated spectra (modeled such as to simulate imperfect experimental spectra) that SpecGlobX improves the identification of complex and unanticipated modifications after reprocessing the PSMs lists provided by OMS software. Next, we demonstrated the usefulness of SpecGlobX on a subset of a well-known experimental spectra dataset downloaded from PRIDE, highlighting the presence of a large proportion of non-aligned masses due to neutral losses, charge estimation error, and the presence of dimers.

SpecGlobX is freely available to all users, including the source code.

Availability and requirements

Project name: SpecGlobX

Project home page: https://github.com/bibs-lab/SpecGlobX

Operating system(s): Platform independent

Programming language: Java

Other requirements: 1.8 or higher

License: GNU GPL

Any restrictions to use by non-academics: None

Availability of data and materials

This published article and its supplementary files give all information required to reproduce the simulated dataset DSim. The HEK293 spectra dataset supporting the conclusions of this article was downloaded from the PRIDE repository, PXD001468, file b1922_293T_proteinID_02A_QE3_122212.raw https://www.ebi.ac.uk/pride/archive/projects/PXD001468

References

  1. Griss J, Perez-Riverol Y, Lewis S, Tabb DL, Dianes JA, Del-Toro N, et al. Recognizing millions of consistently unidentified spectra across hundreds of shotgun proteomics datasets. Nat Methods. 2016;13(8):651–6.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  2. Bogdanow B, Zauber H, Selbach M. Systematic errors in peptide and protein identification and quantification by modified peptides. Mol Cell Proteomics. 2016;15(8):2791–801.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  3. Creasy DM, Cottrell JS. Unimod: protein modifications for mass spectrometry. Proteomics. 2004;4(6):1534–6.

    Article  CAS  PubMed  Google Scholar 

  4. den Ridder M, Daran-Lapujade P, Pabst M. Shot-gun proteomics: why thousands of unidentified signals matter. FEMS Yeast Res. 2020;20(1):foz088.

    Article  Google Scholar 

  5. Colaert N, Degroeve S, Helsens K, Martens L. Analysis of the resolution limitations of peptide identification algorithms. J Proteome Res. 2011;10(12):5555–61.

    Article  CAS  PubMed  Google Scholar 

  6. Bugyi F, Szabó D, Szabó G, Révész Á, Pape VFS, Soltész-Katona E, et al. Influence of post-translational modifications on protein identification in database searches. ACS Omega. 2021;6(11):7469–77.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  7. Savitski MM, Nielsen ML, Zubarev RA. ModifiComb, a new proteomic tool for mapping substoichiometric post-translational modifications, finding novel types of modifications, and fingerprinting complex protein mixtures. Mol Cell Proteomics. 2006;5(5):935–48.

    Article  CAS  PubMed  Google Scholar 

  8. Riffle M, Hoopmann MR, Jaschob D, Zhong G, Moritz RL, MacCoss MJ, et al. Discovery and visualization of uncharacterized drug-protein adducts using mass spectrometry. Anal Chem. 2022;94(8):3501–9.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  9. Kong AT, Leprevost FV, Avtonomov DM, Mellacheruvu D, Nesvizhskii AI. MSFragger: ultrafast and comprehensive peptide identification in mass spectrometry-based proteomics. Nat Methods. 2017;14(5):513–20.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  10. Chen Y, Chen W, Cobb MH, Zhao Y. PTMap–a sequence alignment software for unrestricted, accurate, and full-spectrum identification of post-translational modification sites. Proc Natl Acad Sci U S A. 2009;106(3):761–6.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  11. Horlacher O, Lisacek F, Müller M. Mining large scale tandem mass spectrometry data for protein modifications using spectral libraries. J Proteome Res. 2016;15(3):721–31.

    Article  CAS  PubMed  Google Scholar 

  12. Cifani P, Li Z, Luo D, Grivainis M, Intlekofer AM, Fenyö D, et al. Discovery of protein modifications using differential tandem mass spectrometry proteomics. J Proteome Res. 2021;20(4):1835–48.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  13. Solntsev SK, Shortreed MR, Frey BL, Smith LM. Enhanced global post-translational modification discovery with MetaMorpheus. J Proteome Res. 2018;17(5):1844–51.

    Article  CAS  PubMed  Google Scholar 

  14. Na S, Kim J, Paek E. MODplus: robust and unrestrictive identification of post-translational modifications using mass spectrometry. Anal Chem. 2019;91(17):11324–33.

    Article  CAS  PubMed  Google Scholar 

  15. Chi H, Liu C, Yang H, Zeng WF, Wu L, Zhou WJ, et al. Comprehensive identification of peptides in tandem mass spectra using an efficient open search engine. Nat Biotechnol. 2018;36(11):1059–61.

    Article  CAS  Google Scholar 

  16. Devabhaktuni A, Lin S, Zhang L, Swaminathan K, Gonzalez CG, Olsson N, et al. TagGraph reveals vast protein modification landscapes from large tandem mass spectrometry datasets. Nat Biotechnol. 2019;37(4):469–79.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  17. Burke MC, Mirokhin YA, Tchekhovskoi DV, Markey SP, Heidbrink Thompson J, Larkin C, et al. The hybrid search: a mass spectral library search method for discovery of modifications in proteomics. J Proteome Res. 2017;16(5):1924–35.

    Article  CAS  PubMed  Google Scholar 

  18. Bittremieux W, Meysman P, Noble WS, Laukens K. Fast open modification spectral library searching through approximate nearest neighbor indexing. J Proteome Res. 2018;17(10):3463–74.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  19. Lysiak A, Fertin G, Jean G, Tessier D. SpecGlob: rapid and accurate alignment of mass spectra differing from their peptide models by several unknown modifications. bioRxiv. 2022; doi: https://0-doi-org.brum.beds.ac.uk/10.1101/2022.05.31.494131.

  20. Pevzner P, Dancik V, Tang C. Mutation-tolerant protein identification by mass spectrometry. J Comput Biol. 2000;7(6):777–87.

    Article  CAS  PubMed  Google Scholar 

  21. Pevzner PA, Mulyukov Z, Dancik V, Tang CL. Efficiency of database search for identification of mutated and modified proteins via mass spectrometry. Genome Res. 2001;11(2):290–9.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  22. Bandeira N, Tsur D, Frank A, Pevzner PA. Protein identification by spectral networks analysis. 2007.

  23. David M, Fertin G, Rogniaux H, Tessier D. SpecOMS: a full open modification search method performing all-to-all spectra comparisons within minutes. J Proteome Res. 2017;16(8):3030–8.

    Article  CAS  PubMed  Google Scholar 

  24. Chick JM, Kolippakkam D, Nusinow DP, Zhai B, Rad R, Huttlin EL, et al. A mass-tolerant database search identifies a large proportion of unassigned spectra in shotgun proteomics as modified peptides. Nat Biotechnol. 2015;33(7):743–9.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  25. Cliquet F, Fertin G, Rusu I, Tessier D, editors. Comparison of spectra in unsequenced species. 4th Brazilian Symposium on Bioinformatics (BSB 2009); 2009; Porto Alegre, Brazil.

  26. Griss J, Reisinger F, Hermjakob H, Vizcaíno JA. jmzReader: a Java parser library to process and visualize multiple text and XML-based mass spectrometry data formats. Proteomics. 2012;12(6):795–8.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  27. Yates AD, Achuthan P, Akanni W, Allen J, Alvarez-Jarreta J, Amode MR, et al. Ensembl. Nucleic Acids Res. 2020;48(D1):D682–8.

    CAS  PubMed  Google Scholar 

  28. Chambers MC, Maclean B, Burke R, Amodei D, Ruderman DL, Neumann S, et al. A cross-platform toolkit for mass spectrometry and proteomics. Nat Biotechnol. 2012;30(10):918–20.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  29. Watrous J, Roach P, Alexandrov T, Heath BS, Yang JY, Kersten RD, et al. Mass spectral molecular networking of living microbial colonies. Proc Natl Acad Sci U S A. 2012;109(26):E1743–52.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  30. Bastian M, Heymann S, Jacomy M. Gephi: An Open Source Software for Exploring and Manipulating Networks. 03, 2009.

  31. Jacomy M, Venturini T, Heymann S, Bastian M. ForceAtlas2, a continuous graph layout algorithm for handy network visualization designed for the Gephi software. PLoS ONE. 2014;9(6): e98679.

    Article  PubMed  PubMed Central  Google Scholar 

  32. Giese SH, Belsom A, Sinn L, Fischer L, Rappsilber J. Noncovalently associated peptides observed during liquid chromatography-mass spectrometry and their effect on cross-link analyses. 2019.

Download references

Acknowledgements

We express our sincere gratitude to Colette Larré, Michel Zivy, and Chantal Brossard for their help to set up the deepProt project and for the enriching discussions we have shared.

Funding

This work was supported by the French National Research Agency (ANR-18-CE45-004), ANR DeepProt.

Author information

Authors and Affiliations

  1. INRAE, PROBE Research Infrastructure, BIBS Facility, 44300, Nantes, France

    Grégoire Prunier, Mehdi Cherkaoui, Albane Lysiak, Virginie Lollier, Hélène Rogniaux & Dominique Tessier

  2. INRAE, UR1268 Biopolymères Interactions Assemblages, 44316, Nantes, France

    Grégoire Prunier, Mehdi Cherkaoui, Virginie Lollier, Hélène Rogniaux & Dominique Tessier

  3. Nantes Université, CNRS, LS2N, UMR 6004, 44000, Nantes, France

    Albane Lysiak, Emile Benoist, Géraldine Jean & Guillaume Fertin

  4. Université Paris-Saclay, INRAE, CNRS, AgroParisTech, GQE - Le Moulon, PAPPSO, 91190, Gif-Sur-Yvette, France

    Olivier Langella & Mélisande Blein-Nicolas

Authors
  1. Grégoire Prunier
    View author publications

    You can also search for this author in PubMed Google Scholar

  2. Mehdi Cherkaoui
    View author publications

    You can also search for this author in PubMed Google Scholar

  3. Albane Lysiak
    View author publications

    You can also search for this author in PubMed Google Scholar

  4. Olivier Langella
    View author publications

    You can also search for this author in PubMed Google Scholar

  5. Mélisande Blein-Nicolas
    View author publications

    You can also search for this author in PubMed Google Scholar

  6. Virginie Lollier
    View author publications

    You can also search for this author in PubMed Google Scholar

  7. Emile Benoist
    View author publications

    You can also search for this author in PubMed Google Scholar

  8. Géraldine Jean
    View author publications

    You can also search for this author in PubMed Google Scholar

  9. Guillaume Fertin
    View author publications

    You can also search for this author in PubMed Google Scholar

  10. Hélène Rogniaux
    View author publications

    You can also search for this author in PubMed Google Scholar

  11. Dominique Tessier
    View author publications

    You can also search for this author in PubMed Google Scholar

Contributions

All the authors have contributed to the quality of this work (e.g. AL, GF, GJ, DT conceived the algorithm; GP, MC, AL, HR, DT adapted SpecGlob to experimental data; GP, AL, DT implemented SpecGlobX; MC, HR, EB, OL, VL, MBN tested, provided feedback and suggested improvements. All authors wrote and reviewed the manuscript).

Corresponding author

Correspondence to Dominique Tessier.

Ethics declarations

Ethics approval and consent to participate

Not applicable, human data used was publicly available.

Consent for publication

Not applicable.

Competing interests

The authors declare that they have no competing interests.

Additional information

Publisher's Note

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Supplementary Information

Additional file 1

. Complementary information concerning the evaluation of SpecGlobX. Test parameters and additional results on simulated spectra are detailed. Some spectral alignments are also presented.

Additional file 2

. Comparative interpretation of a subset of spectra by three software.

Rights and permissions

Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/. The Creative Commons Public Domain Dedication waiver (http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated in a credit line to the data.

Reprints and permissions

About this article

Check for updates. Verify currency and authenticity via CrossMark

Cite this article

Prunier, G., Cherkaoui, M., Lysiak, A. et al. Fast alignment of mass spectra in large proteomics datasets, capturing dissimilarities arising from multiple complex modifications of peptides. BMC Bioinformatics 24, 421 (2023). https://0-doi-org.brum.beds.ac.uk/10.1186/s12859-023-05555-y

Download citation

  • Received:

  • Accepted:

  • Published:

  • DOI: https://0-doi-org.brum.beds.ac.uk/10.1186/s12859-023-05555-y

Keywords

BMC Bioinformatics

ISSN: 1471-2105

Contact us