Early cell-wall stress signatures enable 20-min ATR-FTIR discrimination of MRSA and MSSA Authors: Wilbur Hudson, Michael Nelson, Caroline C Taylor, Alexander Marchesani, Diyali Sil, Chayan Dutta, Eric S Gilbert, Gary Hastings, Yi Jiang Published: 2026 ABSTRACT AbstractRapid identification of methicillin-resistant Staphylococcus aureus (MRSA) is crucial for early optimization of antibiotic treatment, but current routine susceptibility testing typically requires 48–72 h. Attenuated total reflectance Fourier-transform infrared (ATR-FTIR) spectroscopy has emerged as a promising approach for bacterial identification and has recently been used to distinguish MRSA from methicillin-sensitive S. aureus (MSSA) after 60–120 min of β-lactam exposure. Here, we test whether ATR-FTIR can resolve MRSA versus MSSA within the first hour of antibiotic challenge. We exposed three MSSA (ATCC 6538, WKZ1, RN4220) and four MRSA (ATCC 43300, USA300-JE2, WKZ2, CA629) strains to sub-MIC ampicillin (0.5 μg/mL) and acquired spectra from 800 to 1800 cm−1 at 0, 20, 30, and 60 min. We compared classification pipelines based on the full spectrum, PCA-reduced features, and LASSO-selected bands, coupled with linear discriminant analysis, partial least-squares discriminant analysis, and support vector machines. Models based on LASSO-selected features achieved the strongest early performance, with strain-aware classification accuracies of 0.91 at 20 min and 0.92 at 30 min. Leave-one-strain-out cross-validation (LOSO-CV) further showed that focusing on mechanistically relevant difference spectra enabled robust across-strain discrimination, with balanced accuracies of 0.91 at 20 min and 0.90 at 30 min. The most informative early bands mapped primarily to peptidoglycan and carbohydrate precursor regions, while later discrimination increasingly involved lipid-associated bands. Transmission electron microscopy and atomic force microscopy at 20 min independently confirmed antibiotic-induced cell-wall thickening and structural disruption in susceptible strains but not in resistant strains. Together, these results establish a proof of concept that early cell-wall stress signatures captured by ATR-FTIR, combined with lightweight and interpretable machine-learning models, can deliver rapid and accurate phenotypic discrimination between MRSA and MSSA.Subject terms: Biotechnology, Microbiology Introduction Rapid identification of antimicrobial resistance remains a major challenge in clinical microbiology. Among Gram-positive pathogens, methicillin-resistant Staphylococcus aureus (MRSA) is associated with increased morbidity and mortality compared with methicillin-susceptible S. aureus (MSSA)1–3. Early optimization of antibiotic therapy is therefore critical, yet routine culture-based susceptibility testing typically requires 48–72 h, leaving clinicians to make initial treatment decisions with incomplete information4. Fourier-transform infrared (FTIR) spectroscopy provides a promising route toward faster phenotypic assessment because it captures holistic biochemical fingerprints of cells, and has been applied to bacterial identification and antimicrobial susceptibility assessment5–8. In attenuated total reflectance (ATR) mode, FTIR is particularly sensitive to the bacterial cell envelope, making it well suited for probing antibiotic-induced changes in cell-wall composition9. Recent studies have shown that ATR-FTIR can distinguish bacterial phenotypes and, after prolonged antibiotic exposure, can differentiate MRSA from MSSA with high accuracy. However, most prior work has focused on measurements obtained after 60–120 min or longer7,10,11, leaving open the question of whether clinically useful discrimination is possible during the much earlier response window. That early window is mechanistically important for β-lactam antibiotics. β-lactams inhibit peptidoglycan cross-linking, which in susceptible strains rapidly produces cell-wall stress, altered precursor accumulation, and structural remodeling before overt growth arrest or lysis becomes dominant. MRSA, by contrast, typically maintains cross-linking via the mecA-encoded penicillin-binding protein PBP2a (1MWS), whereas MSSA exhibits early defects in cell-wall cross-linking under exposure9,12,13. These differences suggest that very early ATR-FTIR measurements may capture mechanistically specific phenotypic responses that separate MSSA from MRSA before large differences in bulk growth emerge. Here, we test that hypothesis using seven strains of S. aureus, comprising three MSSA strains (ATCC 6538, WKZ1, RN4220) and four MRSA strains (ATCC 43300, USA300-JE2, WKZ2, CA629), exposed to sub-MIC ampicillin. We acquired ATR-FTIR spectra in the 800–1800 cm range at 0, 20, 30, and 60 min after exposure and compared three analysis strategies: models trained on the full spectrum, on PCA-reduced representations, and on LASSO-selected bands. Because high apparent accuracy on small spectral datasets can be inflated by strain-specific variation, we combined standard classification analyses with leave-one-strain-out validation to distinguish within-strain reproducibility from across-strain generalization. We further used transmission electron microscopy (TEM) and atomic force microscopy (AFM) at early time points to test whether the discriminative spectral features were consistent with antibiotic-induced cell-wall changes. Our results indicate that accurate MRSA/MSSA discrimination is achievable within 20–30 min of β-lactam exposure and that the most informative early signatures map primarily to peptidoglycan- and carbohydrate-associated spectral regions consistent with early cell-wall stress. Orthogonal TEM and AFM measurements support this interpretation by revealing structural changes in susceptible but not resistant strains at matched timescales. Rather than relying on opaque high-capacity models, this study emphasizes sparse, interpretable feature selection and stringent validation, providing a mechanistically grounded proof of concept for rapid phenotypic discrimination of MRSA and MSSA by ATR-FTIR. Results Early growth divergence reveals phenotypic effects of ampicillin within 20 min We examined seven S. aureus strains, three MSSA (ATCC 6538, RN4220, WKZ1) and four MRSA (ATCC 43300, USA300-JE2, WKZ2, CA629), at 0, 20, 30, and 60 min after exposure to ampicillin. Ampicillin produced rapid and differential effects on growth in MSSA and MRSA. In the absence of antibiotic, all seven strains showed comparable increases in optical density over 60 min. Under ampicillin treatment, however, the MSSA strains showed early lag or reduced OD accumulation relative to untreated control, whereas MRSA strains generally continued to increase over the same interval (Fig. 1). This divergence became apparent by 20 min after antibiotic exposure. Although OD alone does not sufficient for resistance classification, these data establish that antibiotic-induced phenotypic divergence begins by 20–30 min, motivating our focus on early time points for FTIR analysis and classification. Early growth responses of MSSA and MRSA strains to ampicillin. Optical density at 600 nm (OD600) was measured for three MSSA strains, ATCC6538, RN4220, and WKZ1 (a), and four MRSA strains, ATCC 43300, USA300 JE2, WKZ2, and CA629 (b), at 0, 20, 30, and 60 min after exposure to ampicillin (0.5 μg/mL). Dashed lines indicate untreated controls and solid lines indicate ampicillin-treated cultures. MSSA strains show early stagnation or decline in OD600 under ampicillin, whereas MRSA strains continue to increase in OD600 over the same interval. Sample-level cross-validation overestimates discrimination of unseen strains To quantify how well ATR-FTIR spectra can distinguish MRSA from MSSA over time, we compared three feature representations: the full preprocessed spectrum (487 features), PCA-reduced features (first 10 principal components), and LASSO-selected bands. Each representation was paired with four classifiers: linear discriminant analysis (LDA), partial least squares–discriminant analysis (PLS-DA), linear support vector machine (linear SVM), and radial basis function SVM (RBF SVM). Across all feature representations and classifiers, discrimination accuracy increased after antibiotic exposure compared with 0 min (Fig. 2). Models trained on the full spectrum or PCA-reduced features typically achieved mean accuracies of about 0.65–0.75 at 20 and 30 min, with similar or slightly higher values at 60 min (Fig. 2a,b). In contrast, models trained on LASSO-selected bands performed substantially better. For example, the linear SVM achieved an accuracy of 0.93 (95% CI 0.89–0.97) at 20 min with a similarly high value at 30 min, while slightly lower (0.90) at 60 min(Fig. 2c). Similar patterns were observed across the other classifiers. Sample-level classification performance across time and feature representations. Accuracy of MRSA/MSSA classification at 0, 20, 30, and 60 min using four classifiers (LDA, PLS-DA, linear SVM, RBF SVM) under a strain-aware cross-validation scheme. Bars show mean accuracy and black lines indicate 95% confidence intervals. (a) Models trained on all preprocessed wavenumbers. (b) Models trained on the first 10 principal components from PCA. (c) Models trained on LASSO-selected wavenumbers. LASSO-based models achieve the strongest early discrimination, with peak accuracies at 20–30 min and no significant improvement at 60 min. The overlap in the 95% confidence intervals between 20 and 30 min indicates that these time points offer statistically equivalent discrimination under this evaluation scheme. Because the 30-min performance is more consistent across methods, it represents a practical target for rapid testing. Notably, linear classifiers performed as well as or better than nonlinear alternatives, indicating that the relevant class structure is close to linearly separable once informative bands are selected. However, this sample-level evaluation does not fully test whether the learned discrimination generalizes to previously unseen strains. Even with strain-aware balancing, models trained on all available strains may still exploit stable inter-strain differences that are unrelated to the antibiotic response of interest. To assess true across-strain generalization, we performed leave-one-strain-out cross-validation (LOSO-CV), in which all samples from one strain were excluded from training and used only for testing. Under LOSO-CV, performance dropped substantially at early time points, particularly for the closely related WKZ1 (MSSA) and WKZ2 (MRSA) strains (Fig. 3). Mean accuracy across held-out strains was 0.61 at 20 min and 0.71 at 30 min, compared with much higher values under sample-level cross-validation. These results indicate that conventional classification pipelines partly rely on strain-specific spectral structure and therefore overestimate performance on unknown strains. This observation motivates a more targeted analysis focused on spectral changes that reflect the shared phenotypic response to ampicillin. Leave-one-strain-out cross-validation (LOSO-CV) reveals limited early generalisation across unseen strains. Heatmap showing LOSO-CV accuracy for each held-out strain at 0, 20, 30, and 60 minafter ampicillin exposure. Values above each column indicate mean accuracies across all held-out strains at that time point. Performance is lowest at early time points, especially for the closely related WKZ1 (MSSA) and WKZ2 (MRSA) strains, demonstrating that sample-level evaluation overestimates generalization to previously unseen strains. LASSO-selected bands implicate early cell-wall stress as the key discriminatory mechanism To identify the biochemical basis of early discrimination, we inspected the bands selected by LASSO at each time point (Fig. 4). At 0 min, the mean MSSA and MRSA spectra are nearly indistinguishable, consistent with the absence of antibiotic effects. After 20–60 min of ampicillin exposure, subtle but reproducible differences emerged, especially in cell-wall-, carbohydrate-, and amide-associated regions. Time-resolved FTIR signatures and LASSO-selected discriminatory bands. (a) Mean second-derivative spectra (800–1800 cm−1) for combined MSSA (red) and combined MRSA (blue) at 0, 20, 30, 60 min after ampicillin exposure. (b) Zoomed-in sections of representative spectral regions showing time-dependent differences between MRSA and MSSA. (c) LASSO feature stability across wavenumbers and time points. Bar height indicates the percentage of bootstrap runs in which each band was selected; a red dashed line marks the 20% stability threshold. Frequently selected bands at early times cluster in cell-wall and carbohydrate regions, with lipid-associated bands becoming prominent by 60 min. Data are shown for n = 16 MRSA spectra and n = 12 MSSA spectra per time point. LASSO stability analysis highlights which of these bands contribute most consistently to MRSA/MSSA discrimination. At 0 min, the LASSO features show the natural variance between strains. We focus on regions not associated with these non-mechanistic differences. At 20 min, the most stable features clustered around 1494–1506 cm−1 and 1022 cm−1. These bands correspond to regions associated with peptidoglycan and carbohydrate precursor accumulation, consistent with early cell-wall stress in susceptible strains under β-lactam exposure10,11. By 30 min, the selected set expands to include amide-associated bands, suggesting broader structural changes in proteins and cell-wall–associated components. At 60 min, LASSO selection increasingly shifted toward the tyrosine-associated band at 1510 cm−1, consistent with later membrane perturbation9. The earlier cell-wall and carbohydrate precursor bands at 1027 cm−1 remained present, but less prominent. Overall, the temporal evolution of LASSO-selected features is consistent with the expected progression of β-lactam action: early disruption of cell-wall synthesis, followed later by broader envelope and membrane-associated changes. These results support early cell-wall stress as the dominant biochemical signature distinguishing susceptible from resistant strains in the first 20–30 min and suggest that a mechanism-focused analysis may generalize better across strains than models trained on the full spectrum. TEM and AFM confirm early cell-wall remodeling in MSSA To validate the cell-wall–centric interpretation suggested by the FTIR and LASSO analyses, we performed transmission electron microscopy (TEM) and atomic force microscopy (AFM) imaging on representative MSSA and MRSA strains after 20 min of ampicillin exposure. In MSSA, TEM revealed thickened and irregular cell walls, disrupted envelope architecture, and increased periplasmic space under antibiotic treatment compared to untreated controls. In contrast, MRSA strains showed minimal structural changes at the same time point (Fig. 5a,b). TEM and AFM confirm early cell-wall remodeling in susceptible strains. (a) Representative TEM images of MSSA and MRSA cells at 20 min of growth with or without ampicillin; example wall-thickness measurements are marked. (b) Quantification of cell-wall thickness (mean ± SD) for untreated and treated MSSA and MRSA at 20 min; lighter bars show untreated controls, red bars show ampicillin-treated MSSA, and blue bars show ampicillin-treated MRSA. Statistical significance is annotated as ns (p>0.05) or *** (p < 0.001). (c) Representative AFM height images of cells on glass for each strain at 20 min, with untreated cells on the left and ampicillin-treated cells on the right. (d) Quantification of cell heights (mean ± SD) for untreated and treated MSSA and MRSA cells; lighter bars indicate controls, red bars indicate treated MSSA, and blue bars indicate treated MRSA. Statistical significance between treated and control is denoted as ns (p>0.05) or * (p < 0.05). AFM measurements corroborated these findings. Ampicillin-treated MSSA cells exhibited decreased apparent height and altered surface topography, consistent with cell-wall weakening or restructuring, whereas MRSA cells maintained relatively uniform height and smoother surfaces (Fig. 5c,d). Quantitative analysis of wall thickness and cell height showed significant differences between treated and untreated MSSA, but not between treated and untreated MRSA (Fig. 5c,d). These orthogonal measurements strengthen the interpretation that the early discriminatory FTIR signal reflects cell-wall remodeling in susceptible strains rather than nonspecific late-stage damage. Mechanism-focused features improve cross-strain discrimination The previous analyses suggested that early discrimination should depend primarily on a narrow set of bands associated with antibiotic-induced cell-wall stress, rather than on broad spectral differences between strains. To visualize this hypothesis directly, we compared PCA projections built from three feature sets: the full spectrum, the LASSO-selected features, and a restricted mechanistic feature set spanning the peptidoglycan precursor and cell-wall stress regions (1025–1015 cm and 1520–1505 cm) (Fig. 6). Using the full spectrum, samples clustered mainly by strain rather than by resistance class. LASSO-selected features showed some class-associated patterns at 20 and 30 min, but no clear separation. In contrast, the restricted mechanistic bands produced clearer class-level organization in the early time window, indicating that these regions capture a more generalizable antibiotic-response signal. Borderline behavior remained most evident for WKZ1 and WKZ2, consistent with their biological similarity and the partial resistance phenotype of WKZ1. Mechanism-focused features yield clearer early class structure than global spectral features. Principal component analysis at 20, 30 and 60 min using 3 feature sets: all preprocessed spectral features (left column), LASSO-selected features (middle column), and restricted mechanistic bands associated with peptidoglycan precursor accumulation and cell-wall stress (right column). MRSA strains are shown in cool colors, and MSSA strains are shown in warm colors. Full-spectrum projections are dominated by strain-level structure, whereas the restricted mechanistic bands show clearer MRSA/MSSA organization at early time points. We then used these mechanistic bands for direct classification, combining band-restricted raw spectra with difference spectra designed to isolate antibiotic-induced change relative to untreated and time-zero controls (Supplemental Fig. 1). Under LOSO-CV, this mechanism-focused representation achieved balanced accuracies of 0.91 at 20 min and 0.90 at 30 min (Fig. 7a), substantially improving generalization relative to earlier full-spectrum analyses. Even the challenging WKZ1 and WKZ2 strains remained separable at 30 min, with our sub-MIC dosing of 0.5 μg/mL, with balanced accuracies of 0.82 and 0.90, respectively. Mechanism-focused LOSO-CV improves early discrimination across unseen strains. Heatmaps show balanced accuracy for each held-out strain using classifiers trained on original and difference spectra restricted to the cell-wall associated mechanistic bands. Values above each column indicate mean balanced accuracy across all held-out strains. (a) All strains were dosed equally with 0.5 μg/mL ampicillin. (b) WKZ1 and WKZ2 were given a dose of ampicillin equal to their MIC (highlighted in red) at 40 μg/mL. Increasing the dose improves discrimination for the borderline WKZ1/WKZ2 pair, while the overall highest across-strain performance remains within the 20–30 min window. When WKZ1 and WKZ2 were tested at higher doses matched to their MIC of 40 μg/mL, LOSO-CV performance improved further, reaching 0.95 at 20 min and 0.98 at 30 min (Fig. 7b). This result suggests that optimal antibiotic challenge conditions may improve discrimination for borderline or partially resistant strains and should be explored in future work. Performance declined modestly at 60 min in both dosing schemes. This result is consistent with the reduced prominence of the early polysaccharide precursor signal at later times and suggests that the most informative biochemical window for rapid prediction lies within the first 30 min of β-lactam exposure. Discussion Our results identify an early biochemical window in which the response of S. aureus to β-lactam exposure is both mechanistically specific and spectrally detectable. Within 20–30 min of ampicillin challenge, ATR-FTIR spectra contained sufficient information to distinguish MRSA from MSSA, and the most informative early features mapped primarily to peptidoglycan- and carbohydrate-associated bands. TEM and AFM measurements at 20 min independently confirmed structural remodeling in MSSA cells but not in MRSA cells, supporting the conclusion that early discrimination is driven predominantly by cell-wall stress and remodeling in susceptible strains, rather than by nonspecific late-stage damage. Capitalising on this window with interpretable machine learning models yielded MRSA/MSSA discrimination that is faster and more generalisable than previous approaches. Methodologically, we deliberately emphasize low-complexity, transparent models for a small, high-dimensional dataset. Moderate preprocessing and mechanistic feature selection by LASSO, followed by linear classifiers, deliver both performance and interpretability. Each retained band has a clear biochemical assignment, and the decision boundary is nearly linear in this reduced space. Not all classification performance estimates are equally informative. Under sample-level cross-validation, LASSO-based models achieved high apparent accuracy at 20–30 min. However, leave-one-strain-out validation showed that part of this performance was attributable to stable strain-specific spectral structure, which does not necessarily generalize to previously unseen isolates. This distinction is important for small spectral datasets, where sample-level resampling can give an overly optimistic picture of real-world performance. By explicitly separating within-strain reproducibility from across-strain generalization, our analysis shows that stringent validation is not just a technical detail but a necessary part of interpreting discrimination accuracy in biologically heterogeneous datasets. An important methodological development is that restricting the analysis to mechanism-linked spectral regions improved across-strain discrimination. Full-spectrum representations and even LASSO-filtered features still retained substantial strain-level structure, as illustrated by the PCA projections. In contrast, focusing on narrowly defined peptidoglycan precursor and cell-wall stress bands, together with difference spectra that isolate antibiotic-induced change relative to untreated baselines, substantially improved LOSO-CV performance. This result indicates that the most transferable discriminatory signal lies not in the global spectral fingerprint of a strain, but in the shared biochemical trajectory induced by antibiotic exposure. More broadly, this suggests a useful strategy for small-N phenotypic spectroscopy studies: mechanism-informed feature restriction may yield more robust generalization than purely data-driven dimensionality reduction alone. From a translational perspective, the ability to phenotypically distinguish MRSA from MSSA within 20–30 min of antibiotic exposure is promising. Unlike genotypic assays, which detect predefined resistance genes, ATR-FTIR probes functional biochemical responses of cells to antibiotic challenge and may therefore capture resistance phenotypes that are not easily inferred from genotype alone. In principle, such an approach could substantially shorten the time required for susceptibility assessment. At the same time, the present study should be viewed as a proof of concept rather than a clinically deployable workflow. Horizontal gene transfer is a significant contributor to antimicrobial resistance in S. aureus14,15. The mobile genetic element SCCmec (staphylococcal chromosome cassette mec) is a well-characterized determinant of methicillin resistance16. The only difference between the genome structures of MSSA strain WKZ-1 and MRSA strain WKZ-2 is the presence of SCCmec type IV in the latter, most likely resulting from a gene transfer event originating with Staphylococcus epidermidis17. The ability of the mechanism-focused ATR-FTIR analysis to distinguish this pair within 20–30 min suggests that the method can detect functional biochemical consequences of resistance-associated genetic acquisition, rather than relying only on broad species- or strain-level spectral differences. With further validation and optimisation, the present workflow: a 20-min antibiotic exposure, preparation and drying, and scanning timeline, suggests a potential route toward same-visit susceptibility information for S. aureus isolates. This method could shorten the decision gap currently bridged by empiric or broad-spectrum therapy. Because our method uses interpretable bands rather than opaque high-capacity models, it is more amenable to quality control, troubleshooting, and regulatory validation. Several limitations define the scope of our conclusions. The study uses seven strains, a single β-lactam antibiotic, one primary dosing condition, and only four time points up to 60 min; performance and mechanistic patterns may differ across clinical isolates, drug classes, growth conditions, and sampling schedules. The small number of strains also means that, even under conservative validation, uncertainty in accuracy estimates remains substantial. In addition, the stronger performance obtained for WKZ1 and WKZ2 under higher-dose challenge suggests that discrimination may depend not only on resistance class but also on the relationship between antibiotic dose and strain-specific susceptibility. The prevalence of blaZ-mediated resistance in clinical S. aureus isolates18,19 means that ampicillin challenge conditions will need to be optimized by dose adjustment. Looking forward, we see several promising directions. Extending this approach to larger, genetically diverse collections of clinical isolates as well as strain mixtures, and to other antibiotic classes (e.g., glycopeptides, lipopeptides) will test whether early cell-wall stress signatures generalize across drugs and resistance mechanisms. Systematic evaluation of instrument-to-instrument and site-to-site variation will be essential to determine how much calibration or domain adaptation is needed for deployment. Finally, combining ATR-FTIR with automated sample handling and standardized, open-source analysis pipelines could pave the way for routine, high-throughput phenotypic susceptibility testing. In summary, we show that early cell-wall stress signature captured by ATR-FTIR can support rapid discrimination between MRSA and MSSA within 20–30 min of β-lactam exposure. The main contribution is not only the early time scale, but also the demonstration that mechanistically targeted spectral features provide more generalizable discrimination than broader sample-level classifiers. By combining sparse, interpretable modeling with stringent validation and orthogonal structural validation, this work establishes a mechanistically grounded proof of concept for rapid FTIR-based susceptibility assessment and a framework for future testing in clinically diverse isolate sets. Methods Workflow and pipeline Seven Staphylococcus aureus strains were used in this study: three MSSA strains (ATCC 6538, RN4220, WKZ1) and four MRSA strains (ATCC 43300, USA300-JE2, WKZ2, CA629). These included standard laboratory reference strains as well as strains with broader genetic backgrounds. RN4220 is a restriction-deficient derivative of NCTC8325-4 commonly used in genetic studies, and USA300-JE2 represents a major community-associated MRSA lineage. Strains were challenged with ampicillin and analyzed by ATR-FTIR spectroscopy at defined early time points. The resulting spectra were preprocessed and used in two complementary analysis pipelines: a data-driven pipeline comparing full-spectrum, PCA-reduced, and LASSO-selected features, and a mechanism-focused pipeline restricted to bands associated with early cell-wall stress (Fig. 8). Experimental workflow and analytical pipeline for ATR-FTIR-based discrimination of MSSA and MRSA. MRSA (blue) and MSSA (red) strains were grown and challenged with ampicillin, dried on an ATR crystal, and analyzed by FTIR spectroscopy over the 1800–800 cm−1 range. Second-derivative spectra were used in three analysis pipelines based on all spectral features, PCA-reduced features, and LASSO-selected features, followed by classification using linear discriminant analysis (LDA), partial least squares-discriminant analysis (PLS-DA), and support vector machine (SVM). The workflow illustrates how raw spectral measurements were converted into predictive models for rapid phenotypic discrimination between MRSA and MSSA. Growth conditions and sample collection For each strain, a single overnight culture was grown in Luria-Bertani (LB) broth at 37 °C under aerobic conditions. Mature cells were diluted in sterile LB broth to an initial OD of 0.20 and used to create four independent replicates per strain. Ampicillin was added (0.5 μg/mL), and samples were harvested at 0, 20, 30, and 60 min post-exposure. Cells were pelleted by centrifugation at 13,000 rpm, washed with sterile distilled deionized water (ddHO), and resuspended in 10 μL of ddHO for FTIR analysis. This design resulted in 4 biological replicates per strain per time point, for a total of 112 samples. ATR-FTIR spectroscopy FTIR measurements were acquired using a Bruker ALPHA II Compact FT-IR spectrometer. Before each measurement, the diamond ATR crystal was cleaned with methanol and lens paper, and a new background spectrum at 4 cm−1 resolution was collected. A 2 μL aliquot of pelleted cells was deposited onto the ATR crystal. This volume of 2 μL was chosen to minimize drying time while maintaining adequate signal-to-noise ratio. Care was taken to avoid bubble formation and to ensure uniform drying. The scanning resolution was set to match the background spectrum. Because water absorption bands overlap with spectral regions of interest in this study, hydration was monitored spectroscopically during drying. Samples were allowed to dry on the ATR crystal until the absorbance spectrum stabilized. Stability was assessed from the O–H stretching region between 2700 and 3600 cm−1. When the standard deviation between measurements fell below 5% of the absorbance magnitude in this region, the sample was considered sufficiently dry for comparison. Additional details and examples of this method can be found in the Supplementary Materials of Hastings et al.8. For early time-point samples, this preparation and drying process typically takes approximately 15 min. Each measurement used 10 scans with four interferograms per scan. The absorbance curves were processed in Bruker OPUS software and subsequent analyses were performed using in-house Python scripts. Spectra were min-max normalized over the 1800–1000 cm−1 range to reduce variation caused by sampling differences. The normalized FTIR spectra were smoothed using a 13-point Savitzky-Golay smoothing algorithm with a second-order polynomial, because it offers sufficient noise reduction while retaining the peaks in the Amide I and Amide II regions. The smoothed spectra were converted to second derivatives, which were used for all downstream analyses. Transmission electron microscopy (TEM) For orthogonal structural validation, TEM was performed on representative MSSA (ATCC 6538) and MRSA (ATCC 43300) strains harvested after 20 min of ampicillin exposure along with their untreated controls. Samples were fixed in 2.5% glutaraldehyde in 0.1M cacodylate buffer (pH 7.4) for 24–48 h at 4°C. Between each subsequent step, samples were pelleted by centrifugation at 2000–5000 rpm for 2 min. After primary fixation, samples were washed twice with 0.1M cacodylate buffer and then post-fixed with 1% osmium tetroxide in 0.1M cacodylate buffer for 1 h at room temperature. Following two washes with distilled water, samples were dehydrated through a graded ethanol series (25%, 50%, 75%, 95%, and two changes of 100%) for 15 min per step, followed by two changes of propylene oxide (PPO). Infiltration was performed with gradually increasing concentrations of Eponate 12™ epoxy resin in PPO (2:1, 1:1, and 1:2 PPO:resin ratios), culminating in two changes of 100% resin. Samples were then embedded in fresh resin in labeled Beem™ capsules and polymerized at 60 °C for 72 h. Ultrathin sections (70–80 nm) were cut using a Leica Enuity™ Ultramicrotome with a Diatome™ diamond knife, collected on 200-mesh copper grids, and post-stained with 5% uranyl acetate and Reynolds’ lead citrate. Grids were imaged using a JEOL JEM 1400 transmission electron microscope operating at 80 kV, and digital images were acquired with a Gatan Ultrascan 2K × 2K camera. Cell wall thickness was measured manually from digital images using ImageJ (National Institutes of Health); for each group, 8 cells were selected, and each cell wall was measured 8 times with values averaged to yield a single thickness per cell. Treatment effects were assessed using two-tailed Student’s t-tests. Only cells not in the process of binary fission and with sharp cell wall boundaries were selected for measurements. Atomic force microscopy (AFM) Atomic force microscopy (AFM) imaging was performed on representative MSSA and MRSA strains (ATCC 6538 and ATCC 43300) harvested after 20 min of ampicillin exposure together with untreated controls. Images were recorded on a TT-2 AFM Workshop operating in tapping mode using silicon probes (Tap190Al-G, BudgetSensors) at a scan size of 4 μm. Data were acquired at room temperature and were processed in Gwyddion, an open-source image processing software. Four images were acquired per group. Images in which the substrate surface was visible were selected to allow accurate baseline (z = 0) calibration. Height maps were smoothed and peak heights were determined from local maxima, with detected peaks verified against visible cell counts in each image. Spectral standardization and PCA All machine learning was performed in Python. Before classification, all 487 spectral features were standardized using StandardScaler from scikit-learn, which transforms each feature to have zero mean and unit variance (). PCA was used both for dimensionality reduction and for visualization. For predictive modeling, PCA was applied to samples within each time point separately, and the first 10 principal components (explaining >95% cumulative variance) were retained (Supplemental Fig. 3) . For visualization, PCA was performed on each of the three feature sets, and the first three principal components were used. LASSO-feature selection LASSO (least absolute shrinkage and selection operator) with bootstrap stability analysis was applied to the full standardized spectrum. Across 100 bootstrap iterations, features were selected repeatedly using LASSO, and selection frequency was used as a measure of feature stability. Stable features were then ranked by selection frequency, and the retained set was chosen to capture 80% of the cumulative bootstrap-weighted feature importance, providing a robust and interpretable feature subset for downstream classification. Unlike PCA, which forms linear combinations of all features, LASSO directly identifies individual wavenumbers associated with class discrimination and therefore supports biochemical interpretation. LASSO performs variable selection and regularization by imposing an L1-norm penalty on the linear model coefficients, promoting sparsity. The LASSO estimator is defined as: where denotes the L1 norm and is the regularization parameter. LassoCV automatically selects λ via 10-fold cross-validation, yielding λ = 0.05. Unlike PCA, LASSO directly selects features relevant for class discrimination. Mechanism-focused feature selection A mechanism-focused feature set was defined from spectral regions associated with early cell-wall stress and peptidoglycan precursor accumulation: 1015–1025 cm−1 and 1505–1520 cm−1. These regions were chosen based on prior literature assignments10, and their relevance to MRSA/MSSA discrimination was independently supported by the LASSO stability analysis performed on the full-spectrum features. Two complementary feature representations were used within these restricted bands. First, raw second-derivative spectra preserved the band position and intensity. Second, subtracted spectra were computed to isolate the antibiotic-induced change. Each sample spectrum was referenced to the mean untreated control spectrum of the same strain at the same time point, removing growth-related drift unrelated to antibiotic exposure. The resulting spectrum was then referenced to the mean time-zero control spectrum of the same strain, removing stable baseline inter-strain differences. The result was a spectrum reflecting the antibiotic-induced biochemical trajectory of each sample relative to its own untreated baseline. Raw and double-subtracted features within the mechanistic bands were then concatenated to form the final mechanism-focused feature set (Supplemental Figs. 1 & 2). Classification methods and validation Classification was performed using LDA, PLS-DA, and SVM-based models across four feature sets: full spectrum features, PCA-reduced features, LASSO-selected features, and the mechanism-focused band-restricted feature set. For each classifier, hyperparameters were tuned within the training data only. For LDA, the shrinkage parameter was selected by grid search over 11 evenly spaced values between 0 and 1 using the least-squares solver, which is appropriate for small sample sizes where full singular value decomposition may be numerically unstable. For linear SVM, the regularization parameter C was selected by grid search over 0.01, 0.1, 1, 10, 100. For PLS-DA, the number of latent components was selected by cross-validation within the training dataset. Balanced accuracy was used as the tuning metric to account for class imbalance. Tuning was performed independently within each LOSO fold, ensuring that no information from the held-out strain influenced hyperparameter selection. Two validation schemes were used. For the initial comparison of feature representations and classifiers, we used a strain-aware, strain-balanced cross-validation strategy designed to provide robust within-strain performance estimates while limiting domination by any single strain. Performance was summarized as mean accuracy with 95% confidence intervals across repeated random splits. This method provides a more conservative and realistic estimate of generalizable performance compared to conventional validation on limited strain sets. To assess generalization to previously unseen strains, we used leave-one-strain-out cross-validation (LOSO-CV). In each fold, all samples from one strain were held out as the test set while all remaining strains formed the training set. Feature selection, scaling, and hyperparameter tuning were performed using training data only, with no information from the held-out strain entering model construction. Balanced accuracy was recorded for each held-out strain at each time point. To account for variability introduced by the bootstrap feature selection, the full LOSO-CV pipeline was repeated across 10 independent random seeds and results were averaged. Software Data processing and analysis were performed using OPUS, LibreOffice Calc, Python, and standard scientific Python libraries including NumPy, pandas, SciPy, scikit-learn, and matplotlib. The Savitzky-Golay smoothing and second-derivative calculation were implemented using scipy.signal, and classification and preprocessing were implemented in scikit-learn. Supplementary Information Supplementary Information. Acknowledgements We acknowledge the Frady Whipple Endowed Professorship (to YJ). D.S. Recognizes the Molecular Basis of Disease (MBD) for graduate fellowship. Research Innovation and Scholarly Excellence (RISE) challenge grant from Georgia State University (to G.H. and Y.J.). TEM experiment was supported by the Emory University Robert P. Apkarian Integrated Electron Microscopy Core Facility (RRID: SCR_023537), which is subsidized by the School of Medicine and Emory College of Arts and Sciences. Additional support was provided by the Georgia Clinical & Translational Science Alliance of the National Institutes of Health. Author contributions G.H., E.G., Y.J. conceived the experiment(s); W.H., M.N., A.M., D.S., and C.C.T. conducted the experiment(s); W.H. performed the analysis; W.H. and Y.J. wrote the manuscript. All authors reviewed and edited the manuscript. Data availability All analysis codes and sample data are available at https://github.com/Jiang-Lab/FTIR-MRSA.git. Declarations Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Michael Nelson, Caroline C. Taylor, and Alexander Marchesani contributed equally to this work. Supplementary Information The online version contains supplementary material available at 10.1038/s41598-026-57254-2. References Associated Data Supplementary Materials Supplementary Information. Data Availability Statement All analysis codes and sample data are available at https://github.com/Jiang-Lab/FTIR-MRSA.git.