Rapid identification of community-associated methicillin-resistant Staphylococcus aureus by Fourier transform infrared spectroscopy

Available online at www.sciencedirect.com

Diagnostic Microbiology and Infectious Disease 70 (2011) 157 – 166 www.elsevier.com/locate/diagmicrobio

Bacteriology

Rapid identification of community-associated methicillin-resistant Staphylococcus aureus by Fourier transform infrared spectroscopy☆ Nassim M. Amialia,b,⁎, George R. Goldingc , Jacqueline Sedmanb , Andrew E. Simord , Ashraf A. Ismailb a Quelab Laboratories Inc., Montreal, QC, Canada McGill IR Group, McGill University, Montreal, QC, Canada c National Microbiology Laboratory, Public Health Agency of Canada, Winnipeg, MB, Canada d Sunnybrook Health Science Centre, Toronto, ON, Canada Received 30 July 2010; accepted 17 December 2010 b

Abstract The emergence of community-associated methicillin-resistant Staphylococcus aureus (CA-MRSA) carrying Panton–Valentine leukocidin is a worldwide problem. Their identification is based currently on costly and complicated molecular methods. This article describes a simple method for differentiating CA-MRSA from hospital-associated (HA) epidemic MRSA pulsed-field gel electrophoresis types using Fourier transform infrared (FTIR) spectroscopy. The 47 CA-MRSA isolates included 3 Southwest Pacific (resembling USA1100), 24 CMRSA7 (resembling USA400/MW2), 19 CMRSA10 (resembling USA300), and 1 European ST80, while HA-MRSA were represented by 27, 16, 11, 15, 7, and 8 Canadian epidemic isolates CMRSA1 through CMRSA6 respectively, plus 25 nontyped Canadian HA-MRSA. Principal component analysis (PCA), self-organized maps (SOMs), and the K-nearest neighbor (KNN) method were used to cluster the isolates based on chemometric analysis of FTIR spectra of dried films of stationary-phase cells grown on Que-Bact® Universal Medium No. 2 (Quelab Laboratories, Montreal, QC, Canada). First-derivative normalized data from a single narrow spectral region (1361–1236 cm−1, suggesting differences in protein amide III and nucleic acid phosphodiester contents) allowed 98% correct classification by KNN, 93% by SOMs, and 92% by PCA. FTIR spectroscopic analysis of cells grown on Que-Bact® Universal Medium No. 2 offers a rapid and simple alternative to molecular methods for routine identification of CA-MRSA epidemic isolates. © 2011 Elsevier Inc. All rights reserved. Keywords: Fourier transform infrared spectroscopy; Community-associated MRSA; Hospital-associated MRSA; S. aureus

1. Introduction Known for decades as a major cause of nosocomial infections, methicillin-resistant Staphylococcus aureus (MRSA) in recent years has evolved into a highly virulent form known as community-associated MRSA (CA-MRSA). This emerging pathogen infects skin and soft tissue and is associated with increased mortality and morbidity (Mulvey et al., 2005; Wenzel et al., 2007). ☆

Conference or congress proceedings and brief communications: A part of this work was presented as a poster at the 49th Annual Interscience Conference on Antimicrobial Agents and Chemotherapy (ICAAC), San Francisco, CA. Session: 079, D-741. ⁎ Corresponding author. Ph: +1-514-277-2558ext. 13; fax: +1-514-2774714. E-mail address: [email protected] (N.M. Amiali). 0732-8893/$ – see front matter © 2011 Elsevier Inc. All rights reserved. doi:10.1016/j.diagmicrobio.2010.12.016

Most CA-MRSA clones are distinguishable from typical hospital-associated MRSA (HA-MRSA) on the basis of genotype and phenotype (David et al., 2008; Strommenger et al., 2008). They have narrower antibiotic resistance patterns, carry smaller mobile staphylococcal chromosomal cassettes (i.e., SCCmec types IV and V, in contrast with types I–III, Vandenesch et al., 2003), different toxin resistance determinants (Kluytmans-Vandenbergh and Kluytmans, 2006; Vandenesch et al., 2003), and in most cases, the Panton–Valentine leukocidin or PVL genes lukS and lukF, believed to increase virulence (Genestier et al., 2005; Vandenesch et al., 2003). Certain CA-MRSA clones initially appeared continent-specific, such as the North American pulsed-field gel electrophoresis (PFGE) types CMRSA7/ USA400/MW2 (ST1-MRSA-IV) and CMRSA10/USA300 (ST8-MRSA-IV) (Christianson et al., 2007; Tenover et al.,

158

N.M. Amiali et al. / Diagnostic Microbiology and Infectious Disease 70 (2011) 157–166

2006; Zhang et al., 2008), the European clone ST80MRSA-IV (Larsen et al., 2007; Strommenger et al., 2008), and Oceania Western Samoan/USA1100 (ST30-MRSA-IV) clones (Vandenesch et al., 2003). However, recent studies have shown their intercontinental spread, with predominance in distinct geographic regions (Tietz et al., 2005, 2007). The spread of MRSA within the community and from the community into hospitals can be monitored using genotypic analysis and measurement of the genetic relatedness of strains in different geographic regions (Enright et al., 2002; Gomes et al., 2006). PFGE, considered the “gold standard” method for MRSA strain typing (Tenover et al., 1995), molecular analysis of SCCmec (Ito et al., 1999), multilocus sequence typing (MLST) (Enright et al., 2000), polymerase chain reaction (PCR) amplification of target genes (Zhang et al., 2008), and spa typing have all improved the value of epidemiologic studies of CA-MRSA (Hallin et al., 2007). However, these techniques all run into limitations. Specific spa types may occur in more than one CA-MRSA clone (Deurenberg et al., 2007; Strommenger et al., 2008) and even PVL is not entirely reliable (Rossney et al., 2007). Additional methods to identify markers or combinations of markers are necessary for certain recognition of CA-MRSA. A combination of PFGE, MLST, SCCmec analysis and PCR (PVL gene characterization) might be adequate, but would be too timeconsuming and expensive (both the equipment and highly skilled personnel required) for most clinical laboratories. Monitoring the epidemiology of CA-MRSA clones and distinguishing them from HA-MRSA are important especially in empirical therapy, since CA-MRSA clones are more likely to be clindamycin-susceptible (David et al., 2008). One interesting alternative to molecular methods is FTIR spectroscopy, which has been studied over the past decade as a means of identifying microbial species and subspecies (Helm et al., 1991; Mariey et al., 2001). This nondestructive method provides a quantitative profile of the overall biochemical composition (DNA, RNA, proteins, membrane and cell wall components) of intact cells and thus “whole organism fingerprinting” in the form of an infrared absorbance spectrum. By analyzing spectra using a variety of chemometric techniques, closely related strains can be differentiated (Mariey et al., 2001; Lamprell et al., 2006; Van der Mei et al., 1993). FTIR spectroscopy is thus a promising tool for epidemiologic typing (Seltmann et al., 1994; Irmscher et al., 1999). We have shown recently that FTIR in combination with chemometrics provides reliable discrimination between the 5 strains of HA-CMRSA (Amiali et al., 2007a). FTIR typing appears to offer advantages over PFGE in terms of speed, no requirement for reagents, ease of interpretation, reproducibility, and reliability for long-term, nationwide epidemiologic surveillance (Amiali et al., 2007a; Irmscher et al., 1999). The aim of this study was to determine an FTIR spectral region or combination of regions that reflects a biochemical feature peculiar to CA-MRSA and thus provide a substitute

for descriptive epidemiology in the definition of CA-MRSA strain types.

2. Materials and methods 2.1. CA-MRSA and HA-MRSA isolates From 1995 to 2004, 38 hospitals belonging to the Canadian Nosocomial Infection Surveillance Program (CNISP) collected over 9300 MRSA isolates, which were typed using PFGE for national surveillance purposes. A Canadian epidemic PFGE strain type has been defined as one that is clinically significant and isolated from 5 or more hospital sites or from 3 or more geographical regions across Canada (Simor et al., 1999). Initial surveillance from 1995 to 1999 identified 6 epidemic types of MRSA in Canada, CMRSA1 to CMRSA6 (Simor et al., 2002), the “C” meaning Canadian and not to be confused with “communityassociated.” Since 1999, 4 new epidemic types (CMRSA7– CMRSA10) have been identified (Mulvey et al. 2005; Christianson et al., 2007; Simor et al., 2010). Based on PFGE, CMRSA1 to CMRSA6, CMRSA8, and CMRSA9 are recognized as HA-MRSA, while epidemic strains CMRSA7 and CMRSA10 are recognized as CA-MRSA. The isolates used in this study (Table S1), kindly provided by the CNISP, are from the PGFE-typed collection (Simor et al., 2001). The CA-MRSA group included 47 isolates, namely, 3 southwest pacific (resembling USA1100), 24 CMRSA7 (resembling USA400/MW2), 19 CMRSA10 (resembling USA300), and 1 European ST80. The HA-MRSA group consisted of Canadian epidemic isolates CMRSA1 through CMRSA6 (27, 16, 11, 15, 7, and 8, respectively, for a total of 84). The feasibility study was done with CA-MRSA isolates 1–20 and HA-MRSA isolates 21–47 and the spectral regions thus found effective for distinguishing them were tested in the validation study with the inclusion of isolates 48 through 131 (48–74 CA-MRSA; 75–131 HA-MRSA). Twenty-five Royal Victoria HAMRSA isolates were also included in the validation for a total of 156. All isolates were kept as stock cultures stored at −70°C in brain–heart infusion broth with 15% glycerol. 2.2. Molecular typing Strains were typed according to the Canadian standardized PFGE method as previously described (Mulvey et al., 2001). For each of the 8 CMRSA strain types, a single isolate was chosen arbitrarily to represent the characteristic PFGE fingerprint pattern. Multilocus sequence typing was performed using primers and PCR conditions described previously (Enright et al., 2000). Primers were synthesized and sequences were determined at the DNA Core Facility of the National Microbiology Laboratory (Winnipeg, Manitoba, Canada). Sequences for each allele were compared to those in the current database of alleles available at http:// www.mlst.net. SCCmec typing was performed using primers

N.M. Amiali et al. / Diagnostic Microbiology and Infectious Disease 70 (2011) 157–166

and conditions described previously (Oliveira and de Lencastre, 2002). 2.3. FTIR spectroscopic methods Bacterial cells were cultured on Que-Bact® Universal Medium No. 2 (Quelab Laboratories, Montreal, QC, Canada), suspended in sterile saline (0.9% NaCl), and fixed to a zinc selenide (ZnSe) optical window as described previously (Amiali et al., 2007a,b, 2008). Four windows were prepared for each isolate. The windows were examined in the transmission mode over the 4000 to 400 cm−1 (mid-infrared) range at 4 cm−1 resolution on a Bomem MB FTIR spectrometer (ABB-Bomem, Quebec, QC, Canada) equipped with a non-hygroscopic ZnSe beam splitter and deuterated triglycine sulfate (DTGS) detector and continuously purged with dry air. Each FTIR spectrum was obtained by co-adding 64 scans with averaging and comparison to an open-beam background to obtain the equivalent absorbance spectrum. 2.4. Mathematical preprocessing and processing Spectra were processed as described previously (Amiali et al., 2007a,b, 2008). Briefly, spectra converted to MATLAB files were normalized to unit height by vector transformation and transformed to the first derivative to maximize band resolution and minimize the impact of baseline shift. Spectral features of interest were then selected using the singular value decomposition (SVD) algorithm (Golub and van Loan, 1996) with randomly chosen CA-MRSA/HA-MRSA pairs and confirmed by visual examination of the normalized nonderivative spectra. Principal component analysis (PCA) based on the nonlinear iterative partial least squares algorithm was then used to reduce the dimensionality of the multivariate data to a manageable range consisting of a smaller set of representative numbers called scores (Jolliffe, 1986). A nonlinear projection of the first 2 principal components (PC1 and PC2) by self-organizing map (SOM) was then done to provide a sort of clustering diagram (Kohonen, 1995). Finally, cluster analysis was applied using a supervised procedure, namely, the Knearest neighbor (KNN) algorithm (Adams, 1995). The FTIR spectrum of a biological system such as bacterial cells is quite complex and consists of broad bands that arise from the superimposition of absorption by various macromolecules. The broad bands were interpreted as follows: 3200–3100 cm−1 => N-H of amide A in proteins; 3000–2800 cm−1 => C-H of CH2 in fatty acids and CH3 in lipids and proteins; 1750–1715 cm−1 => C=O from lipid esters, carboxyl groups, and nucleic acids; 1700–1500 cm−1 => C=O and N-H in amide I and amide II from proteins; 1470–1240 cm−1 => C-H of CH2 in lipids and C=O of COO− in amino acids and fatty acids and the amide III band in proteins; 1250–1200 cm−1 => P=O in nucleic acids and phospholipids; 1200–900 cm−1 => various polysaccharide-associated absorptions; 900–600

159

−1

cm => fingerprint region with specific band assignments of dipicolinic acid and P-O mainly in nucleic acids. Tentative assignments to CA-MRSA or HA-MRSA were based on systematic comparison of these major regions of the resolution-enhanced microbial spectra with spectra of known components of intact cells (Naumann, 2000). 3. Results and discussion 3.1. Spectral reproducibility of spectral data The identification of microbial organisms by FTIR spectroscopy has been reported to require strict control of growth medium composition and standardization of growth temperature and time (Naumann, 1984; Bourne et al., 2001) as well as high-quality spectral reference databases (Amiali et al., 2007a,b, 2008). The biochemical composition of microbial cells is particularly dependent on growth phase and younger cells differ significantly from older cells. Since the growth medium has a direct influence on cell composition and hence the spectrum, it is best to avoid using blood-based agar, of which the exact composition may vary considerably. In the present study, this was achieved by using Que-Bact® Universal Medium No. 2. Spectral reproducibility also requires minimizing variability due to sampling methodology (transfer of bacteria from agar to the ZnSe optical window, drying of the bacterial suspension on the optical window) and instrument effects such as baseline drift. The influence of most of these parameters on identification and discrimination has been investigated in detail for a few organisms, and standards have been proposed in the literature (Amiali et al., 2007a,b; Helm et al., 1991; Lefier et al., 1997; Oust et al., 2004; van der Mei et al., 1996) in the hope of achieving data compatibility between different laboratories. For routine microbiological analyses, the steps must be simple, quick, and easy to reproduce. Growing cells on agar, suspending them in physiological saline, ensuring that the suspension is homogeneous and the particle size uniform, and finally placing a drop of suspension to form a thin, transparent film on the optical window constitute a procedure that can meet these criteria. After a large number of studies, we are convinced that the main sources of variability are entirely controllable. An FTIR spectrometer equipped with a non-hygroscopic (insensitive to moisture) ZnSe beam splitter improves reproducibility for long-term intra- and interlaboratory comparisons (Amiali et al., 2007a,b) and is suitable for laboratories in which low humidity cannot be ensured (e.g., due to frequent use of an autoclave). The combination of the standardized sampling procedure, blood-free Que-Bact® Universal medium, and a hygroscopic FTIR spectrometer equipped with a ZnSe beam splitter makes the analysis highly reproducible. Reproducibility was examined first by comparison of the spectra of the 4 optical windows prepared for a given sample. The use of 4 different culture plates ensured that the cell

160

N.M. Amiali et al. / Diagnostic Microbiology and Infectious Disease 70 (2011) 157–166

suspension was representative of the original colony of each strain or isolate. Although there was variability due to baseline shift and factors such as dried film thickness, the spectra showed excellent reproducibility in terms of relative peak intensities. Pairwise comparison of quadruplicate spectra yielded an average correlation coefficient of r = 0.97. These variations were minimized by peak height normalization to unit height and by first-order derivation, which highlight spectral shapes and contours and remove the effects of baseline shift. 3.2. Spectral feature selection and interpretation Selection of spectral regions for the differentiation of HA-MRSA from CA-MRSA was performed by SVD, which reveals subtle differences based on specific spectral features or narrow spectral regions rather than on the whole spectrum or large portions of it that are difficult to distinguish visually. SVD of randomly chosen pairs of spectra of HA-MRSA and CA-MRSA isolates showed significant differences in 3 spectral regions: (i) the narrow region 941–932 cm−1 (Fig. 1A), assignable to C-O-C and C-O-P symmetric stretching in cell wall oligosaccharides or polysaccharides (Naumann, 2000); (ii) the broader 1366– 1240 cm−1 (1361–1236 cm−1) region (Fig. 2A), assignable to complex absorption profiles arising from CH2 and CH3 bending modes of lipids and proteins, from amide III (∼1312 cm−1) of proteins (Naumann, 2000) and from asymmetric stretching of the phosphodiester backbone (∼1242 cm−1) of DNA and RNA (Nelson, 1991; Wong et al., 1991); and (iii) the narrow region 1400–1410 cm−1 (Fig. 3A), which may be attributable to C=O symmetric stretching vibrations of COO− functional groups of amino acid side chains and free fatty acids (Naumann, 2000). Visual inspection of the spectra confirmed that there were distinct differences in both regions, as shown in Figs. 1A, 2A, and 3A, most notably a single band centered at ∼936 cm−1, the intensity of which is higher in the spectra of the CA-MRSA isolates (Fig. 1A). 3.3. Differentiation of CA-MRSA and HA-MRSA based on region 941–932 cm−1 3.3.1. Principal component analysis PCA is commonly employed to reduce the dimensionality of spectral data and to obtain preliminary information about data distribution. The 941–932 cm−1 region was selected to compare the spectra of 20 CA-MRSA and 27 HA-MRSA in a preliminary study. The plot of PC1 scores against PC2 scores revealed 2 distinct clusters (Fig. 1B), which were confirmed to represent complete separation (100% correct classification) of CA-MRSA and HA-MRSA. Since PCA is an unsupervised technique, this separation was obtained without using the information that the samples belong to 3 different groups and is indicative of definite spectral differences between the CA-MRSA and HA-MRSA.

Fig. 1. (A) FTIR spectra of CA-CMRSA and HA-CMRSA including region 941–932 cm−1; (B) scores plot of the first 2 PCs; (C) SOM obtained using the first 2 PCs.

N.M. Amiali et al. / Diagnostic Microbiology and Infectious Disease 70 (2011) 157–166

Fig. 2. (A) FTIR spectra of CA-CMRSA and HA-CMRSA including the region 1361–1236 cm−1; (B) scores plot of the first 2 PCs; (C) SOM obtained using the first 2 PCs.

161

Fig. 3. (A) FTIR spectra of CA-CMRSA and HA-CMRSA in the region 1410–1400 cm−1; (B) scores plot of the first 2 PCs; (C) SOM obtained using the first 2 PCs.

162

N.M. Amiali et al. / Diagnostic Microbiology and Infectious Disease 70 (2011) 157–166

3.3.2. Self-organizing map Information about spatial relationships and data clustering can be obtained by visual inspection of an SOM generated by an unsupervised neural network algorithm. The SOM of size [12×6] was generated by nonlinear projection of PC1 and PC2 as input, using 4 iterations for rough training and 16 epochs for the fine-tuning phase (data not shown), giving a final quantization error of 0.168 and a final topological error of 0.043. The learning rate decreased linearly to zero during the fine-tuning phase. Visual inspection of the labeled SOM (Fig. 1C) indicated the presence of 2 clusters corresponding to CA-MRSA and HA-MRSA with 100% correct classification. 3.3.3. Clustering by the KNN algorithm Supervised cluster analysis was performed by applying the KNN algorithm to the response in region 941–932 cm−1 using half of the data set (100 spectra) as the training set and the other half as the prediction set. All of the spectra in the prediction set were thus classified correctly at K = 1 to K = 8 (100% correct classification, as in PCA). 3.4. Differentiation of CA-MRSA and HA-MRSA based on region 1361–1236 cm−1 3.4.1. Principal component analysis The first 2 principal components accounted for over 99.53% of the total variance, with PC1 and PC2 alone accounting for 99.13% and 0.40%, respectively. The scores plot for PC1 versus PC2 (Fig. 2B) revealed a clustering of the data that corresponded to a clear and complete separation of CA-MRSA and HA-MRSA into 2 distinct groups with 100% correct classification. The 1361–1236 cm−1 region thus appears to be as reliable as the 941–932 cm−1 region for the differentiation of CA-MRSA and HA-MRSA. 3.4.2. Self-organizing map Two distinct clusters of CA-MRSA and HA-MRSA were clearly visible on the PCA plot and on an SOM of size [14×5] generated by nonlinear projection of the PC1 and PC2 scores for this region using 4 iterations for rough training and 15 epochs for the fine-tuning phase (data not shown). The final quantization error was 0.155 with a final topological error of 0.043. Based on visual inspection of the labeled SOM (Fig. 2C), CA-MRSA and HA-MRSA were separated into 2 clusters corresponding to 100% correct classification. The SOM classification of CA-MRSA and HA-MRSA based on region 1361–1236 cm−1 is similar to that of PCA, although SOM may be considered a nonlinear generalization of PCA. 3.4.3. Clustering by the K-nearest neighbor algorithm The classification obtained by KNN using the region 1361–1236 cm−1 with half of the data set (100 spectra) as the training set and the other half as the prediction set yielded 100% correct classification. All of the spectra in the prediction set were thus classified correctly at K = 1 and K = 2, as they were with PCA and SOM.

3.5. Differentiation of CA-MRSA and HA-MRSA based on region 1410–1400 cm−1 3.5.1. Principal component analysis The scores plot for PC1 versus PC2 (Fig. 3B) revealed a clustering of the data that corresponded to a clear separation of CA-MRSA and HA-MRSA into 2 distinct groups. However, one CA-MRSA isolate (WSPP 1) fell within the HA-MRSA cluster, yielding a classification only 98% correct. The 1410–1400 cm−1 region thus appears to be somewhat less reliable than the 1361–1236 cm−1 and 941– 932 cm−1 regions for the differentiation of CA-MRSA and HA-MRSA. 3.5.2. Self-organizing map An SOM of size [18×4] generated by nonlinear projection of the PC1 and PC2 scores for this region using 4 iterations for rough training and 16 epochs for the fine-tuning phase (data not shown) showed 2 distinct clusters of CA-MRSA and HA-MRSA clearly visible on the PCA plot. The final quantization error was 0.093 with a final topological error of 0.037. Based on visual inspection of the labeled SOM (Fig. 3C), CA-MRSA and HA-MRSA were separated into 2 clusters corresponding to the same 98% correct classification with the same misclassified CA-MRSA strain (WSPP 19) as for PCA. 3.5.3. Clustering by the K-nearest neighbor algorithm The classification obtained by KNN using the region 1410–1400 cm−1 was 100% correct at K = 1 to K = 4. As expected, this was higher than obtained with the exploratory tools (PCA and SOM) since the KNN algorithm performs the grouping task by incorporating prior knowledge that the test samples belong to one of the groups in the training set. KNN is a well-known powerful technique for clustering any type of data. Based on this feasibility study and preliminary results, the 1361–1236 and 941–932 cm−1 regions should be considered optimal and more appropriate for the discrimination of CAMRSA from HA-MRSA than the region 1410–1400 cm−1. 3.6. Validation of the selected of spectral regions The validity of the selection of regions 1361–1236 and 941–932 and 1400–1410 cm−1 as the basis for differentiating CA-MRSA and HA-MRSA was challenged with the inclusion of the spectra of another collection of CA-CMRSA and HA-CMRSA in the data set. Classification of the enlarged set was 90%, 93%, and 94% correct for PCA and SOM and KNN, respectively, using region 941–932 cm−1. Twelve HA strains (CMRSA types 1, 2, 4, and 6) were misclassified as CA and 4 CA strains were misclassified as HA using PCA. The SOM method misclassified 7 HA and 4 CA, while 6 HA and 4 CA were misclassified using KNN. For region 1361–1236 cm−1, classification was 92%, 93%, and 98% correct for PCA, SOM, and KNN, respectively, with 11 HA strains (3 CMRSA1, 2 CMRSA2, 1 CMRSA4, and 5 CMRSA6) and a CA-

N.M. Amiali et al. / Diagnostic Microbiology and Infectious Disease 70 (2011) 157–166

CMRSA7 misclassified using PCA. The same 11 HACMRSA strains were misclassified by SOM. Only 3 strains of HA-CMRSA (CMRSA6) were misclassified by KNN, yielding the best classification with this region. For region 1410–1400 cm−1, classification was 88%, 89%, and 95% correct for PCA, SOM, and KNN, respectively, with 15 HA (3 CMRSA1, 6 CMRSA2, 1 CMRSA4, and 5 CMRSA6) and 3 CA (2 CMRSA7 and WSPP1) misclassified by PCA. The same 15 HA strains and 2 strains of CA-CMRSA7 were misclassified by SOM. Four strains each of the CA type (CMRSA10) and the HA type (1 CMRSA2, 1 CMRSA4, and 2 CMRSA6) were misclassified using KNN. Among the 3 regions selected by SVD and validated, 1361–1236 cm−1 was found the most appropriate for distinguishing CA-MRSA from HA-MRSA epidemic strain types. 19 CMRSA10 isolates, 11 CMRSA3 isolates, and 7 CMRSA5 isolates as well as the European ST80 and Western Samoan phage pattern isolates were thus correctly classified. However, 11 HA strains (3 CMRSA1, 2 CMRSA-2, 1 CMRSA4, and 5 CMRSA6) and 1 CA strain (CMRSA7) were misclassified using PCA. The same 11 HA strains were misclassified by SOM. Only 3 HA strains (CMRSA6) were misclassified by KNN, yielding the best classification with this region (Table 1). Based on the PFGE fingerprint patterns and the PFGE dendrogram of these isolates (Fig. 4), the reasons for these discrepancies are not clear. For example, the same PFGE fingerprint patterns in some cases were classified incorrectly using this spectral region. Two of the 3 CMRSA1 isolates misclassified by PCA and SOM (95S-0752 and 05S-1830, Table S1) bore PFGE pattern type 1, while 15 additional isolates were correctly classified using these 2 methods. Indistinguishable PFGE patterns do not preclude genetic

Table 1 CA-CMRSA and HA-CMRSA isolates misclassified by FTIR spectroscopy using spectral region 1361–1236 cm−1 Epidemic Isolate type

PVL PFGE PFGE names (+/−) type

PCA SOM KNN

CMRSA7 CMRSA1 CMRSA1 CMRSA1 CMRSA2

− + − − −

176 1 1 5 1594

HA CA CA CA CA

CA CA CA CA CA

CA HA HA HA HA

CMRSA2 98S-1241 −

30

CA

CA

HA

CMRSA4 98S-1237 ND

50

CA

CA

HA

CA CA CA CA CA

CA CA CA CA CA

HA HA CA CA CA

CMRSA6 CMRSA6 CMRSA6 CMRSA6 CMRSA6

97S-0057 05S-1830 95S-0752 06S-0744 06S-0645

99S-0769 05S-1832 05S-1903 05S-1916 05S-1934

ND = not determined.

− + + + +

68 519 519 519 953

USA400/MW2 USA600 USA600 USA600 USA100/ 800/NY USA100/ 800/NY USA200/ EMRSA16

163

Fig. 4. PFGE dendrogram of MRSA isolates used in this study.

differences able to contribute to misclassification. Such differences might not appear on the gel, especially if associated with the larger bands. Since CA-MRSA strains now occur in the health care setting, the “community-associated” designation based solely on any molecular typing method could cause infection management problems. The FTIR technique described in this study is proposed for identifying particular groups of strain types only and should be used in conjunction with epidemiologic data in the health care context. Using QueBact Universal Medium No. 2, the technique appears to offer a rapid and simple alternative to molecular methods for routine identification of CA-type epidemic MRSA and could play an important role in guiding early empirical treatment. Among the 3 regions selected by SVD and validated, 1361– 1236 cm−1 was found the most reliable for distinguishing the CA from the HA type, giving 98% (153/156) correct classification by KNN, 93% (145/156) by SOM, and 92% (144/156) by PCA. 3.7 Use and potential benefits of FTIR spectroscopy in the clinical setting The FTIR spectroscopic technique used in the present study provides a reliable and easily performed method of

164

N.M. Amiali et al. / Diagnostic Microbiology and Infectious Disease 70 (2011) 157–166

identification of isolated bacterial strains within 19 h. This should make it attractive for routine clinical use. However, concerns persist regarding the ability of laboratory personnel to manage the following: (i) Rigorous control of the growth conditions, in particular culture medium composition, growth time, and temperature, is essential. While investing in proper equipment deals with the latter 2 conditions, culture media are likely to remain a source of anxiety. (ii) Control of ambient humidity is necessary to prevent interference with certain spectral regions. This requires a combination of capital investment and training of personnel and may be viewed as onerous. (iii) The quality (especially the reproducibility) of the spectrum depends mainly on sample homogeneity, especially on particle size and concentration and the thickness of the dried film. Anomalous diffraction by large particles in poorly homogenized samples must be eliminated. Even contamination of the suspension with traces of agar could affect the spectra. All of this implies either automation or more training of personnel. (iv) Since spectral regions corresponding to different components often overlap, mathematical processing must be done to highlight spectral features. However, the precise combination of pre-processing tools used for this purpose must be optimized for each specific type of discrimination task. The influence of growth parameters on bacterial identification and discrimination has been investigated in detail for several genera, and standards have been proposed in the literature (Helm et al., 1991; Lefier et al., 1997; Oust et al., 2004; van der Mei et al., 1996). The exact composition of a synthetic medium is more controllable than that of blood or other ill-defined components. Using Que-Bact® Universal Medium no. 2 for the culture subsequent to reactivation of the frozen specimen on blood agar allowed us to minimize batch-to-batch variations in bacterial growth over a period of 3 years. By streaking a single colony onto 4 plates, we obtain a bacterial cell multiplication factor that dilutes any effect that previous growth on blood agar may have had. Rather than multiply the number of colonies examined to obtain statistical representation of each reactivated isolate, we multiplied the number of isolates to obtain representative sampling of each phenotype. This is closer to actual clinical practice, which must reach a diagnosis based usually on single cultures from single clinical samples. Over 3 years of experience with reactivating frozen specimens has indicated that changing from one sheep blood lot to the next has no appreciable impact on the FTIR spectra of bacteria subsequently grown on Que-Bact® Universal Medium No. 2 (data not shown).

Most FTIR spectrometers are equipped with a KBr beam splitter, which is sensitive to humidity and places temporal limits on the optical stability of the instrument. In practice, a background spectrum is collected at the beginning of a series of analyses and is used to calculate all subsequent sample spectra. Changes in water vapor concentration in the optical path and temperature fluctuations can increase instability over time, thus affecting data quality. For these reasons, subtraction of water effects and correction of the baseline are essential. However, subtraction procedures increase spectral noise and limit interpretation since band contours are distorted. Subtraction factors may even vary over time, thus complicating the procedure. The Bomem MB-104 FTIR spectrometer is equipped with a nonhygroscopic ZnSe beam splitter and a DTGS detector. This greatly improves stability and signal-to-noise ratio in laboratory environments where low humidity cannot be guaranteed. In our experience, neither the variance associated with spotting the suspension on the optical windows nor that associated with growth on different plates interfered with phenotypic classification. In view of colonial variability, we prepared each suspension from several colonies and therefore spotted a very large number of cells (109) on the ZnSe optical window. Variations in sample thickness appear as baseline shifts and differences in integrated peak area. Peak-height normalization (between 0 and 1) and extracting the first-order derivative remove these effects and enhance spectral features. IR spectra of bacterial cells reflect vibrations of molecules in the capsule (if present), the cell wall, the membrane, and the cytoplasm. Despite the availability of several resolution enhancement techniques, it remains difficult to separate overlapping spectral features. Most structural and functional components of different bacteria are identical and therefore produce the same signals, making their spectra look very much alike. The differences are in the quantity and distribution of the different functional groups. If the spectra are not of high quality to begin with, all bets are off. However, even with highquality spectra, data compression and pattern recognition techniques are necessary. Derivative spectra, especially the second derivative, enhance differences but also increase noise. They may also contain artefacts. The first derivative combined with peak-height normalization was found suitable for the differentiation of Staphylococcus phenotypes (Amiali et al., 2007a,b, 2008). Classification of bacteria based on analysis of complex spectral signatures by evaluating peak intensities, frequencies, or half-width differences from a few bands that are resolvable by some means, will generally fail. Furthermore, since thousands of spectra must be analyzed each time, the data evaluation techniques must be efficient. These should include pretreatment algorithms such as spectral quality testing, filtering, and normalization as well as multivariate statistical techniques to achieve data reduction and finally pattern classification. Since most

N.M. Amiali et al. / Diagnostic Microbiology and Infectious Disease 70 (2011) 157–166

supervised statistical methods based on a priori knowledge of grouping or group structure (e.g., artificial neural networks) tend to cause overfitting, a nonsupervised method such as PCA is used, which best conserves the original information for cluster analysis. Even if the above issues were resolved, some clinical practitioners would still object to the fact that IR spectroscopy detects phenotypic rather than genotypic differences, measuring only the biochemical expression of genes under a specific set of conditions rather than the presence of specific genes. This was discussed in our evaluations of FTIR spectroscopy for the rapid identification of glycopeptideintermediate S. aureus (Amiali et al., 2008), epidemiologic typing of MRSA (Amiali et al., 2007a), and identification of coagulase-negative staphylococci (Amiali et al., 2007b). Effort is underway to facilitate the use of FTIR spectroscopy in the clinical diagnostic laboratory, in particular through automation of the chemometric procedure, notably by optimizing spectral pre-processing and the selection of spectral features that are “relevant” to the differentiation of the different categories in question. We are currently developing an expert system to automate both data processing and updating of the differentiation models and thus obtain a more robust overall performance. In addition, we are developing a synthetic broth medium to shorten the culture time from 18 to 4–6 h. Using optical windows of the disposable KCl type rather than ZnSe crystal decreases the cost per sample to $10. A single test provides subspecies identification, epidemiologic typing, and antimicrobial resistance characteristics, information that is otherwise obtained using, respectively, PCR, PFGE, and MLST. The cost of the reagents alone for these 3 methods is about $70. We are also developing a dialog interface to make the procedure easier for personnel who have no spectroscopic training. This interface tests the quality of each measured spectrum based on absorbance values, signal-to-noise ratio, and water vapor line intensity and alerts the user if the spectrum is flawed. In conclusion, microbial identification is presently the most widespread and best developed biomedical application of vibrational spectroscopy. FTIR spectroscopy appears suitable for routine clinical use and should be considered very robust, especially given its potential for automation. No other system currently available offers means of determining subspecies, epidemiologic type, and antimicrobial resistance with comparable throughput and technical simplicity at a price affordable in clinics. The prospect of earlier diagnosis and the promise of nonsubjective screening and diagnostics, combined with advantages over manual fingerprint generating systems in terms of standardization and reproducibility, should be irresistible and thus make FTIR spectroscopy very competitive with current commercial diagnostic systems. Supplementary data associated with this article can be found, in the online version, at doi:10.1016/j. diagmicrobio.2010.12.016.

165

Acknowledgments We thank the Canadian Nosocomial Infection Surveillance Program for participating in this research. We also thank Michael R. Mulvey at the National Microbiology Laboratory for making the success of this collaboration possible.

References Adams MJ (1995) Pattern recognition II: supervised learning. In: Chemometrics in Analytical Spectroscopy. Barnett NW, Ed. Letchworth UK: Royal Society of Chemistry Analytical Spectroscopy Monographs, Turpin Distribution, pp. 123–142. Amiali NM, Mulvey MR, Sedman J, Simor AE, Ismail AA (2007a) Epidemiological typing of methicillin-resistant Staphylococcus aureus strains by Fourier transform infrared spectroscopy. J Microbiol Methods 69:146–153. Amiali NM, Mulvey MR, Sedman J, Louie M, Simor AE, Ismail AA (2007b) Rapid identification of coagulase-negative staphylococci by Fourier transform infrared spectroscopy. J Microbiol Methods 68:236–242. Amiali NM, Mulvey MR, Berger-Bächi B, Sedman J, Simor AE, Ismail AA (2008) Evaluation of Fourier transform infrared spectroscopy for the rapid identification of glycopeptide-intermediate Staphylococcus aureus. J Antimicrob Chemother 61:95–102. Bourne R, Himmelreich U, Sharma A, Mountford C, Sorrel T (2001) Identification of Enterococcus, Streptococcus, and Staphylococcus by multivariate analysis of proton magnetic resonance spectroscopic data from plate cultures. J Clin Microbiol 39:2916–2923. Christianson S, Golding GR, Campbell J, Mulvey MR (2007) Comparative genomics of Canadian epidemic lineages of methicillin-resistant Staphylococcus aureus. J Clin Microbiol 45:1904–1911. David MZ, Glikman D, Crawford SE, Peng J, King KJ, Hostetler MA, BoyleVavra S, Daum RS (2008) What is community-associated methicillinresistant Staphylococcus aureus? J Infect Dis 197:1235–1243. Deurenberg RH, Vink C, Kalenic S, Friedrich AW, Bruggeman CA, Stobberingh EE (2007) The molecular evolution of methicillin-resistant Staphylococcus aureus. Clin Microbiol Infect 13:222–235. Enright MC, Day NPJ, Davies CE, Peacock SJ, Spratt BG (2000) Multilocus sequence typing for characterization of methicillin-resistant and methicillin-susceptible clones of Staphylococcus aureus. J Clin Microbiol 38:1008–1015. Enright MC, Robinson D, Randle G, Feil EJ, Grundmann H, Spratt BG (2002) The evolutionary history of methicillin-resistant Staphylococcus aureus (MRSA). Proc Natl Acad Sci USA 99:7689–7692. Genestier AL, Michallet MC, Prevost G, Bellot G, Chalabreysse L, Peyrol S, Thivolet F, Etienne J, Lina G, Vallette FM, Vandenesch F, Genestier L (2005) Staphylococcus aureus Panton–Valentine leukocidin directly targets mitochondria and induces Bax-independent apoptosis of human neutrophils. J Clin Investig 115:3117–3127. Golub GH, van Loan CF (1996) The singular value decomposition and unitary matrices. In Matrix Computations. Johns Hopkins University Press, Baltimore, MD pp. 70-71, 73. Gomes AR, Westh H, de Lancastre H (2006) Origins and evolution of methicillin-resistant Staphylococcus aureus clonal lineages. Antimicrob Agents Chemother 50:3237–3244. Hallin M, Deplano A, Denis O, De Mendonca R, de Ryck R, Struelens MJ (2007) Validation of pulsed-field gel electrophoresis and spa typing for long-term, nationwide epidemiological surveillance studies of Staphylococcus aureus infections. J Clin Microbiol 45:127–133. Helm D, Labischinski H, Schallehn G, Naumann D (1991) Classification and identification of bacteria by Fourier transform infrared spectroscopy. J Gen Microbiol 137:69–79.

166

N.M. Amiali et al. / Diagnostic Microbiology and Infectious Disease 70 (2011) 157–166

Irmscher HM, Fischer R, Beer W, Seltmann G (1999) Characterization of nosocomial Serratia marcescens isolates: comparison of Fouriertransform infrared spectroscopy with pulsed-field gel electrophoresis of genomic DNA fragments and multilocus enzyme electrophoresis. Zb Bakteriol 289:249–263. Ito T, Katayama Y, Hiramatsu K (1999) Cloning and nucleotide sequence determination of the entire mec DNA of pre–methicillin-resistant Staphylococcus aureus N315. Antimicrob Agents Chemother 43:1449–1458. Jolliffe IT (1986) Principal component analysis. Springer series in statistics. New York.: Springer-Verlag. Kluytmans-Vandenbergh MF, Kluytmans JA (2006) Community-acquired methicillin-resistant Staphylococcus aureus: current perspectives. Clin Microbiol Infect 12(Suppl 1):9–15. Kohonen T (1995) Self-organizing maps. Springer series in information sciences, Vol. 30. New York.: Springer-Verlag. Lamprell H, Mazerolles G, Kodjo A, Chamba JF, Noel Y, Beuvier E (2006) Discrimination of Staphylococcus aureus strains from different species of Staphylococcus using Fourier transform infrared (FTIR) spectroscopy. Int J Food Microbiol 108:125–129. Larsen A, Stegger M, Goering R, Sorum M, Skov R (2007) Emergence and dissemination of the methicillin resistant Staphylococcus aureus USA300 clone in Denmark (2000–2005). Euro Surveill http://www. eurosurveillance.org/em/v12n02/1202-222.asp. Lefier D, Hirst D, Holt C, Williams A (1997) Effect of sampling procedure and strain variation in Listeria monocytogenes on the discrimination of species in the genus Listeria by Fourier transform infrared spectroscopy and canonical variate analysis. FEMS Microbiol Lett 147:45–50. Mariey L, Signolle JP, Amiel C, Travert J (2001) Discrimination, classification, identification of microorganisms using FTIR spectroscopy and chemometrics. Vibrational Spectrosc 26:151–159. Mulvey MR, Chui L, Ismail J, Louie L, Murphy C, Chang N, Alfa M, Canadian Committee for the Standardization of Molecular Methods (2001) Development of a Canadian standardization protocol for subtyping methicillin-resistant Staphylococcus aureus (MRSA) using pulsed-field gel electrophoresis. J Clin Microbiol 39:3481–3485. Mulvey MR, MacDougall L, Cholin B, Horsman G, Fidyk M, Woods S, Saskatchewan CA-MRSA Study Group (2005) Community-associated methicillin-resistant Staphylococcus aureus, Canada. Emerg Infect Dis 11:844–850. Naumann D (1984) Some ultrastructural information on intact living bacterial cells and related cell-wall fragments as given by FTIR. Infrared Phys 24:233–238. Naumann D (2000) Infrared spectroscopy in microbiology. In: Encyclopedia of analytical chemistry. Meyers RA, Ed. Chichester: John Wiley and Sons Ltd, pp. 102–131. Oliveira DC, de Lencastre H (2002) Multiplex PCR strategy for rapid identification of structural types and variants of the mec element in methicillin-resistant Staphylococcus aureus. Antimicrob Agents Chemother 46:2155–2161. Oust A, Moretro T, Kirschner C, Narvhus JA, Kohler A (2004) Evaluation of the robustness of FT-IR spectra of lactobacilli towards changes in the bacterial growth conditions. FEMS Microbiol Lett 239:111–116. Rossney AS, Shore AC, Morgan PM, Fitzgibbon MM, O'Connell B, Coleman DC (2007) The emergence and importation of diverse genotypes of methicillin-resistant Staphylococcus aureus (MRSA) harboring the Panton–Valentine leukocidin gene (pvl) reveal that pvl is a poor marker for community-acquired MRSA strains in Ireland. J Clin Microbiol 45:2554–2563. Seltmann G, Voigt W, Beer W (1994) Application of physico-chemical typing methods for the epidemiological analysis of Salmonella enteritidis strains of phage type 25/17. Epidemiol Infect 113:411–424. Simor AE, Boyd D, Louie L, McGeer A, Mulvey MR, Willey BM, Canadian Hospital Epidemiology Committee, the Canadian Nosocomial Infection

Surveillance Program (1999) Characterization and proposed nomenclature of epidemic strains of MRSA in Canada. Can J Infect Dis 10:333–336. Simor AE, Ofner-Agostini M, Bryce E, Green K, Mulvey MR, Paton S, Canadian Nosocomial Infection Surveillance Program, Health Canada (2001) The evolution of methicillin-resistant Staphylococcus aureus in Canadian hospitals: results of 5 years of national surveillance. CMAJ 165:21–26. Simor AE, Ofner-Agostini M, Bryce E, McGeer A, Paton S, Mulvey MR, Canadian Hospital Epidemiology Committee and Canadian Nosocomial Infection Surveillance Program, Health Canada (2002) Laboratory characterization of methicillin-resistant Staphylococcus aureus in Canadian hospitals: results of 5 years of national surveillance, 1995– 1999. J Infect Dis 186:652–660. Simor AE, Gilbert NL, Gravel D, Mulvey MR, Bryce E, Loeb M, Matlow A, McGeer A, Louie L, Campbell J, Canadian Nosocomial Infection Surveillance Program (2010) Methicillin-resistant Staphylococcus aureus colonization or infection in Canada: National Surveillance and Changing Epidemiology, 1995–2007. Infect Control Hosp Epidemiol 31:348–356. Strommenger B, Braulke C, Pasemann B, Schmidt C, Witte W (2008) Multiplex PCR for rapid detection of Staphylococcus aureus isolates suspected to represent community-acquired strains. J Clin Microbiol 46:582–587. Tenover FC, Arbeit RD, Goering RV, Mickelsen PA, Murray BE, Persing DH, Swaminathan B (1995) Interpreting chromosomal DNA restriction patterns produced by pulsed-field gel electrophoresis: criteria for bacterial strain typing. J Clin Microbiol 33:2233–2239. Tenover FC, McDougal LK, Goering RV, Killgore G, Projan SJ, Patel JB, Dunman PM (2006) Characterization of a strain of communityassociated methicillin-resistant Staphylococcus aureus widely disseminated in the United States. J Clin Microbiol 44:108–118. Tietz A, Frei R, Widmer AF (2005) Transatlantic spread of the USA300 clone of MRSA. N Engl J Med 353:532–533. Tristan A, Bes M, Meugnier H, Lina G, Bozdogan B, Courvalin P, Reverdy ME, Enright MC, Vandenesch F, Etienne J (2007) Global distribution of Panton–Valentine leukocidin-positive methicillin-resistant Staphylococcus aureus, 2006. Emerg Infect Dis 13:594–600. van der Mei HC, Naumann D, Busscher HJ (1993) Grouping of oral streptococcal species using Fourier transform infrared spectroscopy in comparison with classical microbiological identification. Arch Oral Biol 38:1013–1019. van der Mei HC, Naumann D, Busscher HJ (1996) Grouping of streptococcus mitis strains grown on different growth media by FT-IR, Infrared. Phys Technol 37:561–564. Vandenesch F, Naimi T, Enright MC, Lina G, Nimmo GR, Heffernan H, Liassine N, Bes M, Greenland T, Reverdy ME, Etienne J (2003) Community-acquired methicillin-resistant Staphylococcus aureus carrying Panton–Valentine leukocidin genes: worldwide emergence. Emerg Infect Dis 9:978–984. Wenzel RP, Bearman G, Edmond MB (2007) Community-acquired methicillin-resistant Staphylococcus aureus (MRSA): new issues for infection control. Int J Antimicrob Agents 30:210–212. Wong PTT, Wong RK, Caputo TA, Godwin TA, Rigas B (1991) Infrared spectroscopy of exfoliated human cervical cells: evidence of extensive structural changes during carcinogenesis. Proc Natl Acad Sci USA 88:10988–10992. Zhang K, McClure JA, Elsayed S, Louie T, Conly JM (2008) Novel multiplex PCR assay for simultaneous detection of USA300 and USA400 community-associated MRSA strains, mecA and Panton– Valentine leukocidin genes with discrimination of Staphylococcus aureus from coagulase-negative staphylococci. J Clin Microbiol 46:1118–1122.