No pipeline is currently available for the automated analysis of MS PG data. Therefore, we sought to replicate a shotgun proteomics approach to create an analysis pipeline dedicated to PG analysis, referred to as ‘peptidoglycomics’ (Wheeler et al., 2014).

To limit misidentifications due to mass coincidences, we established a search strategy relying on an iterative process (Figure 2). A first search was carried out using a database made of reduced disaccharide peptides (monomers) and their theoretical monoisotopic masses. MS data were deconvoluted using the Protein Metrics Byos software to generate a list of observed monoisotopic masses alongside other parameters including retention times and signal intensity. Individual theoretical masses contained in the monomer database (Figure 2, database 1) were compared with observed masses in the experimental dataset. Any observed mass within 10 ppm tolerance was considered as a match and the corresponding inferred structure and theoretical mass were then added to a list of matched structures (Figure 2, library 1). As a second step, we used the list of matched monomers to build another database in silico (Figure 2, database 2), corresponding to dimers and trimers and their theoretical masses. Two types of polymerisation events are included in the original PGFinder version depending on the type of crosslink either through peptide stems or glycan chains. Individual theoretical masses from the in silico database were compared to observed masses to generate a list of matched dimers and trimers (Figure 2, library 2). As a third step, we combined the lists of matched monomers and multimers to generate a final library of modified muropeptides (Figure 2, library 3). The final library contained only modified muropeptides corresponding to matched monomers, dimers, and trimers. The modifications accounted for include the presence of anhydro groups, deacetylated sugars, amidated amino acids, and modifications resulting from N-acetylglucosaminidase or amidase activities (loss of GlcNAc and lack of peptide stems, respectively). In-source decay products (loss of GlcNAc) and Na⁺/K⁺ salt adducts were also added to library 3. All three libraries corresponding to observed monomers, dimers, trimers, and their modified variants were combined to search the MS data for masses matching theoretical values within a 10 ppm mass accuracy window. This search generated results processed by PGFinder to carry out a ‘clean up step’. The intensities of in-source decay products and salt adducts were combined with that from parent ions when found within close retention time (a 0.5 min time window). The output of this final step is a matched table written to a .csv format file. It contained all the inferred structures identified within the specified mass and retention time windows with an extracted-ion chromatogram (XIC) signal intensity for quantification.

The performance of the matching script was tested using the well-characterised PG from E. coli as a proof of concept. UHPLC-MS/MS data were acquired for three independent PG samples (biological replicates; Appendix 1—figure 1). Following MS1 spectral deconvolution (a process calculating masses from observed m/z values), observed masses were matched to theoretical muropeptides masses according to the strategy described above (Figure 2). A first search was carried out using a minimal mass database made of 10 simple PG fragments including three glycan chains (di-, tetra-, and hexasaccharides) and seven monomers (Table 1—source data 1). Due to the .csv format of the database, the diamino acid in position 3 could not be assigned to a symbol or Greek letter and had to be one of the 26 letters already assigned by the IUPAC-IUB Joint Commission on Biochemical Nomenclature. We used the letter J for mDAP for the initial search and replaced it by the letter m in the final table. The output of the automated search is a .csv file per dataset; all files corresponding to biological replicates were collated into one Excel file (Table 1—source data 2). Each search output contained approximately 3000 rows of masses and corresponding parameters. Depending on the dataset analysed, 41–48% of the total ion intensity was assigned to PG structures. As anticipated, inferred structures were frequently found with multiple retention times, reflecting the existence of stereoisomers, with one species accounting for most of the intensity. In some cases, observed masses matched with more than one inferred structure. The output of the automated search was consolidated as described in Supplementary file 1. Retention times were assigned to individual structures based on the elution of the most abundant stereoisomer. For example, >97% of the most abundant monomer GM-AEJA was eluted at an average retention time of 10.04 ± 0.04 min. Data consolidation revealed an unprecedented muropeptide composition complexity compared to recent LC-MS analyses of E. coli (Kühner et al., 2014; Table 1—source data 2). Sixty PG fragments were identified (Table 1): these included glycan chains lacking peptide stems (4.38%), monomers (63.14%), dimers (29.54%), and trimers (2.94%) (Figure 3). Based on the abundance of multimeric PG fragments, we report a crosslinking index of 15.69%, which is slightly lower than the value previously reported of 23.1% (Glauner, 1988).

The automated and unbiased search revealed several muropeptides that were expected but never reported to date for E. coli. These included (i) PG fragments resulting from amidase activity (4.55%), found as ‘denuded glycans’ (disaccharides and tetrasaccharides) and modified variants or muropeptide stem with an extra disaccharide residue; (ii) a low-abundance (0.23%) PG fragments containing deacetylated GlcNAc residues; and (iii) PG fragments resulting from glucosaminidase activity (0.12%). Deacetylated muropeptides were not expected since no E. coli PG deacetylase has been identified in this organism to date. All the structures identified for the first time (glycan chains, monomers containing deacetyl groups, and muropeptides lacking a GlcNAc residue) were confirmed by MS/MS analysis (Table 1—source data 3). The proportion of muropeptides with anhydroMurNAc groups identified (4.55%) was in line with previous studies, yielding an average chain length of 36.05. This value is higher than an earlier study that reported a predominant chain length of 5–10 disaccharide units (Harz et al., 1990), but agreed with recent work that reported long glycan chains in E. coli (Turner et al., 2018). Overall, the quantification of muropeptides across biological replicates was very consistent, with Pearson’s correlation coefficients >0.96 (Appendix 1—figure 2). The most pronounced variations in quantification were observed with low-abundance muropeptides accounting for less than 1% of the species identified.

We further explored the structural diversity of E. coli PG, performing a more complex search with a mass database made of glycan chains and all possible monomers containing di-, tri-, tetra-, and pentapeptide stems (Table 1—source data 4; 224 structures in total). Several monomers with tetra- and pentapeptide stems containing unusual amino acids were identified. Only four structures could be confirmed by MS/MS analysis and corresponding multimers were retained (GM-AEJF, -AEJN, AEJK, -AEJAK, and GM-AEJKD). Collectively, muropeptides containing unusual amino acids accounted for ca. 7.5% of the 80 structures identified (Table 1—source data 5) using a complex mass database.

We showed with E. coli data that PGFinder is suitable to characterise the high-resolution structure of PGs using a ‘bottom-up’ approach. However, this requires a careful analysis of the search output to confirm the identity of muropeptides identified and discriminate between multiple structures that can be assigned to a unique observed mass. A more basic application is the use of PGFinder in organisms that have already been studied in detail to either compare PG composition or quantify the abundance of specific structures. This application accounts for most PG analyses described in the literature, comparing PG structures between different isolates, isogenic mutants, or cells grown in different conditions.

We chose the PG of C. difficile as a proof of concept to demonstrate how PGFinder can carry out a straightforward comparative analysis. We prepared PG samples corresponding to biological triplicates from two clinical isolates, R20291 and M7404. To illustrate the versatility of the software, PG samples were digested by mutanolysin, and disaccharide peptides were beta-eliminated to generate lactyl-peptides (Tipper, 2002) and analysed by UHPLC-MS (Figure 4—figure supplement 1). Since the high-resolution PG structure of C. difficile has been described based on the MS/MS analysis of muropeptides (Bern et al., 2017), we used data previously published to build a PG fragment database for a ‘one-off’ matching step. The database included monomers identified by MS/MS, containing unusual amino acids and the corresponding dimers resulting from either D,D- or L,D-transpeptidation (with a AEJA or AEJ peptide as donor stem). To limit the complexity of the search output, we limited the list of trimers, tetramers, and pentamers to those containing the most abundant peptide stems found in dimers (AEJA, AEJ, or AEJG). The database of theoretical masses contained 74 PG structures described in Figure 4—source data 1. MS data were deconvoluted using Byos Feature Finder, and observed monoisotopic masses were matched to theoretical masses using PGFinder. To perform the matching operation in its simplest form, all options offered by the software were deactivated.

All monomers and dimers searched were identified, except for two muropeptides containing a peptide stem with a methionine residue in position four, both present in low abundance in C. difficile strain 630 (ca. 0.10%). 25 out of the 31 possible trimers, tetramers, and pentamers searched were found (Figure 4—source data 2). Comparison of biological replicates revealed a high reproducibility, both in retention times and quantification. A high correlation was found between biological replicates, confirming the robustness of the quantification method (Figure 4a). Both strains contained a similar amount of mono-, tri-, tetra-, and pentamers, but strain M7404 contained a significantly lower proportion of monomers and a higher proportion of dimers than strain R20291 (p=0.022 and p=0.005, respectively; Student’s t-test; Figure 4b). We next performed a Student’s t-test using permutation-based FDR to identify statistically significant differences in the abundance of individual muropeptides between the two strains. The p-value was plotted on a volcano plot against the fold change in abundance between the two samples (Figure 4c). Two muropeptides were significantly less abundant (Lac-AEJ[AG] and Lac-AEJ-Lac-AEJA), and four others were significantly more abundant in strain R20291 (Lac-AEJV- Lac-AEJA, Lac-AEJ[L/I]-Lac-AEJA, Lac-AEJAA-Lac-AEJA and the trimer (Lac-AEJA)3). These differences are likely to reflect different substrate specificities for PBPs and the Ddl ligases in these strains. Therefore, combining the output of PGFinder with statistical analysis of muropeptide abundance offers a robust workflow to identify differences in PG composition.

The analysis of E. coli PG established a proof of concept, showing that our matching script is suitable for the automated analysis of PG MS data. We next sought to evaluate the robustness of our peptidoglycomics pipeline using available datasets described in the literature. The most suitable publication was a recent study by Anderson et al. describing a PG analysis of P. aeruginosa planktonic cells (Anderson et al., 2020b). Unlike most (if not all) studies published to date, this work provided datasets from biological and technical triplicates. Unlike our E. coli samples, analysed on a Q Exactive Focus Orbitrap (Thermo), P. aeruginosa samples were analysed on an Agilent Q-TOF mass spectrometer. Spectra were deconvoluted using the Byos Feature Finder module, and observed masses were matched using a mass tolerance of 25 ppm as described by Anderson et al. To limit the occurrence of mass coincidences and misidentification at this slightly lower mass accuracy, we carried out a search with PG modifications (anhydroMurNAc residues, deacetylation, lack of peptide stem resulting from amidase activity and amidation) but only using the most frequent combinations of modifications (double anhydroMurNAc and anhydroMurNAc and deacetyl).

PGFinder identified 63 muropeptides out of the 71 reported by Anderson et al., matching our search criteria (Table 2). The eight muropeptides that were not identified were absent from the list of deconvoluted masses, indicating that the problem was not associated with the script, highlighting that the deconvolution step is a source of variability. Interestingly, the observed masses calculated using Byos Feature Finder were closer to the theoretical value (6.5 ppm versus 10.7 ppm on average), reflecting another source of variability associated with data deconvolution. Four muropeptides previously identified containing the AEJAG pentapeptide stem were matched with distinct structures due to a mass coincidence between K and AG. A careful analysis of MS/MS spectra suggested that these muropeptides contained a K residue rather than the AG dipeptide (the y2 ion being 143 ppm away from the expected mass), showing the added value of an unbiased search. This conclusion is supported by the retention times of the corresponding muropeptides since the tetrapeptide AEJK elutes before EAJA whilst the pentapeptide AEJAG elutes later in the chromatography (Bern et al., 2017). It is worth noting that our search also identified a large number of muropeptides that were not reported previously (Table 2—source data 1). These results collectively show that our PG analysis pipeline can identify an unprecedentedly large number of muropeptide structures based on MS1 data, including all those previously reported (Anderson et al., 2020a; Anderson et al., 2020b).

The automated identification of P. aeruginosa muropeptides using PGFinder and the strategy reported previously (Anderson et al., 2020a; Anderson et al., 2020b) relied on commercially available deconvolution software (ProteinMetrics Byos or Agilent MassHunter, respectively). We sought to identify a free alternative software for mass deconvolution to make our PG analysis pipeline accessible to everyone. MaxQuant was selected as a tool of choice since it represents a widely used software package to analyse high-resolution mass-spectrometric data for shotgun proteomics (Cox and Mann, 2008). MaxQuant has native support for Thermo MS data (RAW) and Sciex (WIFF) file formats and supports the open MS data format mzXML. Virtually any proprietary MS data file can be converted to the mzXML open format using freely available tools such as Proteowizard and TOPPAS, making this workflow universally applicable. As proof of concept, we converted P. aeruginosa data to an mzXML file (Appendix 2—figure 1), processed it using MaxQuant for mass deconvolution (Appendix 2—figure 2), and analysed it using PGFinder for PG structure and composition identification. We were able to identify all the expected muropeptides (Table 2). This result confirms that the automated analysis of PG datasets can be carried out using the MaxQuant freeware and our open-source script PGFinder (Table 2 and Table 2—source data 2).

E. coli BW25113 was grown at 37°C in LB under agitation (250 rpm). C. difficile strains were cultured in heart infusion supplemented with yeast extract, L-cysteine, and glucose in an atmosphere of 10% H₂, 10% CO₂, and 80% N₂ at 37°C in a Coy chamber or Don Whitley A300 anaerobic workstation.

PG was purified from exponential (E. coli) or late exponential (C. difficile) phase as described previously (Eckert et al., 2006; Glauner, 1988), freeze-dried, and resuspended in distilled water at a concentration of 5 mg/ml.

PG (1 mg) was digested overnight with 25 µg of mutanolysin at 37°C in 150 µl of 20 mM sodium phosphate buffer (pH 5.5). Soluble disaccharide peptides were recovered in the supernatant following centrifugation (20,000 × g for 20 min at 25°C). To reduce muropeptides, equal volumes (200 µl) of the solution of disaccharide peptides and of borate buffer (250 mM, pH 9.0) were mixed. 2 ml of sodium borohydride was added, and the solution was incubated for 20 min at room temperature. The pH of the solution was adjusted to 4.0 with 20% orthophosphoric acid. Beta-elimination was carried out by mixing 200 µl of muropeptides with 64 µl of 32% (w/v) ammonia. After 5 hr at 37 C, the solution was neutralised with 60 µl of acetic acid glacial, freeze-dried, and resuspended in water (Arbeloa et al., 2004; Eckert et al., 2006).

The reduced muropeptides were desalted by reverse-phase high-performance liquid chromatography (rp-HPLC) on a C18 Hypersil Gold aQ column (3 µm, 2.1 × 200 mm; Thermo Fisher) at a flow rate of 0.4 ml/min. After 1 min in water-0.1% formic acid (v/v) (buffer A), muropeptides were eluted with a 6 min linear gradient to 95% acetonitrile-0.1% formic acid (v/v). Muropeptides were freeze-dried and resuspended in 100 µl. An aliquot of the desalted samples was analysed by rp-HPLC on the same column to measure the UV absorbance of the most abundant monomer (no isocratic step, muropeptides were eluted with a 30 min linear gradient to 15% acetonitrile-0.1% formic acid [v/v]). Samples were diluted to contain 150 mAU/µl of the major monomer and 10 µl were injected. Based on the dry weight of the PG sample, we estimated that this corresponded to approximately 50 µg of material.

An Ultimate 3000 Ultra High-Performance Chromatography (UHPLC; Dionex/Thermo Fisher Scientific) system coupled with a high-resolution Q Exactive Focus mass spectrometer (Thermo Fisher Scientific) was used for LC/HRMS analysis. Muropeptides were separated using a C18 analytical column (Hypersil Gold aQ, 1.9 µm particles, 150 × 2.1 mm; Thermo Fisher Scientific), column temperature at 50°C. Muropeptides elution was performed by applying a mixture of solvent A (water, 0.1% [v/v] formic acid) and solvent B (acetonitrile, 0.1% [v/v] formic acid). After 10 µl sample injection, MS/MS data were acquired during a 40 min step gradient: 0–12.5% B for 25 min; 12.5–20% B for 5 min; held at 20% B for 5 min, and the column was re-equilibrated for 10 min under the initial conditions.

The Q Exactive Focus was operated under electrospray ionization (H-ESI II)-positive mode. Full scan (m/z 150–2250) used resolution 70,000 (FWHM) at m/z 200, with an automatic gain control (AGC) target of 1 × 10⁶ ions and an automated maximum ion injection time (IT).

Data-dependent MS/MS were acquired on a ‘Top 3’ data-dependent mode using the following parameters: resolution 17,500; AGC 1 × 10⁵ ions, maximum IT 50 ms, NCE 25%, and a dynamic exclusion time 5 s.

Byos search parameters to get .ftrs file Protein Metrics Byos v.3.9–32 was used to identify and compute the XICs. The parameters used for mass deconvolution using MaxQuant v.2.0.1.0 are described in Appendix 2—figure 2.

Code is available at https://github.com/Mesnage-Org/PGFinder (Patel, 2021). https://github.com/Mesnage-Org/PGFinder/releases/tag/v0.02 will take you to the archived release used in this paper. We used Python 3 to write the MS1 package and demonstrate its functionality using demo scripts. PGFinder can be run through an interactive Jupyter Notebook hosted on mybinder for ease of use by those less familiar with Python code. A conda environment is provided to ensure reproducible execution. Regression testing has been implemented to ensure changes to code do not cause changes to important results. The GitHub contains an interactive version to run user’s analysis and an end-to-end demo using samples data provided with the script (Interactive PGFinder). The sample data is a MaxQuant deconvolution output from the E. coli MS data analysed in the paper. The current version of the script can handle both .txt (MaxQuant) or .ftrs (Byos) deconvoluted data and offers the possibility for the user to include several modifications in the search. The time window for the ‘clean up step’ (in-source decay and salt adducts) as well as ppm tolerance for matching can also be defined by the user; the default values corresponding to these parameters used in this work are 0.5 min and 10 ppm.

All E. coli and C. difficile MS datasets generated in this study are available through the GlycoPOST repository (GPST000168; Watanabe et al., 2021). P. aeruginosa MS datasets are accessible via Figshare (Anderson et al., 2020b).