Skip to main content
NCBI home page
NIHPA Author Manuscripts logoLink to NIHPA Author Manuscripts
. Author manuscript; available in PMC: 2024 Dec 2.
Published in final edited form as: Clin Chem. 2022 Nov 3;68(11):1459–1470. doi: 10.1093/clinchem/hvac147

Identifying Clinically Relevant Bacteria Directly from Culture and Clinical Samples with a Handheld Mass Spectrometry Probe

Sydney C Povilaitis 1, Ashish Chakraborty 1, Lindsey M Kirkpatrick 2, Rachel D Downey 3, Sarmistha B Hauger 3,4, Livia S Eberlin 5,*
PMCID: PMC11610516  NIHMSID: NIHMS2033294  PMID: 36103272

Abstract

Background:

Rapid identification of bacteria is critical to prevent antimicrobial resistance and ensure positive patient outcomes. We have developed the MasSpec Pen, a handheld mass spectrometry-based device that enables rapid analysis of biological samples. Here, we evaluated the MasSpec Pen for identification of bacteria from culture and clinical samples.

Methods:

A total of 247 molecular profiles were obtained from 43 well-characterized strains of 8 bacteria species that are clinically relevant to osteoarticular infections, including Staphylococcus aureus, Group A and B Streptococcus, and Kingella kingae, using the MasSpec Pen coupled to a high-resolution mass spectrometer. The molecular profiles were used to generate statistical classifiers based on metabolites that were predictive of Gram stain category, genus, and species. Then, we directly analyzed samples from four patients, including surgical specimens and clinical isolates, and used the classifiers to predict the etiologic agent.

Results:

High accuracies were achieved for all levels of classification with a mean accuracy of 93.3% considering training and validation sets. Several biomolecules were detected at varied abundances between classes, many of which were selected as predictive features in the classifiers including glycerophospholipids and quorum-sensing molecules. The classifiers also enabled correct identification of Gram stain type and genus of the etiologic agent from three surgical specimens and all classification levels for clinical specimen isolates.

Conclusions:

The MasSpec Pen enables identification of several bacteria at different taxonomic levels in seconds from cultured samples and has potential for culture-independent identification of bacteria directly from clinical samples based on the detection of metabolic species.

Keywords: bacteria identification, MasSpec Pen, mass spectrometry, infectious disease

INTRODUCTION

Rapid and accurate pathogen identification is critical to allow selection of targeted antibiotic treatment and improve outcomes for patients with bacterial infections. Empiric broad-spectrum antibiotic regimens can lead to adverse effects including allergic reactions, bacterial resistance, and opportunistic infections including Clostridium difficile colitis(1). Specific and targeted antimicrobial regimens can only be administered if the pathogen is accurately identified. Primary methods for bacterial identification rely on isolating bacteria from specimens by culturing them in media prior to biochemical testing, which can take 1–5 days (2, 3). Many acute bacterial infections develop in sterile sites in the human body, such as osteoarticular infections (OAIs), which are most often caused by hematogenous spread of bacterial pathogens to bone and joints in children and direct inoculation or contiguous spread in adults (4, 5). Certain pathogens such as Kingella kingae, a prevalent cause of pediatric OAIs, are difficult to culture and isolate, which often results in the absence of an isolate thus precluding bacterial identification and susceptibility testing (6). Alternative molecular methods, such as polymerase chain reaction (PCR), have been applied for culture-independent identification of pathogens from patient specimens (3). PCR provides high specificity and accuracy for pathogen identification, but PCR assays require specific reagents and user expertise, which often entails off-site analysis in specialized laboratories.

Several mass spectrometry (MS) technologies have been proposed for clinical identification of pathogens (7–16). The most prominent and commercially-available clinical approach is matrix-assisted laser desorption/ionization-MS (MALDI-MS), which allows rapid (<5 min) and accurate (>90%) bacterial identification of cultured isolates based on the detection of ribosomal protein profiles. MALDI has been used to characterize hundreds of bacterial species through decades of research development (7). Culture-independent MALDI-MS workflows have been developed for urine and cerebrospinal fluid; however, additional sample preparation steps including centrifugation and chemical extraction are required to remove non-bacterial interferents prior to analysis. MALDI-MS is not amenable to three-dimensional samples, and thus, bacterial smears, thin tissue sections, or dried fluid spots have been more commonly analyzed (7, 17–19).

We previously reported the development of the MasSpec Pen (MS Pen) technology as a handheld device integrated to a mass spectrometer for direct molecular analysis of biomolecules of unmodified clinical samples (20, 21). Analysis with the MS Pen is rapid (<30 s), and requires no sample preparation, harsh solvents, or applied voltages. The device is handheld, geometry independent, and autoclavable, which facilitates use by medical professionals within the clinical environment (22, 23). The rich metabolic information acquired with the MS Pen can be used to build robust statistical models (24) that enable classification of human tissues (21), ovarian cancer (20), and meat types (25). Here, we describe a novel application of the MS Pen for rapid identification of bacterial species that are common etiologic agents in OAIs, including the difficult-to-culture K. kingae.

MATERIALS AND METHODS

Clinical samples

Five de-identified samples from four patients (Table 1) were obtained from the clinical microbiology laboratory at Dell Seton Medical Center (DSMC) in accordance with Institutional Review Board protocol #2020–05-0013. Three samples were surgically collected from two patients as part of routine care. The specimens were evaluated by DSMC laboratory using standard clinical culture techniques followed by identification of bacteria using MALDI-TOF-MS. The two other samples were clinical isolates that were re-plated to nutrient agar slants at DSMC for the purpose of our study. All samples were also cultured on blood agar plates in our research laboratory and isolates of these laboratory cultures were analyzed using the MS Pen (Supplementary Material file).

Table 1.

Patient demographic and clinical sample information.

Patient ID Age Sample(s) Diagnosis Site of Infection Bacteria ID
1 64 Synovial fluid Septic arthritis and tibiofemoral osteomyelitis with MRSA bacteremia Right femur synovial fluid MRSA
2 19 Sacral tissue Stage 5 decubitus ulcer Sacral bone MRSA
Tissue smear
3 54 Clinical isolate on agar slant Cellulitis, right lower extremity Fibula Group B Streptococcus
4 63 Clinical isolate on agar slant Septic arthritis with MRSA secondary to prosthetic knee infection Left knee synovial fluid MRSA
Open in a new tab

Bacteria, Culture Conditions, and Sample Generation

Strains of S. aureus, S. epidermidis, Streptococcus pyogenes (Group A Streptococcus), Streptococcus agalactiae (Group B Streptococcus), K. kingae, P. aeruginosa, E. coli, and S. enterica were obtained from American Type Tissue Collection (ATCC) and BEI Resources. Strain designation, serotype, and other information are listed in Supplementary Table 1. Bacteria were received as frozen cultures in 0.5X Tryptic Soy Broth supplemented with 10% glycerol or freeze dried. Freeze dried bacteria were cultured in Tryptic Soy Broth overnight and aliquots were flash frozen with 10% DMSO to prepare frozen stocks which were stored at −80°C until use. For each batch of analyses, bacteria from frozen stock were streaked on blood agar and cultured overnight (12–16 h) at 37°C. For MS Pen analysis of agar cultures, colonies were removed from the agar plate with a sterile plastic inoculating loop and smeared onto a glass slide. To compare molecular profiles of bacteria grown on MacConkey and chocolate agar, bacteria were prepared similarly (Supplementary Material file).

For MS analysis of clinical samples, 10–20 μL of sample was aliquoted onto a glass slide using a sterile inoculating loop and allowed to dry at room temperature in a Class II biological safety cabinet (~5–10 min). An additional aliquot was streaked on blood agar and cultured for 24 h at 37°C. Resulting colonies were prepared for analysis in the same manner as described above.

MS Pen Instrumentation and Analysis

All procedures were performed in accordance with IBC protocol #2018–00224. Bacterial smears were analyzed in seven batches with the MS Pen system in random order. The MS Pen is comprised of a handheld sampling probe, microcontroller unit, and mass spectrometer (21). All analyses were performed on a ThermoFisher Scientific QExactive HF Orbitrap mass spectrometer in negative-ion mode from m/z 120–1500 with a resolving power of 120,000 and inlet capillary temperature of 350°C. Additional experimental details can be found in the Supplementary Material file.

Data Processing and Statistical Analysis

Mass spectral data was extracted to .csv and imported to R for data pre-processing. The least absolute shrinkage and selection operator (lasso) algorithm (24) (glmnet package) was used to generate classifiers. Model performance was evaluated using leave-one-out cross-validation on training sets and prediction validation and test sets. Volcano plots and one way analysis of variance (ANOVA) were also used to evaluate the data (26). The Supplementary Material file contains additional details regarding data processing. All metabolites described herein have been annotated using high mass accuracy and/or tandem MS and are thus putative annotations per the reporting standards proposed by Sumner et al (27).

RESULTS

A total of 237 MS Pen analyses were performed on 43 strains of bacteria from eight species that are relevant to OAIs (Supplementary Fig. 1, Supplementary Tables 1 and 2). Water was selected as the solvent for analysis because it maximized molecular coverage and S/N of biomolecules when compared to other evaluated solvents, and yielded a reproducibility of 12% relative standard deviation (%RSD), which was comparable to values previously reported for tissue analyses (8% RSD) (Supplementary Material file) (21). The molecular profiles obtained in the negative-ion mode presented a range of small metabolites, lysophospholipids, and glycerophospholipids (Fig. 1), which were used to generate statistical classifiers (24).

Fig. 1.

Fig. 1.

Open in a new tab

Representative mass spectra of Group A and B Streptococcus, S. aureus, S. epidermidis, K. kingae, P. aeruginosa, E. coli, and S. enterica from m/z 120-900. Ions putatively annotated as small metabolites (green), quorum-sensing molecules (red), PG lipids (blue), PE lipids (orange), and rhamnolipids (purple) are labeled with colored dots.

Gram stain category

Identification of Gram stain category is a critical first step in the treatment of patients with bacterial infections (28). The molecular profiles obtained using the MS Pen from G- bacteria differed qualitatively from G+ bacteria, specifically within the lipid profile range. In G+ bacteria, phosphatidylglycerols (PG) with odd- and even- fatty acyl chains were predominantly detected, whereas both phosphatidylethanolamine (PE) and PG lipids with even fatty acyl chains were predominantly detected in the mass spectra of G- bacteria (Supplementary Fig. 2). Volcano plot analysis revealed that 281 ions differed significantly (FC>2, p-adj<0.05) between G- and G+ molecular profiles (Fig. 2a). These include ions putatively annotated as acetyl-aspartic acid [M-H]– (m/z 174.040), phosphocholine [M-H₂O-H]– (m/z 165.055), and acetyl-asparagine [M-H]– (m/z 173.057), which were observed at a higher relative abundance in G+ bacteria, as well as deoxy-manno-octulosonate [M-H]– (m/z 237.061), which was observed at a higher relative abundance in G- bacteria.

Fig. 2.

Fig. 2.

Open in a new tab

Lasso classification performance and volcano plot analysis for (A) gram+ vs gram−, and (B) Staphylococcus vs Streptococcus. For volcano plots, dashed lines indicate cutoffs for features with a normalized intensity FC >2 (vertical) and adjusted (adj.) P value <0.05 (5% FDR) (horizontal). Features with a higher normalized intensity in g− or Staphylococcus are shown on the upper left (blue), and features with a higher normalized intensity in g+ or Streptococcus are shown on the upper right (green) of the volcano plots.

The lasso classifier generated to distinguish G- and G+ bacteria (Fig. 2a, train/test split shown in Supplementary Table 1) was based on 34 predictive features including small metabolites and phospholipids. Features weighted towards G- bacteria include ureidoglycine [M+Cl]– and phospholipids PG 14:0_14:1 [M-H]– (m/z 663.424) and PE 16:1_18:1 [M-H]– (m/z 714.508). Features weighted towards G+ bacteria include phosphocholine [M-H₂O-H]– (m/z 165.055), orotidine [M-H]– (m/z 287.051), and glycineamideribotide (GAR) [M-H]– (m/z 285.052) (Supplementary Fig. 3a). The classifier performed with accuracies of 95.7% in the training set (nstrain = 28, ntotal = 163), and 95.9% in the validation set (nstrain=15, ntotal=74). Notably, four analyses of S. gordonii, a Viridans group Streptococcus spp., were used as an independent test strain and were all classified correctly, demonstrating that the model could be predictive of G+ species not included in the training set (Supplementary Fig. 4).

Staphylococcus vs. Streptococcus genera

Staphylococcus (Staph.) and Streptococcus (Strep.) spp. are etiologic bacterial agents for many diseases, including OAIs (6). Because some antibiotics are only active against one genus or some species in one genus, genus-level classification is important to inform selection of more targeted antibiotics (29). Through MS Pen analyses, we found that a total of 116 ions differed significantly between the analyzed Staphylococcal and Streptococcal molecular profiles based on volcano plot analysis (FC>2, p-adj<0.05) including taurine [M-H]– (m/z 124.006) and acetyl-glutamine [M-H]– (m/z 187.071), which were detected at a significantly higher relative abundance in the analyzed Staphylococcus spp., and adenosine diphosphate ribose [M-H₂O-H]– (m/z 540.054) and glutathione [M-H]– (m/z 306.077), which were detected at significantly higher relative abundance in the analyzed Streptococcus spp. (Fig. 2b).

The mass spectra from 138 analyses of 20 strains were used to generate a lasso classifier to distinguish Staph. and Strep. spp. (Fig. 2b). The classifier was composed of seventeen molecular features, including taurine [M-H]– (m/z 124.006) and phosphoglycolic acid [M-H]– (m/z 154.974) weighted towards analyzed Staph. spp. and glutathione [M-H]– (m/z 306.077) and PG 16:0_18:1 [M-H]– (m/z 747.519) weighted towards analyzed Strep. spp. (Supplementary Fig. 3b). The classifier performed with a high accuracy of 92.4% in the training set (nstrain=15, ntotal=105) and 100% in the validation set (nstrain=5, ntotal=43). Four analyses of S. gordonii, were used as an independent test set and all classified correctly, demonstrating that this model can be predictive of Streptococcus spp. not included in the training set (Supplementary Fig. 4).

Group A vs. Group B Streptococcus species

Group A Streptococcus (GAS) and Group B Streptococcus (GBS) are both common causes of OAIs in children younger than five (6). Although most antibiotics are active against both species, the time required for bactericidal activity for beta lactam antibiotics is much longer for GBS when compared to GAS (30); thus, species-level identification is beneficial to inform therapeutic decisions. While the molecular profiles obtained using the MS Pen from the closely related Group A and B Streptococcus spp. appeared to be qualitatively similar (Fig. 1), 57 molecular features differed significantly between these two groups, including glutathione [M-H]– (m/z 306.077) and biotin [M-H]– (m/z 243.081) weighted towards GBS and acetyl-asparagine [M-H]– (m/z 173.057) weighted towards GAS (Fig. 3a).

Fig. 3.

Fig. 3.

Open in a new tab

Lasso classification performance and volcano plot analysis for (A) Group A vs Group B Streptococcus, and (B) S. aureus vs S. epidermidis. For volcano plots, dashed lines indicate cutoffs for features with a normalized intensity FC >2 (vertical) and adjusted P value <0.05 (5% FDR) (horizontal). Features with a higher normalized intensity in GAS or S. aureus are shown on the upper left (blue), and features with a higher normalized intensity in GBS or S. epidermidis are shown on the upper right (green) of the volcano plots.

The lasso classifier generated from 67 analyses of 10 strains (Fig. 3a) yielded a high accuracy of 97.1% in the training set (nstrain=5, ntotal=35), and accuracy of 84.8% in the validation set (nstrain=5, ntotal=32). The lower accuracy achieved in the validation set is primarily due to a single strain of GBS (ATCC 51487) for which 3 out of 4 analyses were misclassified as GAS. The classifier for GAS vs. GBS. is comprised of three molecular features weighted towards GBS, including glutathione [M-H]– (m/z 306.077) (Supplementary Fig. 3c).

S. aureus vs. S. epidermidis

Within Staphylococcus species, S. aureus and S. epidermidis are common causes of OAIs. Widespread resistance and variable anti-microbial activity to these species may affect therapeutic options (6, 31). Molecular profiles obtained from S. aureus and S. epidermidis appeared to be qualitatively similar, yet 65 features differed significantly between the species as identified by volcano plot analysis (Fig. 1, Fig. 3b). Many of these ions were identified as small metabolites including acetyl-aspartic acid [M-H]– (m/z 174.040) and acetyl-tyrosine [M-H]– (m/z 222.076), which were at a significantly higher relative abundance in S. epidermidis, and pentose phosphate [M-H]– (m/z 421.075) and PG 15:0_18:0 [M-H]– (m/z 735.519), which were present at higher relative abundances in S. aureus. (Fig. 3b). The lasso classifier generated from 70 analyses of 10 strains (Fig. 3b) yielded an accuracy of 94.3% in the training set (nstrain=5, ntotal=35), and accuracy of 97.1% in the validation set (nstrain=5, ntotal=35). Six features were selected for the model, all weighted towards S. aureus, including hexosamine [M-H]– (m/z 178.010), pentose phosphate [M-H]– (m/z 421.075), and ¹³C isotope of PG 16:0_18:1 [M-H]– (m/z 749.526) (Supplementary Fig. 3d).

Gram negative bacteria

Gram negative spp. comprise a large proportion of pathogenic bacteria. Several Gram negative bacteria, including K. kingae, P. aeruginosa, S. enterica, and E. coli, are prevalent causes of OAIs.(6) Varied molecular profiles were observed between the species analyzed. Notably, K. kingae, a pathogen that is difficult to culture and a common cause of OAI in young children, presented rich lipid profiles with a suite of PE’s and PG’s observed (Fig. 1). In the mass spectra of P. aeruginosa, 12 alkyl-quinolone quorum sensing molecules were detected, including Pseudomonas quorum sensing signal (PQS) [M+Cl]– (m/z 294.127), heptyl-quinolone [M-H]– (m/z 242.115), and undecylquinolone (UDQ) [M-H]– (m/z 298.127). From a one-way ANOVA, 441 ions were significantly different between the groups (p-adj < 0.05, 5% FDR), many of which were exclusive to one species. For example, ions identified as alkyl-quinolone quorum sensing molecules, six rhamnolipids, and the iron chelator pyochelin [M-H]– (m/z 323.053) were only observed in P. aeruginosa, while several phospholipids with 16:1, 16:0, or 14:0 acyl chains were observed exclusively in K. kingae. Ions identified as acetyl-methionine [M-H]– (m/z 190.052), PE 16:0_17:1 (m/z 702.509), PG 16:0_17:1 (m/z 733.503) were observed in the enterobacteria E. coli and S. enterica but not from other G- species. Furthermore, a few ions were detected in all species except one. For example, PE and PG lipids with 16:0 and 18:1 acyl chains including PE 16:0_18:1 [M-H]– (m/z 716.524) and PG 16:0_18:1 [M-H]– (m/z 747.519), were observed in all species except K. kingae, possibly reflecting a different membrane fatty acid composition in this relatively-uncharacterized organism.

The lasso classifier that was generated from 101 analyses of 23 strains yielded accuracies of 84% in the training set (nstrain=17, ntotal=77) and 91.7% in the validation set (nstrain=6, ntotal=24) (Fig. 4). Forty-four features were selected for the model including quorum sensing molecules PQS and UDQ, pyochelin, and rhamnolipid Rha-C12-C10 [M-H]– (m/z 531.534), weighted towards P. aeruginosa and phospholipids PE 14:0_14:0 [M-H]– (m/z 634.446), and PG 14:0_14:0 [M-H]–, (m/z 665.441), weighted towards K. kingae. Features weighed towards S. enterica include hexosamine [M-H]– (m/z 178.071), and phospholipids PE 16:0_17:1 [M-H]– (m/z 702.509) and PG 16:0_18:1 [M-H]– (m/z 747.519). Lysophospholipids LPE 17:1 [M-H]– (m/z 464.278) and LPG 20:4 [M-H]– (m/z 455.241), acetyl-methionine [M-H]– (m/z 190.052), and PG 14:0_18:1 [M-H]– (m/z 719.488), PG 16:0_16:1 [M-H]– (m/z 719.488) were among features weighted towards E. coli. (Supplementary Fig. 3e).

Fig. 4.

Fig. 4.

Open in a new tab

(A), Lasso classification performance; and (B), ANOVA analysis results for gram-negative species.

Effect of growth media

We also evaluated the effect of three growth media on the molecular profiles detected and on the performance of lasso classification for S. aureus and E. coli. (Supplementary Material file). Detailed results are included in the Supplementary Material file and Supplementary Fig. 5.

Identification of bacteria directly from clinical specimens and isolates

We next analyzed five clinical samples obtained from four patients diagnosed with OAIs as an independent test set of samples to evaluate if the classifiers trained on bacterial smears from the commercially-acquired strains could identify bacteria directly from patient surgical specimens without culture and from clinical isolates (Table 1). Consistent with the high prevalence of S. aureus in OAIs (6), the etiologic agent of the three specimens was determined to be methicillin-resistant Staphylococcus aureus (MRSA) by the DCMC microbiology laboratory. The etiologic agent for the clinical isolated plated on agar slants was clinically determined to be GBS and MRSA.

Fig. 5 shows the molecular profiles obtained for each clinical specimen and their corresponding isolates that were cultured in our research laboratory (2–3 replicates per sample). The ions detected from laboratory-cultured isolates were consistent with the mass spectra acquired from the commercially-sourced S. aureus described above, including GAR [M-H]– (m/z 285.053), acetyl-glutamine [M-H]– (m/z 187.071), and PG 15:0_18:0 [M-H]– (m/z 735.519). All laboratory-cultured isolates generated from clinical specimens predicted correctly as Gram positive, Staphylococcus, and S. aureus using the Gram stain type, Staph. vs. Strep. and S. aureus vs. S. epidermidis classifiers, respectively.

Fig. 5.

Fig. 5.

Open in a new tab

Representative molecular profiles and lasso classification results for clinical specimens and corresponding laboratory isolates from 3 patients infected with MRSA. The symbols shown in molecular profiles denote which lasso features were present and the class that the feature was weighted toward. (orange: gram+ vs gram−, blue Staphylococcus vs Streptococcus, green S. aureus vs S. epidermidis) Lasso classification results for the Gram stain type, Staph. vs Strep., and S. aureus vs S. epidermidis classifiers are shown to the right of each spectrum with a majority or minority of replicates classifying correctly denoted by a green check or red “x”, respectively, and a yellow asterisk indicates that no lasso features of either class were observed in a majority of replicates. For additional information on clinical specimens and corresponding isolates, please see the Supplemental Material file.

The mass spectra obtained directly from patient specimens were complex and varied between samples due to the diversity of molecular composition between human synovial fluid and tissue. Several lasso features in our Gram stain type and Staph. vs. Strep. classifiers were detected in the mass spectra of each sample, including taurine, glutathione, and PGs 15:0_17:0 and 16:0_18:1 (Fig. 5), in addition to ions previously described as characteristic of human tissue and biofluids such as hexose [M+Cl]– (m/z 215.033), heme [M-H]– (m/z 615.170), and glycerophosphoserine 36:1 [M-H]– (m/z 788.545) (20, 21). The Gram stain type for all clinical specimens and the genus of the synovial fluid and sacral tissue were correctly predicted. Lasso features from the S. aureus vs. S. epidermidis classifier were detectable (above S/N = 3) in only one replicate of the synovial fluid and tissue smear samples, both of which were correctly classified as S. aureus, and undetected in the sacral tissue sample (Fig. 5). Lastly, the mass spectra obtained from direct MS Pen analysis of the two clinical isolates on agar slants enabled correction prediction of the bacteria Gram stain type and genus, but species-level molecular predictors were not detectable (above S/N = 3) in the mass spectra of the clinical slants, and thus, these samples could not be classified at the species-level.

DISCUSSION

In this study, we demonstrate the use of the MS Pen technology to discriminate eight bacterial species prevalent in OAIs on the Gram stain type, genus, and species level based on metabolic profiles that are detected in seconds and with minimal sample preparation requirements. Further, we highlight the potential of this technique to acquire molecularly predictive information for Gram stain type and genus of etiologic agents directly from patient specimens without culture isolation.

Lasso statistical classifiers were built using mass spectral data obtained from 237 analyses of 43 strains. The lasso algorithm is advantageous when building classification models from untargeted metabolite analysis because it selects only a sparse set of the ions detected in the training set as predictive markers. As such, the models are more interpretable and generally robust since they are not substantially affected by unexpected background ions or mass spectral fluctuations that may be observed between datasets (24, 32). The lasso algorithms developed in this study provided high mean accuracies between training and validation sets. The high accuracies for G+ genus and G+ species are comparable to accuracies obtained for MALDI-TOF MS (7, 33, 34), although more modest for G- than the reported accuracy for MALDI-TOF MS. Of note, the lower accuracy for the G- classifier was largely due to misclassification of S. enterica as E. coli, and the lower accuracy for the GAS vs. GBS classifier was largely due to misclassification of a single strain of GBS (ATCC 51487). These are unsurprising results considering that these species are closely related. As such, future efforts will be dedicated to expanding our datasets and refining our statistical models to improve robustness and prediction accuracy for closely-related species that are highly relevant in the clinical setting. Nevertheless, the accuracies achieved with the MS Pen on bacterial smears are arguably comparable to reported accuracies for 16S rRNA PCR methods (35). Importantly, the correct classification of the test species S. gordonii for the Gram stain category and Staph. vs. Strep. models indicates that the lasso models are predictive of G+ and Streptococcus spp. and may be generalizable to species that were not included in training sets. Yet, further investigation and independent testing are needed to rigorously validate the models with a larger sample and variety of species.

Rich molecular profiles were observed from all eight species investigated, many of which have been previously reported using other ambient MS methods (11, 16, 33, 34). Ninety-five unique molecular features were selected as lasso features across the five models and 411 ions were determined to be significantly different between groups. Metabolites involved in phospholipid biosynthesis such as phosphocholine, glycerol phosphate, and glycerophosphoethanolamine were weighted towards G+ bacteria, Strep. and Staph., respectively, potentially reflecting biological differences in bacterial cell wall structure and assembly (36). Features annotated as metabolites in purine and pyrimidine biosynthesis pathways were present in every classifier, such as GAR, cytosine, and ureidoglycine. A possible mechanism leading to this observation could be different rates of purine biosynthesis or the presence of alternative de novo purine biosynthesis enzymes in some bacteria. GAR, which was weighted towards G+ and Staph., is regulated by fused genes in many G+ bacteria, but for E. coli, S. enterica, or P. aeruginosa, the genes are not fused and an alternative GAR transformylase is available (37). Additionally, a suite of PE and PG phospholipids significantly differed between species and were predictive of bacterial identity reflecting membrane composition differences between G+ and G- bacteria and inter-species differences (36). Quorum-sensing molecules that signal bacterial density and trigger biofilm formation, and rhamnolipids, which have been reported as markers of P. aeruginosa virulence, were also observed and selected as lasso features (38). The observation of these molecules could allow for determination of P. aeruginosa virulence in addition to identification. Interestingly, while most of the ions selected as lasso features were unique to each classifier, some features including GAR, glutathione, and PG 16:0_18:1, were present in multiple classifiers. This observation may result from the presence/absence of biochemical pathways regulating these molecules for higher levels of taxonomy (i.e., Gram stain type or genus). Additionally, acetyl-methionine was selected as a lasso feature for the G- classifier and other acetylated amino acids, including acetyl-tyrosine and acetyl-glutamine, were significantly more abundant in some classes of bacteria. Protein acetylation at lysine residues, and occasionally histidine, serine, and threonine residues, has been established as an important post-translational modification for metabolic regulation (39, 40). To the best of our knowledge, acetylation of the amino acids detected in this study has not been previously reported, and these findings could illuminate novel routes of bacterial metabolic regulation.

Regarding limitations of the MS Pen as a metabolic based method, factors such as growth phase and media can alter bacterial metabolism, which may influence the predictive metabolic profiles obtained with the MS Pen. However, in our study, 100% accuracy was achieved with our species-level classifier built from bacteria grown on blood agar when applied to the data acquired from S. aureus cultured on chocolate agar and E. coli cultured on MacConkey agar, despite the expected low similarity scores between the overall mass spectra profiles (Supplementary Material file). Thus, while a thorough evaluation of the effect of media on lasso features for all bacteria is warranted, the sparse set of predictive features selected by the lasso models may be sufficiently robust to classify bacteria cultured on different media. It also must be noted that the identification of metabolites described in our study are putative, requiring more investigation to substantiate claims of the biological significance of the predictive features (27). In addition, the relatively brief signal (seconds) obtained using the MS Pen compared to other MS and ambient ionization techniques limits the ability to perform extensive structural characterization of ions.

Importantly, the classifiers trained on bacterial smears enabled accurate prediction of Gram stain type and genus of the etiologic agent from the molecularly complex synovial fluid and sacral tissue samples without culture isolation. The detection of predictive molecular features directly from tissue and joint fluid specimens supports that the MS Pen could be further developed for culture-independent bacterial identification directly from patients’ surgical specimens through an analysis that requires minimal pre- and post-analytical steps. While the classifiers performed well at the Gram stain and genus levels, the predictive features for the S. aureus vs. S. epidermidis classifier were detected in a minority of the analyses. The absence of species-predictive features could be a result of ion suppression and/or suggest that our models must include MRSA-related molecular features to robustly detect MRSA in clinical samples, as none of the strains used to train statistical models were methicillin resistant. Nevertheless, correct species identification was achieved for all analyses for which predictive metabolic markers were detected, which demonstrates that the models developed enable species-level identification when markers are detectable. It must be noted that the number and types of clinical samples included in this study was limited, and thus, our results must be further validated on additional clinical samples and species.

From a clinical perspective, the rapid analysis time (<30 s) with the handheld MS Pen and the high prediction accuracies obtained with statistical classifiers built from bacterial smears show the potential of the MS Pen for identification of bacteria from culture and directly from clinical specimens based on metabolic profiles. By accelerating analysis time to seconds and eliminating sample preparation steps associated with matrix application and in vacuum analyses employed in MALDI, identification of bacteria with the MS Pen could provide an attractive alternative MS-based method for clinical use. Although the scope of strains included in this study are limited relative to the extensive range of bacteria for the well-developed MALDI-TOF systems, the coverage and predictive performance of the MS Pen could be further improved with expansion of the bacteria library. In addition, further exploration of specific classifiers for resistance genes in both G+ and G- bacteria may increase identification accuracy and provide early guidance for anti-microbial selection. While the limitations warrant further investigation and technical development, our results establish the MS Pen technology as a rapid and accurate platform for bacterial identification based on metabolites detected from bacterial isolates and directly from clinical surgical tissue samples.

Supplementary Material

Supporting Information

ACKNOWLEDGMENTS

We thank ATCC, BEI Resources, and Dr. Olja Simoska for providing the samples. We thank Michael Keating for independently performing statistical analysis to verify results. L.S. Eberlin is grateful to Dr. Laura Baer for insightful discussions.

Research Funding:

This work was supported by the NIH/NICHD under award number R21HD106614, the NCI/NIH under award number 1R33CA229068–01A1, the Gordon and Betty Moore Foundation through grant GBMF8049, and the Welch Foundation (F-1895 to L.S. Eberlin).

Role of Sponsor:

The funding organizations played no role in the design of study, choice of enrolled patients, review and interpretation of data, preparation of manuscript, or final approval of manuscript.

Nonstandard Abbreviations:

OAI

osteoarticular infection

MS

mass spectrometry

MALDI-TOF

matrix assisted laser desorption ionization – time of flight

MS Pen

MasSpec Pen

K. kingae

Kingella kingae

Staph

Staphylococcus

Strep

Streptococcus

S. aureus

Staphylococcus aureus

S. epidermidis

Staphylococcus epidermidis

DSMC

Dell Seton Medical Center

GAS

Group A Streptococcus

GBS

Group B Streptococcus

P. aeruginosa

Pseudomonas aeruginosa

E. coli

Escherichia coli

S. enterica

Salmonella enterica

ATCC

American Type Tissue Collection

IBC

Institutional Biosafety Committee

m/z

mass-to-charge ratio

S/N

signal to noise ratio

TIC

total ion count

Lasso

least absolute shrinkage and selection operator

S. gordonii

Streptococcus gordonii

FC

fold change

FDR

false discovery rate

p-adj

adjusted p-value

ANOVA

analysis of variance analysis

G-

Gram-negative

G+

Gram-positive

RSD

relative standard deviation

PG

phosphatidylglycerol

PE

phosphatidylethanolamine

GAR

glycine amide ribotide

PQS

Pseudomonas quinolone signal

UDQ

undecylquinoline

Rha

rhamnolipid

Footnotes

Employment or Leadership: L.S. Eberlin, chief scientific officer of MS Pen Technologies; L.M. Kirkpatrick, Research Affairs Committee; S.B. Hauger, Texas Pediatric Society Co Chair Infectious Disease Committee.

Consultant or Advisory Role: None declared.

Stock Ownership: L.S. Eberlin, shareholder in MS Pen Technologies.

Honoraria: L.S. Eberlin, MD Anderson Cancer Center; S.B. Hauger, CDC Presentation on Covid 19 and Infection Prevention Texas Pediatric Society April 2021.

Expert Testimony: None declared.

Patents: S.C. Povilaitis, L.M. Kirkpatrick, and L.S. Eberlin are inventors in US Patent No. 10,643,832 and/or in other patent applications related to the MS Pen Technology licensed by the University of Texas to MS Pen Technologies, Inc. L.S. Eberlin, US2012312979-A1, US2012295276-A1, US2013273560-A1, U.S. Serial Nos. 62/383,234, 62/411,321, and 62/462,524, WO2011072130-A1, US Patent No. 10,643,832.

Other Remuneration: L.S. Eberlin, royalties/licenses from Purdue Research Foundation; L.M. Kirkpatrick, support for attending meetings and/or travel from Indiana University School of Medicine.

Authors’ Disclosures or Potential Conflicts of Interest: Upon manuscript submission, all authors completed the author disclosure form. Disclosures and/or potential conflicts of interest:

REFERENCES

  • 1.Maxson T, Mitchell DA. Targeted treatment for bacterial infections: prospects for pathogen-specific antibiotics coupled with rapid diagnostics. Tetrahedron 2016;72:3609–24. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 2.Váradi L, Luo JL, Hibbs DE, Perry JD, Anderson RJ, Orenga S, Groundwater PW. Methods for the detection and identification of pathogenic bacteria: past, present, and future. Chemical Society Reviews 2017;46:4818–32. [DOI] [PubMed] [Google Scholar]
  • 3.Maugeri G, Lychko I, Sobral R, Roque ACA. Identification and antibiotic-susceptibility profiling of infectious bacterial agents: a review of current and future trends. Biotechnol J 2018;14:1700750. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 4.Yi J, Wood JB, Creech CB, Williams D, Jimenez-Truque N, Yildirim I, et al. Clinical epidemiology and outcomes of pediatric musculoskeletal infections. J Pediatr 2021;234:236–44.e2. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 5.Romanò CL, Romanò D, Logoluso N, Drago L. Bone and joint infections in adults: a comprehensive classification proposal. European Orthopaedics and Traumatology 2011;1:207. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 6.Schmitt SK. Osteomyelitis. Infect Dis Clin North Am 2017;31:325–88. [DOI] [PubMed] [Google Scholar]
  • 7.Hou TY, Chiang-Ni C, Teng SH. Current status of MALDI-TOF mass spectrometry in clinical microbiology. J Food Drug Anal 2019;27:404–14. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 8.Braga PAC, Tata A, Gonçalves dos Santos V, Barreiro JR, Schwab NV, Veiga dos Santos M, et al. Bacterial identification: from the agar plate to the mass spectrometer. RSC Adv 2013;3:994–1008. [Google Scholar]
  • 9.Cameron SJS, Bodai Z, Temelkuran B, Perdones-Montero A, Bolt F, Burke A, et al. Utilisation of Ambient Laser Desorption Ionisation Mass Spectrometry (ALDI-MS) improves lipid-based microbial species level identification. Sci Rep 2019;9. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 10.Cody RB, McAlpin CR, Cox CR, Jensen KR, Voorhees KJ. Identification of bacteria by fatty acid profiling with direct analysis in real time mass spectrometry. Rapid Commun Mass Spectrom 2015;29:2007–12. [DOI] [PubMed] [Google Scholar]
  • 11.Hsu CC, ElNaggar MS, Peng Y, Fang JS, Sanchez LM, Mascuch SJ, et al. Real-time metabolomics on living microorganisms using ambient electrospray ionization flow-probe. Anal Chem 2013;85:7014–8. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 12.Jarmusch AK, Pirro V, Kerian KS, Cooks RG. Detection of strep throat causing bacterium directly from medical swabs by touch spray-mass spectrometry. Analyst 2014;139:4785–9. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 13.Bolt F, Cameron SJS, Karancsi T, Simon D, Schaffer R, Rickards T, et al. Automated high-throughput identification and characterization of clinically important bacteria and fungi using rapid evaporative ionization mass spectrometry. Anal Chem 2016;88:9419–26. [DOI] [PubMed] [Google Scholar]
  • 14.Kocurek KI, Stones L, Bunch J, May RC, Cooper HJ. Top-down LESA mass spectrometry protein analysis of Gram-positive and Gram-negative bacteria. J Am Soc Mass Spectrom 2017;28:2066–77. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 15.Havlikova J, May RC, Styles IB, Cooper HJ. Direct identification of bacterial and human proteins from infected wounds in living 3D skin models. Sci Rep 2020;10:11900. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 16.Watrous J, Roach P, Heath B, Alexandrov T, Laskin J, Dorrestein PC. Metabolic profiling directly from the petri dish using nanospray desorption electrospray ionization imaging mass spectrometry. Anal Chem 2013;85:10385–91. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 17.Arroyo MA, Denys GA. Parallel evaluation of the MALDI Sepsityper and Verigene BC-GN assays for rapid identification of Gram-Negative bacilli from positive blood cultures. J Clin Microbiol 2017;55:2708–18. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 18.Ferreira L, Sánchez-Juanes F, González-Ávila M, Cembrero-Fuciños D, Herrero-Hernández A, González-Buitrago JM, Muñoz-Bellido JL. Direct identification of urinary tract pathogens from urine samples by matrix-assisted laser desorption ionization-time of flight mass spectrometry. J Clin Microbiol 2010;48:2110–5. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 19.Segawa S, Sawai S, Murata S, Nishimura M, Beppu M, Sogawa K, et al. Direct application of MALDI-TOF mass spectrometry to cerebrospinal fluid for rapid pathogen identification in a patient with bacterial meningitis. Clinica Chimica Acta 2014;435:59–61. [DOI] [PubMed] [Google Scholar]
  • 20.Sans M, Zhang J, Lin JQ, Feider CL, Giese N, Breen MT, et al. Performance of the MasSpec pen for rapid diagnosis of ovarian cancer. Clin Chem 2019;65:674–83. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 21.Zhang J, Rector J, Lin JQ, Young JH, Sans M, Katta N, et al. Nondestructive tissue analysis for ex vivo and in vivo cancer diagnosis using a handheld mass spectrometry system. Sci Transl Med 2017;9:eaan3968. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 22.Zhang J, Sans M, DeHoog RJ, Garza KY, King ME, Feider CL, et al. Clinical translation and evaluation of a handheld and biocompatible mass spectrometry probe for surgical use. Clin Chem 2021;67:1271–80. [DOI] [PubMed] [Google Scholar]
  • 23.King ME, Zhang J, Lin JQ, Garza KY, DeHoog RJ, Feider CL, et al. Rapid diagnosis and tumor margin assessment during pancreatic cancer surgery with the MasSpec Pen technology. Proceedings of the National Academy of Sciences 2021;118:e2104411118. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 24.Tibshirani R Regression shrinkage and selection via the Lasso. J R Stat Soc Series B Stat Methodol 1996;58:267–88. [Google Scholar]
  • 25.Gatmaitan AN, Lin JQ, Zhang J, Eberlin LS. Rapid analysis and authentication of meat using the MasSpec Pen technology. J Agric Food Chem 2021;69:3527–36. [DOI] [PubMed] [Google Scholar]
  • 26.Chong J, Wishart DS, Xia J. Using MetaboAnalyst 4.0 for comprehensive and integrative metabolomics data analysis. Curr Protoc Bioinformatics 2019;68:e86. [DOI] [PubMed] [Google Scholar]
  • 27.Sumner LW, Amberg A, Barrett D, Beale MH, Beger R, Daykin CA, et al. Proposed minimum reporting standards for chemical analysis. Metabolomics 2007;3:211–21. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 28.Zgurskaya HI, López CA, Gnanakaran S. Permeability barrier of Gram-begative cell envelopes and approaches to bypass it. ACS Infect Dis 2015;1:512–22. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 29.Krause KM, Serio AW, Kane TR, Connolly LE. Aminoglycosides: an overview. Cold Spring Harb Perspect Med 2016;6:a027029. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 30.Schauf V, Deveikis A, Riff L, Serota A. Antibiotic-killing kinetics of group B streptococci. J Pediatr 1976;89:194–8. [DOI] [PubMed] [Google Scholar]
  • 31.Baldwin JN, Strickland RH, Cox MF. Some properties of the beta-lactamase genes in Staphylococcus epidermidis. Appl Microbiol 1969;18:628–30. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 32.Tarazona S, Arzalluz-Luque A, Conesa A. Undisclosed, unmet and neglected challenges in multi-omics studies. Nature Computational Science 2021;1:395–402. [DOI] [PubMed] [Google Scholar]
  • 33.Dean SN, Walsh C, Goodman H, van Hoek ML. Analysis of mixed biofilm (Staphylococcus aureus and Pseudomonas aeruginosa) by laser ablation electrospray ionization mass spectrometry. Biofouling 2015;31:151–61. [DOI] [PubMed] [Google Scholar]
  • 34.Bardin EE, Cameron SJS, Perdones-Montero A, Hardiman K, Bolt F, Alton EWFW, et al. Metabolic phenotyping and strain characterisation of Pseudomonas aeruginosa isolates from cystic fibrosis patients using rapid evaporative ionisation mass spectrometry. Sci Rep 2018;8:10952. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 35.Srinivasan R, Karaoz U, Volegova M, MacKichan J, Kato-Maeda M, Miller S, et al. Use of 16S rRNA gene for identification of a broad range of clinically relevant bacterial pathogens. PLOS ONE 2015;10:e0117617. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 36.Sohlenkamp C, Geiger O. Bacterial membrane lipids: diversity in structures and pathways. FEMS Microbiol Rev 2015;40:133–59. [DOI] [PubMed] [Google Scholar]
  • 37.Chua SMH, Fraser JA. Surveying purine biosynthesis across the domains of life unveils promising drug targets in pathogens. Immunol Cell Biol 2020;98:819–31. [DOI] [PubMed] [Google Scholar]
  • 38.Barr HL, Halliday N, Cámara M, Barrett DA, Williams P, Forrester DL, et al. Pseudomonas aeruginosa quorum sensing molecules correlate with clinical status in cystic fibrosis. Eur Respir J 2015;46:1046–54. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 39.Christensen DG, Xie X, Basisty N, Byrnes J, McSweeney S, Schilling B, Wolfe AJ. Post-translational protein acetylation: an elegant mechanism for bacteria to dynamically regulate metabolic functions. Front Microbiol 2019;10:1604. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 40.Carabetta VJ, Cristea IM, Margolin W. Regulation, function, and detection of protein acetylation in bacteria. J Bacteriol 2017;199:e00107–17. [DOI] [PMC free article] [PubMed] [Google Scholar]

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Supplementary Materials

Supporting Information
NIHMS2033294-supplement-Supporting_Information.docx (2MB, docx)

RESOURCES