2.1.

Choice of Species and Strains

Nine bacterial and one yeast species were selected which all belong to the ones most frequently encountered in clinical microbiology. They include six Gram-negative species comprising three Enterobacteriaceae species (E. coli, E. aerogenes, and E. cloacae) and three nonEnterobacteriaceae species (Acinetobacter baumannii, A. johnsonii, and Stenotrophomonas maltophilia). Some of these species are known to be difficult to identify by the phenotypic methods. Four Gram-positive species were added, including three bacterial species (Bacillus cereus, Staphylococcus aureus, and S. epidermidis) and one yeast species, selected as an eukaryotic outlier (Candida albicans). Eight well-characterized strains were selected per species. They were provided by the American Type Culture Collection (Manassas, VA, USA), by the Centers for Disease Control (Atlanta, GA, USA) or were taken from the bioMérieux culture collection.

2.2.

Sample Preparation

Efforts were made to minimize variation in sample handling: standardized time of growth, short elapsed time between culture and measurement, no storage, and single culture medium originating from the same lot. Strains were stored at $-80°\mathrm{C}$ in broth containing glycerol. Before Raman analysis, a first overnight culture was performed on TSA (bioMérieux Ref. 43011) at 37°C (except for A. baumannii and A. johnsonii that were grown at 30°C). This first culture was stored at 4°C and constituted a “stock culture” which was used as source during the measurement campaign (3 weeks of storage at most). TSA was selected as the preferred artificial solid culture medium as it expressed a weaker fluorescence or less pronounced Raman features than other tested suitable media from bioMérieux: Columbia Agar with 5% Sheep Blood (Ref. 43041); mannitol-salt agar containing $2\mu \text{g}/\mathrm{mL}$ of oxacillin (Ref. 43671); Drigalski agar (Ref. 43341); medium dedicated to cultures of E. coli, Proteus, Streptococci (CPS3, Ref. 43541); lactose agar with Bromocresol purple (Ref. 43021); and $\mathrm{TSA}+5\%$ sheep blood (Ref. 43001).

For the macrocolony study, the preparation consisted in picking up colonies from a stock culture, streaking on TSA and culturing for 24 h. For the microcolony study, an intermediate culture was done by picking up colonies from stock culture on TSA and culturing them overnight to revitalize the bacteria. This overnight culture was followed by a 6-h long culture to obtain microcolonies. The time elapsed between the end of growth and reading did not exceed 30 min. Culture temperature was 37°C for all species, except for A. baumannii and A. johnsonii (30°C).

2.3.

Spectroscopic Device and Measurements

Raman spectra were acquired using a LabRam ARAMIS (Horiba Jobin Yvon, Villeneuve d’Ascq, France) micro-spectrometer equipped with a 532-nm laser (Ventus LP 532 50 mW, Laser Quantum, Stockport, UK) and a $-70°\mathrm{C}$ Peltier-cooled CCD detector (Synapse TE cooled, Horiba Jobin Yvon). The acquisition spectral window ranged from 395 to $3075{\mathrm{cm}}^{-1}$, given the choice of a $600\text{line}/\mathrm{mm}$ grating. The 1024 channels yield a spectral resolution ranging from 3.07 to $2.65{\mathrm{cm}}^{-1}$ in the $[470\text{to}1700]{\mathrm{cm}}^{-1}$ region selected for data processing. Optimal acquisition conditions, established experimentally, appeared to be quite different between macrocolonies and microcolonies. The parameters used, respectively, on macrocolonies and microcolonies are summarized in Table 1.

Table 1

Parameters and conditions of Raman spectra acquisition for microbial macrocolonies and microcolonies.

Petri dishes with bacterial cultures were directly transferred from the incubator to the spectrometer and Raman spectra were recorded directly from the grown colonies without any additional preprocessing. To account for most variations in bacterial samples, as well as to avoid significant variation of the material during Raman measurements, the following criteria were applied:

2.4.

Indicators of Spectra Quality

Three indicators were used to assess the quality of a spectrum: the SNR, the relative standard deviation (STD), and the Pearson coefficient of correlation ($R$). The first quality indicator, SNR, is derived from the signal, defined as the mean of the net spectrum in the region of interest [see Sec. 2.5 (iii)], and the noise, defined as the STD of the net spectrum between 1800 and $2000{\mathrm{cm}}^{-1}$ (a region deemed free of Raman signal). The second index, STD, used to estimate reproducibility between spectra (preferably net spectra) of the same species, is defined as the within-species STD of each spectral channel, averaged on all channels. It directly provides a relative (mean) STD for spectra previously normalized by their own mean intensity. The third quality index, used to evaluate similarity between spectra originating from a given strain or species, is the Pearson correlation coefficient ($R$), calculated for each pair of spectra in a given dataset of spectra. Plotting the distribution of $R$ enabled a quick visualization of the homogeneity of a dataset. The mean value of these coefficients is an indicator of the similarity of spectra within the dataset.

2.5.

Data Preprocessing

The preprocessing of the initial Raman spectra is important for the subsequent data analysis and classification since it eliminates or reduces significantly the impact of the nonbacterial variability (e.g., instrumental or stochastic). Four preprocessing steps were used in this study:

All preprocessing was performed automatically in the $\mathrm{R}$ software environment²⁰ using the existing or developed in-house routines. Elapsed time for preprocessing of 100 spectra did not exceed 20 s, including 7 s for cosmic spikes suppression, and 12 s for the background suppression.

2.6.

Classification

Two types of cross-validation were carried out in the so-called “stringent” and “nonstringent” modes.

When performing a classification at the species level in the stringent mode, all spectra belonging to the strain being classified were previously removed from the reference database, thus preventing an artificially almost perfect match. Our biological model contained eight strains per species and 10 species, so 10 strains (one per species) were randomly chosen and simultaneously removed from the reference spectra to constitute the test group. An eightfold cross-validation was therefore sufficient to test all spectra.

In the nonstringent mode, all strains were represented in the reference database. In this case, one-eighth of the spectra of each strain was randomly chosen, removed from the reference database, and tested. In those conditions, an eightfold cross-validation was also sufficient to test all spectra.

The stringent mode is thought to be more representative of a clinical situation, where the exact microbial strain present in the sample of interest is most often absent from the reference database used for the identification. In each mode, 10 cross-validations were done, with randomly chosen eightfold partitions. It allowed for the calculation of mean and STD for the CIR. It has been verified that these values are not significantly modified when the number of cross-validations is increased (e.g., to 100).

Several classification algorithms were tested:

Classification results of each method are summarized by the mean of the CIRs of all species. They are also given with more details in the form of a confusion matrix which consists in a cross-table of actual and found species membership, with classification rates expressed in percentages of the number of tested spectra in the actual species. Sensitivity and specificity for a given species can be obtained from this confusion matrix, since sensitivity simply is the corresponding CIR, and specificity is given, after removing the row and column of that species, by summing in each row and then averaging. Nevertheless, since the classification scheme consists of choosing one class among 10, the specificity is naturally high (90% for a random classifier) whereas getting a high sensitivity (or CIR) is much more challenging (a random classifier would give 10%).

3.1.

Databases Description

The databases of macrocolonies and microcolonies contain 2533 spectra and 1813 spectra, respectively (Table 2), acquired for the same 80 strains from 10 species. Each raw spectrum is subdivided into 1024 channels corresponding to a Raman shift ranging from 395 to $3075{\mathrm{cm}}^{-1}$.

Table 2

List of species, code, number of strains, and number of acquired spectra per species, in macrocolonies and microcolonies databases.

3.2.

Indicators of Spectral Quality

The mean SNR values per species, for both macrocolony and microcolony bases, are listed in Table 3. They show that single spectra acquired on microcolonies are of better quality than those from macrocolonies since their SNRs are approximately 2.7-fold higher than those of the latter (average SNR of 10.3 for microcolonies versus 3.8 for macrocolonies). Moreover, the relative STD of spectra in each species (Table 4) clearly demonstrates that the within-species reproducibility is significantly higher with microcolonies (STD of 0.10) than with macrocolonies (STD of 0.17). This is further illustrated by the Pearson correlation analysis of the E. coli spectra (Fig. 1) quantifying the observed similarities among macrocolonies, among microcolonies, and between macrocolonies and microcolonies. It was observed that spectra correlated with an average $R$ value of 0.97 and 0.98 for macrocolonies and microcolonies, respectively, while the correlation dropped dramatically to 0.48 when comparing the macrocolonies to microcolonies, which clearly shows that their spectra are very different. The Pearson coefficient distribution was narrower for microcolonies, a further indication of the better reproducibility of the microcolony dataset.

Table 3

Average signal-to-noise ratio (SNR) per species for macrocolonies and microcolonies studies.

Table 4

Mean standard deviation (STD) of normalized net spectra per species for macrocolonies and microcolonies studies.

Fig. 1

Frequency distribution of $R$ Pearson coefficients when correlating normalized net spectra of Escherichia coli (i) from microcolonies $2\times 2$, (ii) from macrocolonies $2\times 2$, and (iii) from microcolonies with spectra from macrocolonies.

3.3.

Macrocolonies

Figure 2 presents raw spectra of E. coli strain API 9203096, before and after normalization. Because of the high variability in spectral background in both cases, it was deemed essential to remove the background before classifying, as is commonly performed in spectroscopic analysis.

Fig. 2

Raw spectra of strain “ESH-COL API 9203096,” (a) on the full Raman-shift range without normalization and (b) restricted and normalized on the region of interest (macrocolonies).

The mean, normalized, preprocessed spectra for each species are shown in Fig. 3, for the (smoothed) raw spectra, the net spectra after background suppression, and the first derivative (second derivative is not shown). It is important to notice that nearly the same peaks are present in all species. The main differences between species lie in the relative abundances, an observation easily explained by the fact that all species have similar biochemical compositions and are therefore characterized by identical chemical bounds, differing mostly by the relative abundance of particular biogroups. One exception is presented by S. aureus, where six of the eight strains had the characteristic golden pigmentation, due to carotenoids and associated with the intense and very specific Raman peaks²⁸ at 1160 and $1520{\mathrm{cm}}^{-1}$, clearly visible in the averaged signal (Fig. 3). The corresponding principal component analysis (PCA) plots (Fig. 3, bottom) suggest that the background subtraction step, by the peak-clipping algorithm or by deriving the spectra, substantially improves the discrimination between species.

Fig. 3

Mean spectra per species for macrocolonies, for (a) raw spectra, (b) net spectra, (c) first derivative, and (d, e, f) corresponding PCA plots (first two components, with one color per species). NB: Orange strains of Staphylococcus aureus have been excluded from PCA to provide more details.

The CIRs of the five classification methods tested on these four sets of raw or preprocessed spectra are shown for stringent and (partially) for nonstringent modes in Table 5 (upper part). Figure 4 shows the confusion matrix obtained for the stringent mode in the best configuration, which is the rQDA method applied to the first derivative spectra. In this case, the CIR is $94.1\pm 0.6\%$. It decreases to $93.3\pm 0.9\%$, when the realignment step is skipped. Same results are presented in terms of sensitivity (i.e., CIR) and specificity in Table 6. As anticipated in Sec. 3, specificities are very high, with an average value of 99.3%.

Table 5

Correct identification rates (CIR) obtained for the macrocolonies and microcolonies in stringent and nonstringent modes with the five classification methods and the four sets of preprocessed spectra.

Fig. 4

Confusion matrix for macrocolonies with rQDA method in stringent mode on first derivative spectra ($\mathrm{CIR}=94.1\pm 0.6\%$).

Table 6

Sensitivity and specificity calculated from confusion matrix of Fig. 4 (rQDA on first derivative spectra for macrocolonies and stringent cross-validation).

We observe that the rQDA method still provides the best CIRs for the net and second derivative signals. It is also noticeable that the CIRs obtained with the raw spectra and the more advanced classification methods (92.0% with LDA, 92.7% with rQDA, and 93.3% with SVM) are barely lower than the best CIR. This is very different from the results obtained with the ED and KNN, where subtracting the background (by the peak-clipping algorithm or by deriving) clearly improves the CIR, as suggested by the PCA plots. This is a direct consequence of the close relationship between PCA and ED.

In the nonstringent mode, the best CIR is $99.8\pm 0.1\%$ and was obtained with SVM. It is worthwhile to notice that it was obtained for the raw spectra and is very close to the one obtained with SVM on first derivative ($99.7\pm 0.1\%$).

3.4.

Microcolonies

Figure 5 shows the same set of raw and preprocessed spectra as in Fig. 3, also with the corresponding PCA plots, here for the microcolonies. We also observed in the PCA plots that the background subtraction by peak-clipping or the derivative improves the discrimination between species, although it is seemingly poorer than with macrocolonies. Interestingly, the yellow strains of S. aureus did not have detectable Raman peaks specific to carotenoids. This is in agreement with the absence of observable pigmentation of those strains at the microcolony stage.

Fig. 5

Mean spectra per species for microcolonies, for (a) raw spectra, (b) net spectra, (c) first derivative, and (d, e, f) corresponding PCA plots (first two components, with one color per species).

The CIRs for the microcolonies are shown in the lower part of Table 5 for the stringent and nonstringent modes. Figure 6 shows the confusion matrix obtained in the best stringent configuration, which is applying the SVM method to the first derivative spectra. The corresponding CIR is $91.5\pm 0.4\%$ (decreasing to $88.7\pm 0.4\%$ when the realignment is skipped). Once again, the ED and KNN are the only classification methods that showed a clear improvement when background signal was removed compared to raw spectra. In the nonstringent mode, the best CIR is $98.0\pm 0.1\%$ and was obtained with SVM on raw spectra, as for macrocolonies.

Fig. 6

Confusion matrix for microcolonies with SVM method in stringent mode on first derivative spectra ($\mathrm{CIR}=91.5\pm 0.4\%$).

4.1.

Misclassifications and Taxonomy

Species that are the most difficult to differentiate by Raman spectroscopy are also the ones being very close in their taxonomic position, as defined by using conventional phenotypic and molecular methods. With macrocolonies as well as with microcolonies, the lowest CIRs were observed inside the Enterobacteriaceae family (for E. aerogenes, E. cloacae, and E. coli) and inside the Acinetobacter genus (A. johnsonii and A. baumannii). Other significant errors occurred with E. cloacae instead of A. johnsonii for macrocolonies and S. maltophilia instead of E. cloacae for microcolonies. Confusions confined under the genus level accounted for 89% of all errors for macrocolonies and 44% for microcolonies. When errors inside the Enterobacteriaceae family are included, the proportion increased to 93% of all errors with macrocolonies and 85% with microcolonies. These results are similar to earlier findings²⁸ showing confusions between Enterococcus faecalis and E. faecium.

4.2.

Comparison Between Macrocolonies and Microcolonies and Influence of the Agar Signal

For any given species, spectra acquired on microcolonies were very different from spectra acquired on macrocolonies. This was illustrated by the Pearson correlation analysis of the E. coli spectra (Fig. 1), which clearly implies that identification is only possible if the spectra of the tested sample are acquired at the same culture age as the reference spectra forming the database, at least for times of culture where the growth stage is expected to be very different as is the case in this study.

These differences are more directly shown in Fig. 7 for species E. coli by comparing the mean net spectra of macrocolonies and microcolonies. The Raman spectrum of TSA is also shown to demonstrate that the difference cannot be due to the contribution of the underlying agar as the TSA medium has a few characteristics unique peaks (although it also clearly shows multiple peaks in the same region of interest as the colonies). We suggest that the large differences observed between microcolonies and macrocolonies are indeed due to the biological differences and not to the underlying growth medium (a result to be expected for microorganisms at different growth stages).

Fig. 7

From top to bottom: average normalized net Raman spectra (offset for clarity) of (a) TSA culture medium, (b) E. coli microcolonies ($\pm 1\mathrm{SD}$ in thin traces), (c) E. coli macrocolonies, and (d) difference between microcolonies and macrocolonies.

The following three arguments support our proposition:

4.3.

Comparison of CIRs Between Macrocolonies and Microcolonies

Another clear difference between macrocolonies and microcolonies lies in the fact that the CIRs from microcolonies were significantly lower than those obtained from macrocolonies (see CIRs in Table 5 and confusion matrices in Figs. 4 and 6). The reasons for this observed drop of performance were not elucidated. Logical explanations could be a lower specific Raman signal from microcolonies because of the smaller quantity of biological material, hence a lower SNR, or a higher contribution of the underlying culture medium, due to the confocal volume extending beyond microcolony depth, hence a less specific signal. These assumptions are denied by the fact that despite the reduced overall signal, the spectra quality actually appeared to be better for microcolonies than for macrocolonies, since they showed a higher SNR (10 versus 4; see Table 3). Moreover, Fig. 1 illustrates (for E. coli) that the normalized net spectra were slightly better correlated within species for the microcolonies than for the macrocolonies and Table 4 shows (for all species) that they had a lower STD on average (0.10 versus 0.17). This is in agreement with the already cited study by Choo-Smith et al.⁷ who observed that microcolonies are more homogeneous in their composition than macrocolonies.

A possible way to reconcile the fact that lower CIRs are observed for microcolonies despite a lower dispersion inside each class and a higher SNR could be to assume that the microcolonies show less chemical composition differences between species than macrocolonies, therefore making the discrimination between species more difficult. Possibly the Raman peaks associated with a high metabolic activity are not that species-specific, otherwise discrimination would be improved as the relative importance of those marker peaks decreases in the stationary phase.

4.4.

Influence of Background Subtraction on Classification

It was already mentioned that the best classification results (in stringent mode) were obtained for the derivative spectra, which are deemed to be background subtracted, with the rQDA and SVM methods (for macrocolonies and microcolonies, respectively), but that the same methods give barely lower CIRs on raw spectra (see Table 5). This seems contradictory with the PCA plots of Figs. 3 and 5 which suggest that the background subtraction substantially improves discrimination between species, as actually observed in the CIRs obtained with the ED and KNN methods. This logically questions the astonishingly small improvement in the classification performance of LDA, rQDA, and SVM methods between raw and background-subtracted spectra. Or, one could equivalently ask why they proceed so well with the raw spectra. A part of the answer lies in the fact that, contrary to ED and KNN (closely related with PCA through the ED), those methods are not based on constant weights for all channels throughout the spectrum. They instead determine optimized weights (by taking account of the common or specific variance-covariance matrix for LDA and rQDA, or by searching for the separating hyperplane with maximum margin and minimum classification errors for SVM), which shows equivalent performance in terms of classification to explicitly subtracting the background.

Now we come to the question: is it really necessary to subtract the background? In fact, even if raw spectra give good classification results, it seems dangerous to keep background in the signal of interest. Indeed, background depends a lot on experimental conditions. In our study, sources of variability were minimized by for instance choosing a constant acquisition time for spectra for all 80 strains of our experimental collection and selecting a unique culture medium. This is a valuable approach here, as we wanted to evaluate the ultimate discriminatory power of Raman spectroscopy under nearly ideal conditions while minimizing possible confounding factors. For obvious reasons, it might be not very practical, as for instance a common time of acquisition is unlikely to be found as the number of species and strains will be significantly expanded. Variable acquisition time will very likely cause an increase in fluorescence background variability as photobleaching is concurrent to acquisition. Also, multiple media and low temperature storage will certainly be in frequent use in real life. For those reasons, removing the background seems mandatory because it is likely to render the Raman procedure more robust.