Rapid and Reliable Identification of Staphylococcus aureus Capsular Serotypes by Means of Artificial Neural Network-Assisted Fourier Transform Infrared Spectroscopy

Selection of the strain set and optimization of growth conditions for FTIR spectroscopic capsular biotyping.A comprehensive set of strains, including human and veterinary clinical isolates, was compiled (Table 1). All strains were initially serotyped by colony immunoblot, and all NT isolates, showing no immunoreactivity against CP5 and CP8 antibodies, were additionally tested by the more-sensitive double immunodiffusion assay. S. aureus Reynolds CP5 and its isogenic mutants Reynolds CP8 and Reynolds CP− were included as control strains for CP expression (31). A reference strain set was defined, comprising a representative selection of 10 isolates of each CP type (including the Reynolds and its isogenic mutants). For all isolates, the genetic background of the cap-specific allele was determined. cap allele determination of NT isolates showing no CP expression revealed that they carry either the cap allele cap5 or cap8 or none or both alleles. From the 37 NT isolates, 65% carried the cap5 allele and 30% carried the cap8-specific allele. The remaining isolates (5%) did not show any amplification products with the cap allele-specific primers (Table 1). NT isolates for the reference strain set were selected in such a way that the different cap genetic backgrounds were well represented.

The control strains Reynolds CP5, CP8, and CP− were used to determine the culture condition providing the highest discriminatory power for differentiation of CP types by FTIR spectroscopy. Strains were grown for 24 h under the following conditions: (i) TSA at 30°C, (ii) TSA at 37°C, (iii) TSA supplemented with 2% NaCl at 37°C, and (iv) Columbia base agar supplemented with 2% NaCl at 37°C. Bacteria were subsequently subjected to FTIR spectroscopic analysis (for details, see Materials and Methods). TSA at 30°C (condition i) was included since this culture condition is routinely used for bacterial species identification by means of FTIR spectroscopy (38) and has the advantage of allowing species identification and CP biotyping of S. aureus strains using a single FTIR spectrum, recorded only once. Columbia base agar supplemented with 2% NaCl at 37°C (condition iv) was included in the set of tested cultivation conditions because it is well established for immune-based capsule serotyping. Four independent experiments were performed for each growth condition. PCA, an unsupervised multivariate statistical method, was employed to investigate the discriminatory features of the different growth conditions for S. aureus capsule typing. The score plot for PC1 and PC2 revealed a clear clustering of spectral data according to growth conditions and S. aureus CP types, respectively. As shown in Fig. 1, all growth conditions tested are suitable for discrimination of the Reynolds CP5, CP8, and CP− strains. However, the distance between the different strains depended upon the chosen growth conditions. The determination of spectral distance values, which are a measure of dissimilarity corresponding to the nonoverlapping areas of the spectra, confirmed these findings: (i) TSA at 30°C, 0.55; (ii) TSA at 37°C, 0.33; (iii) TSA supplemented with 2% NaCl at 37°C, 0.41; and (iv) Columbia base agar supplemented with 2% NaCl at 37°C, 0.36. Hence, for FTIR spectroscopy-based CP determination, TSA growth medium is superior to Columbia agar base and a growth temperature of 30°C revealed a better resolution than 37°C. Overall, the best discrimination between the different CP types was achieved for growth on TSA at 30°C (Fig. 1; depicted in red), and these conditions were selected for developing the FTIR capsule serotyping system.

• Open in new tab
• Download powerpoint

Fig 1

PCA was performed on 2nd-derivative and vector-normalized spectra in the spectral range of 1,200 to 800 cm⁻¹ recorded from the Reynolds CP5 strain and its isogenic capsule mutants CP8 and CP− grown at the following conditions: (i) TSA at 30°C and 24 h (red), (ii) TSA at 37°C and 24 h (green), (iii) TSA and 2% NaCl at 37°C and 24 h (yellow), and (iv) Columbia base and 2% NaCl at 37°C and 24 h (blue). PC1 was plotted against PC2. PC1 shows a distinct separation of the different growth conditions, whereas PC2 discriminates the Reynolds CP5, CP8, and CP− strains.

Reliable discrimination between encapsulated S. aureus CP5 and CP8 by FTIR spectroscopy using HCA.Hierarchical cluster analysis (HCA) was employed to assess the overall similarity between isolates of the different CP serotypes. In order to extract spectroscopically relevant information and to increase the discriminatory power, the whole recorded spectral range of 4,000 to 500 cm⁻¹ was limited to 1,200 to 800 cm⁻¹. This spectral region contains two blocks of relevant biochemical information: (i) 1,200 to 900 cm⁻¹, the polysaccharide region, and (ii) 900 to 800 cm⁻¹, which forms part of the so-called “fingerprint” region (22). The polysaccharide region is dominated by C-O-C and C-O-P stretching vibrations of polysaccharides and has been used previously to discriminate the E. coli serotypes O:123 and O:4 (27) as well as for the determination of L. monocytogenes serogroups (29). Several spectral differences in this spectral region were also observed between CP5 and CP8 strains. The most prominent differences were located at 1,110 cm⁻¹, 1,097 cm⁻¹, 1,070 cm⁻¹, 1,058 cm⁻¹, 1,030 cm⁻¹, and 975 cm⁻¹ (data not shown). However, it is still unclear which cellular components are responsible for the differences observed.

A detailed manual analysis of the 2nd derivative of the recorded spectra from the different S. aureus strains revealed a highly discriminatory spectral window in the “fingerprint” region, not yet known to be suitable for serotyping. The spectral region of 845 to 810 cm⁻¹ especially showed remarkable differences between CP5 and CP8 strains (Fig. 2a). Specific spectral differences between the two groups are found at 834 cm⁻¹ and 823 cm⁻¹. The same spectral wavenumbers are also highly discriminatory for the encapsulated Reynolds strain (CP5) and its isogenic mutant CP8 (Fig. 2b). In order to gain a better insight into this highly discriminatory spectral region and to test if these differences can be assigned to CP, spectra of the crude polysaccharide capsule extract from the Reynolds strain and its isogenic mutant CP8 were analyzed by FTIR spectroscopy (Fig. 2c) and compared with spectra obtained from intact cells. Since the highest discriminatory power for differentiation of the two serotypes was achieved for the capsule extract, it can be assumed that the major discriminatory components in the spectral range of 845 to 810 cm⁻¹ for discrimination of CP5 and CP8 are related to capsule compounds. Interestingly, it has been described that the spectral region at 860 to 825 cm⁻¹ is specific for an α-anomeric configuration of carbohydrates (39). An anomeric region absorption band located at 834 cm⁻¹ has been reported to be characteristic for α-glycosidic linkage of aldopyranoses in carbohydrates (40). Therefore, it is tempting to speculate that the highly discriminatory spectral band at 834 cm⁻¹ relates to different specific and structural α-and β-glycosidic linkages in S. aureus capsules. And indeed, a specific α-glycosidic linkage type of the N-acetyl-d-fucosamine (D-FucNAc) residues has been previously described for CP8 capsular polysaccharides (41, 42) as follows: type 5, →4)-β-d-ManNAcA-(1→4)-α-l-FucNAc(3OAc)-(1→3)-β-d-FucNAc-(1→, and type 8, →3)-β-d-ManNAcA(4OAc)-(1→3)-α-l-FucNAc-(1→3)-α-d-FucNAc-(1→.

• Open in new tab
• Download powerpoint

Fig 2

FTIR spectra of S. aureus intact cells and capsule extracts in the spectral range of 860 to 800 cm⁻¹, showing the discrimination between CP5 (solid line) and CP8 (dashed line) at 834 cm⁻¹ and 823 cm⁻¹. All isolates were grown under standardized conditions on TSA at 30°C for 24 h. (a) Clinical CP5 (n = 10) and CP8 (n = 10) isolates included in the reference data set. Whole bacteria. (b) Reynolds prototype strain CP5 and its isogenic mutant CP8. Whole bacteria. (c) Reynolds prototype strain CP5 and its isogenic mutant CP8. Crude capsule extracts. (b and c) Spectral data derived from three independent experiments. Each experiment was performed in technical duplicates. A.U., arbitrary units.

As shown in the dendrogram depicted in Fig. 3, HCA of the 2nd derivatives of FTIR spectral data obtained from the compiled S. aureus strain set revealed two major clusters. Cluster A covers all CP5 strains and cluster B covers all CP8 strains, while the NT isolates are interspersed between the clusters. Subcluster A2 is predominated by NT isolates with different cap gene allelic backgrounds (NTcap5, -cap8, and -nt), but some NT isolates are clustered to the CP5 subcluster A1 (one isolate of NTcap5) and to the CP8 subcluster B2 (two isolates of NTcap8).

• Open in new tab
• Download powerpoint

Fig 3

FTIR spectroscopy-based dendrogram of the reference strain set, comprising 10 isolates from serotypes CP5, CP8, and NT. Second derivatives were calculated from 10 replicate spectra of each isolate with a 9-point Savitzky-Golay filter, and subsequent vector normalization was performed for the whole spectral range. Dendrograms were obtained using Ward's algorithm in the spectral range of 1,200 to 800 cm⁻¹ at repro level 30.

Based on their metabolic fingerprints, generated by FTIR spectroscopy, encapsulated CP5 and CP8 strains can be clearly discriminated from each other. But these findings also show that HCA alone is insufficient to discriminate between encapsulated CP5/CP8 and nonencapsulated NT isolates carrying the respective cap5 or cap8 allele or none or both alleles. With respect to implementation in routine diagnostics, an artificial neural network (ANN)-assisted method would also simplify the readout of results because data are clearly classified by the latter method (37).

Identification of S. aureus CP5, CP8, and NT using ANN-assisted FTIR spectroscopy.It has been shown that the discriminatory power of FTIR spectroscopy can significantly be improved by combining it with appropriate chemometrics, especially by employing supervised learning methods, such as ANNs (see, for instance, references 24, 30, and 43). We therefore used an ANN-based approach to optimize our FTIR spectroscopic serotyping system. To establish the ANNs, a panel of human and veterinary clinical strains with known serotypes was compiled (for details, see Materials and Methods). ANNs are simplified models of neural learning processes, which simulate the behavior of biological neural nets. For the used feed-forward three-layer ANN, the information propagates from the input layer, through the hidden layers, to the output layers. The ANN in our study was trained with selected spectral signatures as input data (input neurons) paired with the predefined output classes CP5, CP8, and NT (output neurons). During the iterative training process, the connection weights (numbers of input and hidden neurons) are automatically adjusted until the global error is minimized (44). In this work, several single ANNs and consecutive combinations of ANNs (modular ANNs) were tested to achieve an optimal network performance. The ANN training resulted in a single-level ANN using the 25 most discriminative wavenumbers (input neurons), three hidden neurons, and three output neurons, one for each CP type. For internal validation, one randomly selected spectrum of each isolate from the training set (n = 30) was tested for the correctness of the classification. For 29 out of 30 strains used for the internal validation, a correct classification was achieved. One strain yielded an uncertain result, but it was not incorrectly classified. To assess the ANN performance, an external validation was performed (Table 2). The external validation comprises only isolates (n = 57) which were not included in the training set to develop the ANN method. To simulate routine diagnostic procedures, strains were measured one time for the external validation. In total, 98.2% (56/57) of the isolates were correctly identified. One isolate was not able to be clearly assigned to one of the three CP types (1.8%), but none of the isolates were incorrectly identified. In particular, the strain showing an unknown classification was a nonencapsulated NT isolate.

In general, similar results for internal (96.7%) and external (98.2%) validation of the ANN are a good indicator for the stability of the established ANN model. It also indicates that a significant part of the microbial diversity was covered by the selected strain set (43). This underpins the reliability and robustness of the established typing system.

CP serotyping is commonly performed in combination with cap allele determination, requiring about 3 days for the immune-based detection of CP expression. Usually, isolates are first serotyped by colony immunoblot (1 day) and NT isolates are subsequently confirmed by the double immunodiffusion assay (2 days). FTIR spectroscopy-based capsule biotyping requires only 1 day for growth of bacteria and less than 5 min of sample preparation per strain. In addition, time and resources for animal experiments for CP-specific antibody production are not needed and only a minimum of bacterial biomass (about 10⁸ bacteria) is required. Since the sample preparation procedure is simple and data analysis is automated by the software, no specific training is necessary. Operating costs for FTIR spectroscopy are very low because it does not require specific reagents and ZnSe optical plates used for the measurements are reusable. Overall, FTIR spectroscopy is, compared to the current gold standard for CP serotyping by immune-based detection, much faster, cheaper, and easier to perform. In addition, due to its high-throughput capacities, it may also provide valuable epidemiological information about the chronicity of specific S. aureus infections.

In conclusion, this study demonstrated that ANN-assisted FTIR spectroscopy is a suitable tool for a rapid, cost-effective, and reliable discrimination of S. aureus capsular serotypes CP5, CP8, and NT. Since it is based on the metabolic fingerprint of the bacterium itself, it enables a routine high-throughput screening and can be applied to monitor pathogen adaptation in host environments. Biotyping by FTIR spectroscopy therefore might not only represent an interesting tool for clinical diagnostics but also may contribute to the better understanding of the adaptation mechanisms of S. aureus to the host, which would allow the development of improved treatment regimes in the future, especially for chronic infections.