Introduction

Enterococcus faecalis are Gram-positive paired or chained cocci. They are part of the normal human gastrointestinal tract. However, in some cases, they can cause urinary tract infections, endocarditis and seldom septicaemia. Because enterococci have several natural resistances against a wide range of antibiotics, vancomycin is often used in enterococcal infection as a so-called “antibiotic of last resort”. Vancomycin is a glycopeptide antibiotic and active against aerobe and anaerobe Gram-positive bacteria. It entered clinical practice in the 1950s, but was soon substituted with alternatives that were considered to be less toxic or more efficacious [1]. Increasing resistances of pathogens against those alternative drugs resulted in a return of vancomycin [1–3].

Vancomycin inhibits the last stage, stage III, of the bacterial cell wall synthesis by binding to the d-alanyl-d-alanine terminus of the peptidoglycan precursor UDP-MurNAc-pentapeptide or nascent glycan chains. During the interaction, the peptide moiety of vancomycin forms five hydrogen bonds with the d-Ala-d-Ala dipeptide from the cell wall precursor. This bulky complex prevents glycan polymerization (transglycosylation) as well as cross-linking of the individual peptidoglycan chains (transpeptidation). This leads to an instable bacterial cell wall resulting in lysis of the cell [4].

In the 1980s, the first vancomycin-resistant enterococci (VRE) emerged which can be associated with the continuous use of vancomycin and broad spectrum antibiotics [5]. Nowadays, VRE have spread extensively worldwide and cause increasing problems in clinical practice [6–10]. However, not only the VRE themselves, but also the risk of transferring vancomycin-resistance genes to other pathogens, for example methicillin-resistant Staphylococcus aureus (MRSA), marks a major reason of concern [2, 11–13]. In order to quickly identify, isolate and efficiently treat patients suffering from VRE, fast diagnostic methods are needed. Current standard microbiological methods are very time-consuming as they require bacterial cultivation for detection of bacterial growth or growth behaviour [14, 15]. Therefore, research efforts are undertaken to speed up analysis time. Such alternative approaches comprise polymerase chain reaction (PCR)-based methods that identify the resistance genes [16–18], mass spectrometry [19–21], flow cytometry [22, 23] as well as combinations of phenotypic and genotypic approaches [19, 24, 25]. Among the spectroscopic methods, vibrational spectroscopy, such as IR absorption and Raman scattering, gained a lot of importance as they require only minimal sample amounts down to a single bacterial cell. This enables insights into cellular heterogeneity and the possibility to study difficult-to-culture pathogens with easy sample preparation [14, 26–28]. Moreover, vibrational spectroscopic techniques are label-free and non-destructive, allowing further analysis after spectroscopic characterization [29, 30]. The “vibrational spectroscopic fingerprint” is characteristic for each bacterial species due to the unique cellular composition made up of various nucleic acids, lipids, proteins, polysaccharides and carbohydrates. During recent years, Raman spectroscopy was already successfully applied to characterize bacteria, discriminate Gram-positive and Gram-negative bacteria as well as identify and classify various pathogens [26, 28, 31–37] and also pathogens under antibiotic influence [27]. An advance offers the combination of Raman spectroscopy with microfluidic devices with further functionalities, such as dielectrophoresis for the enrichment of bacteria directly in suspension, making sample preparation faster and very easy [30, 38, 39]. Furthermore, Raman spectroscopy can be successfully applied to investigate bacteria–drug interactions with antibiotics as was shown for the interaction of fluoroquinolones with Gram-positive bacteria [40, 41], the antibiotic effects on Pseudomonas aeruginosa biofilms [42] as well as the interaction of clinical relevant antibiotics with Escherichia coli [43–45]. Vancomycin–bacteria interactions have been employed to capture bacteria on silver nanoparticles for further spectroscopic characterization [46] as well as to differentiate resistant and sensitive enterococci [38].

In this contribution, we present a systematic characterization of the interaction of vancomycin with enterococci by means of Raman spectroscopy. The spectral changes upon efficient action of the drug is analyzed under different experimental conditions. The aim is to achieve a robust and fast differentiation of vancomycin treated and untreated E. faecalis. Inter-personal operator variation was taken into account by training a classification model with samples prepared and measured by one operator and testing its prediction accuracy with biological replicates prepared and measured later on by another operator.

Such drug–bacteria interaction studies provide the knowledge base for the interpretation of experiments with resistant bacteria, such as various VRE strains. Furthermore, they are also valuable for the assignment of the mode of action when studying the effect of different antibiotics on bacteria.

Materials and methods

Cultivation and treatment of microorganisms

E. faecalis ATCC® 29212™ was cultivated in CASO broth (ROTH GmbH) overnight at 37 °C while shaking at 90 rpm. The overnight culture was used to inoculate a starter culture with an optical density (OD) of 0.1 (d = 1 cm) at λ = 600 nm. To monitor bacterial growth, the OD was measured over time at 600 nm (with pathlength 1 cm) using a Cary 60 UV-Vis spectrophotometer (Agilent Technologies Cary 60). At an OD of 0.6–0.8, vancomycin hydrochloride hydrate (SIGMA-ALDRICH, Steinheim, Germany) was added to the bacteria culture giving a final vancomycin concentration of either 100 μg/ml (batches 1–4) or 10 μg/ml (batches 5–11). In the following, this time point is called time point 0 min. Each batch is an independent biological replicate (new overnight culture) that comprises a vancomycin-treated sample and a control where no vancomycin was added, but which was otherwise treated identically.

Raman sample preparation

Bacteria sampling for Raman spectroscopic investigations was carried out according to previously published protocols [40, 47]: Samples of the bacterial culture were taken directly (0 min) and 30, 60 and 90 min after addition of vancomycin. At each time point, 1 ml bacterial culture was harvested. The medium was removed by centrifugation (Eppendorf centrifuge 5415, 13,000 rpm, 3 min, room temperature) and the cells were washed twice: first, with 0.9 % sodium chloride solution and, secondly, with deionized water. The pellet was resuspended in 50 μl deionized water and dried at 55 °C for 20 min on a calcium fluoride (CaF2, Crystal GmbH, Berlin, Germany) slide. Raman spectra were taken immediately after sample preparation.

Vancomycin hydrochloride hydrate and peptidoglycan (from S. aureus cell wall component, SIGMA-ALDRICH, Steinheim, Germany) were measured as solids under the same experimental conditions as the bacteria samples.

Micro-Raman spectroscopy

Raman spectra of the bacteria on CaF2 slides were recorded with an upright micro-Raman setup (CRM 300, WITec GmbH, Germany) equipped with a 600 grooves/mm grating and a back-illuminated CCD camera (DV401 BV, ANDOR, 1024 × 127 pixels) cooled down to −60 °C. A frequency doubled Nd:YAG solid-state laser (532 nm, Excelsior 532-60) was used to excite the Raman signal. The laser light was focused onto the sample using a 50x objective (Zeiss EC EpiPlan, NA 0.7) for the batches 1–3, and a 100x objective (Carl Zeiss LD EC EpiPlan-Neofluar, NA 0.75) for batches 4–11. For the 50x objective, chromatic aberrations were encountered when studying thin bacterial films which were not observed with the 100x objective. Laser power was 15 mW at the sample. The 180° backscattered light was collected and delivered to the spectrograph by a multimode optical fibre with 50 μm core diameter. Performance and alignment of the system were checked each day by recording Raman spectra of silicon as well as of 4-acetamidophenol (ACP, paracetamol, SIGMA-ALDRICH, Steinheim, Germany). ACP spectra were used to correct possible shifts in the laser wavelength. Reference band positions were taken from the literature [48].

Two different sampling modes were applied. For batches 1–7, an automated scanning mode was used to collect 49 spectra evenly distributed over an area of 20 μm × 20 μm integrating 20 s per spectrum. However, sample roughness led to varying signal-to-noise ratio within each area. Furthermore, spectra with cosmic spikes in the fingerprint region were discarded prior to analysis. For batches 8–11, 20 single spectra (without cosmic spikes) were recorded at different locations on the sample with an integration time of 30 s. The total numbers of spectra per time point and treatment are summarized in Table 1.

Table 1 Overview of the number of spectra from 11 independent measurements. Data are grouped according to the time point after vancomycin addition and treatment (with or without vancomycin)
Full size table

Data analysis

Statistical analysis was carried out in the statistical language GNU R, version 3.0.3 [49] using the “hyperSpec” [50] environment for handling the spectra. The package “ggplot2” [51] was employed for graphical display (Fig. 4).

Spectra pre-processing

Spectra were initially truncated to the spectral ranges 350–2000 and 2500–3300 cm−1. Cosmic rays in the Raman maps were removed with an interactive, in-house written spike filter. Baseline correction in the fingerprint region (350–2000 cm−1) was performed with an automated linear baseline fit with 50 support points. For the C–H stretching region, a quadratic baseline was used. The fitting ranges were from 2500 to 2775 cm−1 and from 3150 to 3300 cm−1 where no Raman signal occurs, and the CH region between 2825 and 3050 cm−1 was extrapolated with the fit. Next, “loess” smoothing interpolation [52] was used to downsample the spectra onto an evenly spaced wavenumber axis from 600 to 1750 and 2825 to 3050 cm−1 with data point spacing of 5 cm−1. This interpolation took the laser wavelength correction calculated from the ACP spectra into account and, at the same time, provided an evenly spaced wavenumber axis common to all measurements. The smoothing parameter was set accordingly corresponding to 10 cm−1 spectral resolution. Interpolation was necessary due to slightly varying excitation wavelengths over several months (as measured by ACP spectra). This operation also trades some spectral resolution for a better signal-to-noise ratio. Fig. S1 in the Electronic Supplementary Material (ESM) visualizes the pre-treatment procedure for an exemplarily Raman spectrum of E. faecalis.

For the 50x objective (NA 0.7) employed initially to cover a large area of the bacteria film, significant chromatic aberrations were observed in the spectral region between 600 and 3000 cm−1 (corresponding to a wavelength range of 549 to 633 nm). Those chromatic aberrations were observed in the experimental data as changes in the relative intensities of corresponding Raman bands, such as, e.g. C–H deformation and C–H stretching, with different foci (z-positions) in the thin bacterial film. Similar findings have been described measuring microbial spores [53]. In a principal component analysis (PCA), this aberration initially yielded a component which was essentially the average spectrum with a change in sign between the fingerprint and the C–H stretching regions. This component carried ca. 90 % of the total variance and was uncorrelated with batch, time and treatment. The chromatic aberrations were approximately corrected by normalizing the fingerprint and C–H stretching regions separately to their mean intensity. In addition, from batch 4, another objective with improved colour correction (100x EC Epiplan Neofluar) was used.

Descriptive LDA

Descriptive linear discriminant analysis (LDA) was carried out to reveal spectral differences due to vancomycin treatment with 100 μg/ml vancomycin (four batches). These models were regularized by partial least squares regression (PLS). PLS is a latent variable approach which models the covariance structures of a matrix of predictors X (in our case the Raman spectra) and a matrix of response variables Y (which are the treatment groups “treated” and “untreated” in our example). LDA maximizes the variance between given groups (the treatment) with respect to the variance within these groups. However, in contrast to PLS, LDA needs large sample sizes relative to the number of variates. Therefore, appropriate regularization to few input variables, such as the first (few) PLS scores, is important [54, 55]. These combined PLS-LDA models are provided by the R package “cbmodels” [56] relying on packages “pls” [57] and “MASS” [52]. Separate descriptive PLS-LDA models were calculated for each of the four treatment time points (0, 30, 60 and 90 min) by taking just one PLS latent variable. The coefficients extracted from these models were used to analyse relevant changes in spectra over time.

Setup and validation of a predictive model

In order to train a model for the prediction of efficient bacteria–drug interaction, only those time points for which the PLS-LDA coefficients were steady (30, 60 and 90 min) were used. For this predictive PLS-LDA model, only two batches (batches 1 and 2) prepared and measured by operator 1 (SK) were used for training. Here, a high vancomycin concentration (100 μg/ml) was applied to observe strong drug-induced spectral variations. The first latent variable was sufficient to discriminate treated and untreated spectra. The second latent variable already carried much noise (even though some spectral features are still recognizable). While the first latent variable explains 89 % of the Y-variance in the training data using 37 % of the variance in the spectra, the second latent variable almost doubles the explained variance in the spectra, but adds only 3 % to the explained Y-variance. A closer inspection of the scores reveals that the second latent variable captures mostly variation between the biological replicates. Therefore, only one latent variable was considered for modelling. This also keeps the number of variates within the model as low as possible which is advisable for a dataset with, in total, 11 biological independent replicates [58]. The model performance was evaluated with nine independent biological replicates prepared and measured by operator 2 (CA), two replicates with 100 μg/ml and seven with only 10 μg/ml vancomycin.

Calculation of changes in the relative nucleic acid/protein content

To follow changes in the relative nucleic acid/protein ratio with time and vancomycin treatment, ratios of mean Raman intensities in the spectral range 775–800 cm−1, representing nucleic acid (785 cm−1), to the 995–1020 cm−1 region, representing protein (1005 cm−1), were calculated (Fig. 1b) by numerically integrating the spectrum in the respective spectral range.

Fig. 1
figure 1

a Growth curve of Enterococcus faecalis measured as optical density (OD) at 600 nm over time in the absence and presence of vancomycin. The arrow marks the time point of antibiotic application. b Nucleic acid/protein ratio during cultivation without (0 μg/ml) and with 10 or 100 μg/ml of vancomycin calculated using the integrated intensity of the Raman bands at 785 cm−1 (uracil, cytosine, thymine and PO2 −) and at 1005 cm−1 (phenylalanine). The mean and standard deviation were calculated from the 11 biological replicates

Full size image

Results and discussion

Characterization of the growth of E. faecalis using optical density

The growth of E. faecalis ATCC® 29212™ was measured as optical density of the bacteria culture over a time of 3 h (1.5 h before to 1.5 h after vancomycin addition). Each biological replicate (batch) comprises an E. faecalis control culture without vancomycin and an E. faecalis culture incubated with vancomycin. Figure 1a shows a typical E. faecalis growth curve which can be divided in lag phase (first 30 min of growth), followed by the exponential growth phase and stationary phase (>90 min, not shown). From the slope of the growth curve in the exponential phase the specific growth rate μ of E. faecalis ATCC® 29212™ in CASO at 37 °C was determined to be μ = 1.315 ± 0.07 h−1, the generation time g = ln2/μ = 32 ± 2 min and the dividing rate n = 1/g = 1.9 ± 0.1 h−1. Mean and standard deviation were calculated from six independent biological replicates. These values are in good agreement with other data published in the literature [59–61].

Characterization of vancomycin–E. faecalis interaction using established microbiological methods

The minimum inhibitory concentration (MIC) of E. faecalis ATCC® 29212™ was determined by automated analysis using a VITEK® 2 system (bioMérieux, Nürtingen, Germany) to be 2 μg/ml. This is in good agreement with literature values reporting MICs for vancomycin between 1 and 4 μg/ml [62]. With a MIC of vancomycin ≤4 mg/l, E. faecalis ATCC® 29212™ is a vancomycin susceptible strain according to the EUCAST-approved susceptibility breakpoint [63].

Vancomycin inhibits cell wall synthesis during bacterial growth and division [4]; thus, it is most efficient on metabolically active bacteria. Therefore, in our experiments, vancomycin was always added during the exponential growth phase (90 min after inoculation, time point 0 min). To ensure a strong and fast effect of the antibiotic, a vancomycin concentration of 50 times the determined MIC, i.e. 100 μg/ml, was used to analyse typical spectral changes induced by the action of the drug and also to train the classification model to identify treated bacteria. The model was tested using vancomycin concentrations of 5 times the MIC (10 μg/ml vancomycin) which is the lower range of a typical trough serum concentration recommended for most patients during appropriate vancomycin therapy [1, 64].

The growth curve depicted in Fig. 1a demonstrates that both applied vancomycin concentrations (100 and 10 μg/ml) inhibit further growth of E. faecalis. For the high vancomycin concentration (100 μg/ml), this effect is visible already after 30 min for the lower concentration at 60 min after vancomycin addition.

Raman spectroscopic characterization of vancomycin–E. faecalis interaction

The antibiotic influence of vancomycin on E. faecalis was monitored by recording Raman spectra of the bacteria immediately after adding the drug (0 min) and every 30 min after vancomycin addition over a time span of 1.5 h, i.e. 30, 60 and 90 min of vancomycin–E. faecalis interaction time. The experiment was carried out using two different drug concentrations: 100 and 10 μg/ml vancomycin respectively. E. faecalis cultures without vancomycin treatment served as controls.

Figure 2 summarizes the pre-processed average Raman spectra of E. faecalis at the four different time points after addition of vancomycin, as well as Raman spectra of the untreated controls. Mean spectra including the 5th and 95th percentile are shown in Fig. S2 in the ESM. The Raman spectra show all characteristic features of typical bacteria spectra and are in good agreement with published data [26, 27, 30, 31]. The vibrational bands can be assigned according to contributions from the overall biochemistry content of the bacteria cells. Nucleic acid contributions show up at 785, 1335 and 1575 cm−1. Proteins are responsible for the amide I around 1660 cm−1 and the amide III band around 1250 cm−1 as well as the sharp ring breathing mode of phenylalanine at 1005 cm−1. The band around 1450 cm−1 can be assigned to C–H deformation vibrations and the prominent bands around 2935 cm−1 to C–H stretching vibrations. The Raman spectra of the different treatment methods and the different time points appear very similar at first glance. However, they contain significant differences that reflect metabolic changes. These differences can be extracted and visualized by multivariate statistical analysis methods.

Fig. 2
figure 2

Mean Raman spectra of E. faecalis without (black) and with 10 μg/ml (grey) and 100 μg/ml (light grey) of vancomycin for four different time points of vancomycin–E. faecalis interaction time: immediately after addition of the drug (0 min) and after 30, 60 and 90 min

Full size image

Calculation of the nucleic acid/protein ratio to follow bacterial growth

The ratio of nucleic acid and protein content is known to be characteristic for the different metabolic stages during bacterial growth [41, 43, 65]. This ratio is highest in the exponential growth phase when bacteria have high transcription of genes and decreases in the stationary phase when the bacteria accumulate storing material and show reduced replication activities. Figure 1b shows the calculated nucleic acid/protein band ratios for the three different vancomycin concentrations. As expected, a slight decrease of the ratio is observed for the controls (0 μg/ml vancomycin). During the observation period (0–90 min), these cultures develop from the exponential growth phase towards the stationary phase. For bacteria treated with vancomycin, the nucleic acid/protein ratio stays constant within experimental error (Fig. 1b). This indicates that under the influence of the drug, bacteria cannot progress into another metabolic state and remain with a similar overall chemical composition as when hit by the insult. This observation is in good agreement with the action of the antibiotic: Vancomycin inhibits the cell wall biosynthesis, prevents bacteria to further divide and, ultimately, leads to cell death.

Descriptive LDA to analyze spectral changes due to vancomycin interaction

In order to visualize the spectral differences caused by the action of the drug, the Raman spectra of the untreated controls and the bacteria treated with the high vancomycin concentration (100 μg/ml) were compared for each time point by means of partial least square regression combined with linear discriminant analysis (PLS-LDA). The high vancomycin concentration was chosen for this analysis as the effect is expected to be strong. The analysis is carried out by individual comparison at each time point to unravel at which time point spectral changes happen. The coefficients of the PLS-LDA models are shown in Fig. 3a. Large coefficients (both, positive and negative) indicate wavenumbers that contribute to the differentiation of treated and untreated E. faecalis. Positive scores indicate vibrational bands that are more prominent in vancomycin-treated bacteria while negative scores originate from vibrational bands that are more intense in the untreated controls. The coefficient at time point 0 min shows no significant features, but rather spectral noise. This indicates that the pure presence of the antibiotic vancomycin in the bacterial culture does not have an effect on the Raman spectra of the bacteria. After 30 min of vancomycin interaction, time in the bacterial culture and characteristic spectral features emerge. These features stay the same or become even more prominent as the time elapses. All in all, the coefficients of the three interaction time points 30, 60 and 90 min look very similar. Most prominent positive signals are observed around 790, 1225 and 1490 cm−1. The signals around 1225 and 1490 cm−1 could be assigned protein contributions (amide III and amide II, respectively). Especially the band around 1225 cm−1 has been described to reflect intermolecular hydrogen bonds in bacterial cell wall peptides [66]. Such hydrogen bonds would be in good agreement with vancomycin binding to d-alanyl-d-alanine precursors of the cell wall. Strong negative signals are found around 1445 and 1655 cm−1. These signals can be assigned to CH2 deformation vibrations [31, 67] and to contributions from the amide I vibration from proteins [31] respectively. We attribute these spectral features to the increase in relative protein content with time, as discussed previously for the nucleic acid/protein ratio. However, the majority of the signals are a superposition of bands from several biochemical compounds of the complex bacterial cells allowing only wavenumber estimation instead of a well-defined band assignment. Raman spectra of pure substances are shown in Fig. 3b for comparison. The spectrum of solid vancomycin hydrochloride which is a glycopeptide consisting of amino acids and carbohydrates shows prominent Raman bands around 885, 993, 1232, 1320, 1610 and 2945 cm−1. These bands can be assigned to several vibrations corresponding to peptides/proteins [31, 67]. None of those bands contributes to the PLS-LDA coefficient separating treated and untreated bacteria. Thus again, not the pure presence of vancomycin but rather the action of the drug on the enterococci is detected. Vancomycin inhibits the bacterial cell wall biosynthesis by binding to cell wall precursors and hence, inhibiting peptidoglycan synthesis. The Raman spectrum of peptidoglycan, a polymer made of sugars (N-acetylglucosamine and N-acetylmuramic acid) and amino acids, shows prominent vibrational bands at 1670 and 1448 cm−1 originating from the peptide bonds and C–H deformation vibrations, respectively (Fig. 3b). Both of those vibrations appear as negative coefficients in Fig. 3a, a more pronounced contribution of peptidoglycans to the spectra of untreated bacteria. This is in agreement with an impaired peptidoglycan synthesis in the vancomycin-treated bacteria. However, it is impossible to assign the vibrational bands in the coefficient plot explicitly to changes in the bacterial cell wall. Similar bands could be observed due to changes in other proteins.

Fig. 3
figure 3

a Coefficients of the descriptive PLS-LDA for the different time points after vancomycin addition (100 μg/ml) classifying treated and untreated E. faecalis. b Raman spectra of vancomycin and cell wall component peptidoglycan

Full size image

Classification model for the fast identification of vancomycin interaction

In order to rapidly detect efficient vancomycin-enterococci interaction, a classification model to distinguish treated and untreated bacteria has been trained. Based on the observation from the descriptive LDA that all interaction times >30 min show similar coefficients, all time points after at least 30 min of vancomycin treatment were included into one classification model.

Two different operators carried out the experiments. To verify the model including possible inter-personal variation, data from operator 1 were used to set up the classification model and data recorded by operator 2 were used to evaluate the model's predictive abilities. The preferred validation scheme in literature for those numbers of replicates would be a resampling method such as repeated cross validation or .632+ bootstrap on the level of biological replicates [68, 69]. In the present study, however, we were able to make use of the rare fact that two independent persons were involved in the experiments. This allowed checking the generalization ability for measurements acquired up to 1 year after the acquisition of the training replicates was finished. The prediction of the data (biological replicates in batches 3–11) recorded by operator 2 with this treatment-based classification model is visualized in the score plot of the PLS-LDA model (Fig. 4). Spectra were classified as either control or vancomycin-treated samples. For easier assessment, data are grouped according to true treatment, i.e. control (0 μg/ml vancomycin) and treated with 10 and 100 μg/ml vancomycin respectively. Table 2 summarizes the correct classifications for each time point and treatment. Already after 30 min of vancomycin incubation, the prediction accuracy is above 90 %. Longer incubation times (60 min) lead to some further increase in correct classifications (≥95 %) and stay so also for 90 min incubation time.

Fig. 4
figure 4

Classification model based on antibiotic treatment. Test results for nine independent biological replicates prepared and measured by an independent operator. Single Raman spectra are represented by symbols. Horizontal lines represent 5, 10, 25, 50, 75, 90 and 95 % predicted posterior probability of the spectrum coming from a vancomycin-treated sample

Full size image
Table 2 Confusion matrix of the classification model predicting the treatment of the bacteria. Sensitivity and specificity were calculated for each time point for correct prediction of vancomycin treatment, irrespective of the applied vancomycin concentration
Full size table

For test samples treated with 100 μg/ml vancomycin, already at time point 0 min, many of the treated samples are recognized as such. We attribute this mainly to the fact that due to sample handling, a short but existing duration of vancomycin treatment occurs: While the sample is taken immediately after vancomycin is added, some time is needed for centrifugation and washing during which the drug can affect the bacteria. With the lower vancomycin concentration, this effect is not as pronounced. At time points ≥30 min, the scores for both concentrations are practically the same. As PLS-LDA scores would reflect concentration changes in a linear manner, we again arrive at the conclusion that biochemical changes in the cells due to the interaction of the drug with the cell are observed rather than the pure presence of the drug. Furthermore, the drug-induced biochemical changes reflected in the Raman spectra are the same for vancomycin concentrations 5× and 50× of the MIC. This is in agreement with biological knowledge as the MIC denotes the lowest drug concentration that can inhibit further growth of the bacteria.

The slight trend of the PLS-LDA scores of the untreated control samples towards more negative coefficients as time progresses indicates that also growth-dependent biochemical changes as they occur when evolving from the mid to the late exponential growth phase might contribute to the PLS-LDA to recognize untreated cases.

Summary and conclusion

It was shown that micro-Raman spectroscopy can be used to describe biochemical changes within bacteria caused by the action of a drug as shown for the interaction of vancomycin with E. faecalis. Observed spectral changes can be explained using biological knowledge of the mode of action of the antibiotics. Furthermore, it could be shown that the observed spectral changes are not caused by the pure presence of the drug. Observing the effect rather than the presence of the antibiotic is a necessary (though not sufficient) condition for the future development of an analytical method to detect antibiotic susceptibility of unknown strains and maybe even species.

The spectral differences observed in the presence and absence of vancomycin were used to train a classification model which is able to discriminate between vancomycin-treated and untreated groups of E. faecalis. The quality of the model was assessed with an independent test data set recorded by an independent operator yielding prediction accuracies >90 %. This marks the high potential of the presented method for further diagnostic or pharmacological tests.