Sample preparation

Bacterial sample preparation and mass spectrometry data acquisition has been described in detail previously [15]. Handling of bacterial samples was performed in a Class II biohazard cabinet (Baker Company). Briefly, a bacterial strain was cultured on Luria-Bertani agar (ThermoFisher) overnight at 37°C in a static incubator. One to two μL of cells were harvested with a sterile 1 μL loop and transferred to 300 μL of extraction solution in a 2 mL, O-ring-lined, screw-cap microcentrifuge polypropylene microvials (Biospec Products, Bartlesville, OK). The extraction solution was either HPLC grade water (Fisher Chemical) or 33% acetonitrile (Fisher Chemical), 67% water and 0.2% trifluoroacetic acid (Sigma-Aldrich, St. Louis, MO). Approximately 30 mg of 0.1 mm diameter zirconia/silica beads (Biospec Products) were added to the tube. The tube was tightly capped and agitated with a mini-bead-beater for 2 minutes (Biospec Products). The tube was then centrifuged for 3 minutes at 13,000 rpm (Eppendorf, Germany).

Mass spectrometry

1.5 μL of sample supernatant was spotted onto 384-spot stainless steel MALDI target (Sciex, Redwood City, CA) and allowed to dry. The dried sample spot was then overlayed with 1.5 μL of a saturated solution of sinapinic acid (Life Technologies, ThermoFisher) dissolved in a solution of 33% acetonitrile, 67% water and 0.2% trifluoroacetic acid. Redissolved sample with matrix was then allowed to dry.

MS and MS/MS data was collected on a 4800 MALDI-TOF-TOF mass spectrometer (Sciex, Redwood City, CA) equipped with a pulsed solid-state YAG laser (λ = 355 nm, τ = 5 ns) with a 200 Hz repetition rate. MS data was collected in linear mode. After a brief delay (~1 μs) following the laser pulse, ions were accelerated from the source at 20.0 kV after which they strike the linear detector. The m/z range was 2000 to 20,000. MS data was collected, summed and signal averaged from 1000 laser shots. MS linear mode was externally calibrated with the +1 and +2 charge states of cytochrome-C, myoglobin and lysozyme (Sigma-Aldrich, St. Louis, MO).

MS/MS-PSD data was collected in reflectron mode wherein after a brief delay (~300 ns) following the laser pulse, ions were accelerated from the source at 8.0 kV. Upon reaching the timed-ion selector or TIS (a mass gate that selects the precursor ion based on its m/z and thus its arrival time), the selected precursor ion transits the TIS gate unimpeded where ions arriving outside the TIS window too soon or too late, are blocked. A typical TIS window is manually set to the precursor mass ± 100 Da. The TIS window was narrowed further, when necessary, to exclude fragment ions from neighboring protein ions if present. After the TIS, the mass-selected precursor ion was then decelerated to 1.0 kV after which it enters the collision cell. As no collision gas was introduced into the collision cell, any fragmentation is due to post-source decay (PSD), i.e. delayed fragmentation resulting from internal energy acquired by the ion during the ionization/desorption process in the source. After the collision cell, fragment ions and unfragmented precursor ion were re-accelerated to 15.0 kV. A metastable suppressor (another mass gate) was used to block any unfragmented precursor ion from advancing to the reflectron mirror to increase the detection sensitivity of fragment ions. Fragment ions were reflected nearly 180° by a 2-stage reflectron mirror: mirror #1: 10.515 kV and mirror #2: 18.330 kV) after which ions strike the reflectron detector. The MS/MS m/z range spans from 9.0 to above (+500 to 1000) the m/z of the precursor ion. MS/MS data was collected, summed and signal averaged from 10,000 laser shots. MS/MS reflectron mode was externally calibrated with the PSD fragment ions of singly charged alkylated thioredoxin.

Data was viewed using Data Explorer® software (Version 4.9, Sciex, Redwood City, CA). Raw MS/MS data was processed in the following sequence: Advanced baseline correction (Baseline correction parameters: Peak width: 32; Flexibility: 0.5; Degree: 0.0), Noise removal (Std dev to remove: 2.00) and Gaussian smoothing (Filter width: 31 points). The processed MS/MS data was then centroided and exported as an ASCII spectrum consisting of two columns of data: m/z and absolute intensity. Processed and centroided MS/MS data are provided at https://github.com/jpark837/PSD.

Extraction of protein properties

The protein properties analyzed in this work are sequence and structurally based. We used Alphafold2 (version 2.2.0) to predict the structure of each of the bacterial proteins using the default databases [30]. We then selected bacterial proteins that were pre-identified for which MS/MS-PSD data was available. We wrote a pipeline in python to extract 14 properties for each instance of either a D-, E-, or N-residue from the proteins. We used PyMol (Schrödinger) to count the number of intramolecular backbone and sidechain hydrogen bonds, as well as to check for a salt bridge presence for each residue instance. For hydrogen bonds, we considered electrostatic pairings of the protonated lysine (K) and arginine (R) residues with deprotonated aspartic acid (D) and glutamic acid (E) residues. We chose a bond length range under 4.0 Å for salt bridges [31].

Secondary structure assignment and relative solvent accessible surface area calculations were done using the DSSP program [32]. The remaining numerical properties (degree, clustering coefficient, closeness, betweenness, eigenvector centrality, eccentricity, average nearest neighbor degree and strength) are centrality measurements from residue interaction network (RIN) analysis [33]. We used the Network Analysis or Protein Structure (NAPS) webserver for prediction and centrality analysis of the RIN for each protein [34]. For the NAPS webserver, we used the following options: C-alpha network type, weighted, threshold of 0–7 Å, and residue separation of 1. For comparison between networks, we adjusted eccentricity to be normalized to the protein diameter [34]. The protein diameter is the maximum eccentricity value of the network.

Alphafold2 predicted protein structures and the code used to extract the structural properties and accompanying data are available at https://github.com/jpark837/PSD.

Computational and statistical analyses

All Alphafold2 predictions were run on a GPU node through the USDA-ARS Scientific Computing Initiative (SCINet) Ceres high-performance computing (HPC) cluster.

All statistical analyses and plot generation was done using Python and R.

For multivariate regression analysis, we assumed the response variable Y (signal score) to follow a negative binomial distribution with a mean of E[Y] = μ and let x_p be a set of explanatory variables (extracted properties). μ is then related to the explanatory variables as Eq 1. We scaled the explanatory variables from 0–1 for comparative interpretation before fitting the linear model to our data containing the signal score and property values for each fragment.

(1)

To analyze the significance of the categorical properties (secondary structure, N-terminal adjacent residue, C-terminal adjacent residue and salt bridge presence), we performed the Kruskal-Wallis test to check if any groups within each property deviates significantly. We then performed the pairwise Mann-Whitney U test to identify the group within each categorical property that was significantly different.

For analysis of the categorical properties (N-terminal adjacent residue and C-terminal adjacent residue), we used all 36 bacterial proteins, as they only depend on the protein sequence. For the remaining categories, we removed 3 bacterial proteins that had a poor average predicted local distance difference test (pLDDT) score below 70 (S1 Table), as these properties depend on the predicted protein structure from Alphafold2.

Calculation of signal scores

We selected 36 bacterial proteins for which MS/MS data was available for analysis (S1 Table and at https://github.com/jpark837/PSD). A typical example of MS and MS/MS data is shown in Fig 1 wherein a protein biomarker is identified from its intact mass by MS and its characteristic fragment ions obtained by MS/MS. Each protein in our study was previously identified by top-down proteomic analysis and confirmed by manual inspection comparing observed fragment ions to that of in silico fragment ions of the identified protein sequence. The aspartic acid effect is the dominant fragmentation mechanism of low charge state protein ions that fragment by PSD. Subsequently, the most prominent fragment ions are the result of backbone cleavage on the C-terminal side of D-, E- and N-residues and on the N-terminal side of P-residues, resulting in characteristic backbone b-type and y-type fragment ions. Isobaric protein ions, i.e. protein ions that have the same nominal m/z and are thus not isolatable from each other by our TIS mass gate, would result in a mixture of fragment ions from both protein ions. Such a circumstance was not observed in the 36 proteins analyzed in this study. All the fragment ions of each MS/MS experiment corresponded to a single protein sequence.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Example MS data of a of a strain of Salmonella enterica subsp. enterica serovar infantis.

(A) Linear MS data of bacterial cell lysate. (B) The identified protein sequence (hypothetical/YahO) after removal of its 21-residue signal peptide. An asterisk denotes a site of backbone cleavage with its corresponding b-type and/or y-type fragment ions. (C) MS/MS data of the protein ion at m/z 7666. Fragment ions are identified by m/z (theoretical value in parentheses) and their b- or y-type fragment ion designation. (D) The pre-processed and centroided MS/MS data of the protein ion at m/z 7666. Pre-processed and centroided MS/MS data is shown in Fig 1D.

https://doi.org/10.1371/journal.pone.0299287.g001

The raw MS/MS data for each protein was processed, centroided and exported as an ASCII spectrum and analyzed (Fig 2). GPMAW (version 13.03) was used to predict the average m/z of b- and y-type fragment ions resulting from in silico backbone cleavage on the C-terminal side of D-, E and N-residues for each protein sequence [35]. In silico fragment ions generated by GPMAW are provided at https://github.com/jpark837/PSD. Our script then matched each predicted fragment ion to the highest signal intensity of the MS/MS data within ± 5 m/z. The script also accounted for loss of ammonia (-17 m/z) and water (-18 m/z) for each fragment ion to separate noise from background as much as possible. Once fragment signals were assigned and separated, our script compared the b- and y-type fragment ion intensity for each backbone cleavage position, then considered the larger of the two as the fragment signal (u).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Workflow for analysis of the selected bacterial proteins.

https://doi.org/10.1371/journal.pone.0299287.g002

For each fragment signal, we used Eq 2 to calculate a signal score. The signal score, which we defined as the ratio of the intensity of the fragment signal (u) and the standard deviation (σ) of the background (Eq 2), was our metric for fragmentation efficiency. A higher signal score indicates a higher likelihood of polypeptide backbone cleavage at that residue position, as the resulting fragment ion is more abundant. The standard deviation of the background was to normalize varying noise between MS/MS data.

(2)

Backbone cleavage at E and N-residues have similar efficiencies

Initially, we noticed the distribution of our response variable, the signal score of each fragment, to overlap each other for E- and N-residues (Fig 3A). Plots of the empirical cumulative distribution function (eCDF) of signal scores for D-, E- and N-residues confirmed this observation, as we also saw the eCDFs of E- and N-residues to overlap (Pearson’s correlation coefficient = 0.99) (Fig 3B). This overlap indicates that E- and N-fragments have a similar spread of signal scores. In contrast, the eCDF of D-residues was distinct from E- and N-residues in that they were shifted towards the right, as a larger proportion of D-fragments have higher signal scores. For instance, ~56% of D-residue fragments have a signal score higher than 10 while for E- and N-residues, only ~15% of fragments do (Fig 3B). Together, our results suggest that polypeptide backbone cleavage on the C-terminal side of E- and N-residues have similar efficiencies and are lower than D-residues.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Distribution of D-, E-, and N-fragment signal scores.

(A) Histogram of D-, E-, and N- fragment scores. (B) Empirical cumulative distribution functions of D-, E-, and N-fragment scores.

https://doi.org/10.1371/journal.pone.0299287.g003

Regression analyses reveal several centrality measures to be significant factors

We also noticed that the signal score for all residues were non-normal and heavily positively skewed (Fig 3A). This shape is characteristic of count-based data, for which there exist discrete probability distributions that provide convenient models for analysis [36,37]. We rationalized that by viewing the polypeptide backbone cleavage as an event with a probability of success, we could apply these types of models for our case [36]. The clustering of signal scores of D-, E- and N-fragments near 0, alongside extreme outliers at high signal scores, indicates overdispersion (Fig 3A). For protein properties which were numerical (Table 1), we consequently used negative binomial regression to assess the effect of each property on the signal score. The negative binomial distribution allows its variance to differ from its mean, allowing greater flexibility in handling dispersion [38].

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Multivariate regression analysis results.

https://doi.org/10.1371/journal.pone.0299287.t001

A cross correlation matrix of the explanatory variables showed degree and strength to be strongly correlated with each other, as the pairwise Pearson’s r correlation coefficient between them was 1 (S1 Fig). We subsequently removed strength as an explanatory variable from our regression analysis to reduce redundancy. Our regression results for D-,E-, and N-residues are summarized in Table 1. We found various centrality measurements from residue interaction network (RIN) analysis to be significant. In RIN analysis, proteins are drawn as a network, where residues are considered as nodes while contacts between them are considered as edges [34].

For D- residues, relative solvent accessibility, closeness, and eccentricity were significant (p<0.01) explanatory variables. Relative solvent accessibility describes how exposed or buried a residue is in a protein and is an important factor for determining its stability [39,40]. The positive value suggests that for D-, the less buried the residue is, the higher the signal score probability is up to a certain extent. D-fragments with relative solvent accessibility values that were in the 50–75% quartile had the highest distribution of signal scores. (Fig 4A).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Distribution of D-, E-, and N- fragment signal scores with respect to significant explanatory variables.

(A-C) Empirical cumulative distribution functions (eCDF) of D-fragment signal scores. (D) eCDF of N-fragment signal scores. (E-I) eCDF of E-fragment signal scores. Shades of pink represent percentile of significant explanatory variable values.

https://doi.org/10.1371/journal.pone.0299287.g004

Closeness is defined as the inverse of the shortest path distance (dist(u,v)) of a node (n) to all other nodes (v) (Eq 3). Closeness is an indicator of how close a node (residue) is to all other nodes in the network [34]. A positive coefficient estimate for closeness indicates that residues near other residues path wise are associated with a higher signal score probability, which we also clearly observed in its distribution (Fig 4B).

(3)

Eccentricity is defined as the shortest path distance of the residue to the farthest residue divided by the diameter of the protein (Eq 4). A higher value indicates the residue is closer to the periphery while a lower value indicates the residue is closer to the center [41]. The significant, positive coefficient estimate (p<0.01) for eccentricity indicates that D-residues that are closer to the periphery of the protein, but not at its absolute extremity leads to a higher signal score probability. For eccentricity, D-fragments with values that were the lowest 0–25% and the highest 75–100% quartiles had lower distribution of signal scores compared to those within the 25–50% and 50–75% quartiles (Fig 4C).

(4)

For E-sidechain hydrogen bond count, closeness, eccentricity, eigenvector centrality, and average nearest neighbor degree were significant (p<0.05) explanatory variables. Sidechain hydrogen bond count is the number of potential hydrogen bonds the sidechain of a residue is involved in within a bond length range between 2.5 and 3.2 Å. E-residues with the highest number of sidechain hydrogen bond counts (75–100% quartile) had the highest distribution of fragment signal scores (Fig 4E). Like D-, E-residues also had a positive coefficient estimate and distribution pattern for closeness (Fig 4F). Similarly for eccentricity, E-fragments that were the highest 75–100% quartiles had the highest distribution of signal scores (Fig 4G).

Eigenvector centrality is the eigenvector (x_i) that corresponds to the largest eigenvalue (λ) of the adjacency matrix (A_ij) [34,42] (Eq 5). This centrality metric indicates how connected a node is to other well-connected nodes in the network [34]. The negative coefficient estimate is reflected in its distribution, where E-fragments with eccentricity values in the 25–50% quartiles had the highest distribution of signal scores (Fig 4H).

(5)

Average nearest neighbor degree is the average of the degree (C_d(u)) of a node’s direct neighbors (N(u)) (Eq 6) [34]. This centrality metric quantifies the dependency between degrees of a node and its neighbors [43]. Although the variable was significant (p<0.05) and its coefficient estimate was positive (Table 1), we did not see a clear pattern upon visual inspection of the distribution of E-fragment signal scores with respect to average nearest neighbor degree (Fig 4I).

(6)

For N-residues, only eigenvector centrality was a significant explanatory variable (p<0.05). The coefficient estimate for this variable was positive (Table 1). However, we saw that N-fragments with degree values in the lower 0–25% and 25–50% quartiles had higher distributions of signal scores (Fig 4D), indicating a negative relationship. The lack of significant explanatory variables closeness and eccentricity of N- compared to D- and E- is also interesting. The presence of an amide rather than a carboxylic acid on the side chain may present different behaviors regarding the aspartic acid effect.

Presence of an adjacent C-terminal proline enhances fragmentation

We also analyzed four categorical properties, where we found the C-terminal adjacent residue to be a significant explanatory variable for all three residues (Table 2). The D-G, D-N, D-P, E-L, E-G, N-L, and N-P sequence motifs were found to be significant (p<0.05). Except for the E-L and N-L sequence motifs, the rest led to a higher signal score (Fig 5A–5C). We noticed that when P was present on the C-terminal side of D- and N-residues, the signal score of the fragments were dramatically higher. Indeed, for P-residue fragment ions, the presence of either a D-, E-, or N-residue on the N-terminal side significantly (p<0.00001) led to a higher signal score (48.1 ± 20.1). In contrast, P-residue fragment ions that did not have an adjacent N-terminal D-, E-, or N-residues had a lower signal score of 3.6 ± 1.0 (Fig 5D). E-residue alone did not show the E-P sequence motif to be significant, presumably because there was only one instance of the sequence motif in our dataset.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 5. Analysis of categorical explanatory variables.

(A) Box plot of fragment signal scores grouped by C-terminal residues adjacent to D. (B) Box plot of fragment signal scores grouped by C-terminal residues adjacent to E. (C) Box plot of fragment signal scores grouped by C-terminal residues adjacent to N. (D) Bar graph comparing the proline fragment signal scores whose adjacent N-terminal adjacent residue was D,E,N or non-D,E,N. (E) Box plot of fragment signal scores grouped by secondary structure of E-residues. Significant explanatory variables p<0.05, p<0.01, p<0.00001 is respectively marked by *,**, and ***** based on the Mann-Whitney U test. Bar graph is displayed as mean ± standard error.

https://doi.org/10.1371/journal.pone.0299287.g005

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Kruskal-Wallis test of categorical explanatory variables.

https://doi.org/10.1371/journal.pone.0299287.t002

For glutamic acid, the secondary structure assignment of the residue was also significant (Table 2). T, which stands for turn and designates single helix hydrogen bonds in DSSP, lead to a significantly higher signal score (Fig 5E). In contrast, H, which stands for a 4 residue-turn alpha helix, was significantly lower (Fig 5E) [32,44].