Localizing and Identifying Living Bacteria in an Abiotic Environment by

Oct 11, 2008 - Mario Krause, Petra Rösch, Benno Radt and Jürgen Popp*. Institute .... Shaleen B. Korch , Joshua M. Stomel , Megan A. León , Matt A...
0 downloads 0 Views 1MB Size
Anal. Chem. 2008, 80, 8568–8575

Localizing and Identifying Living Bacteria in an Abiotic Environment by a Combination of Raman and Fluorescence Microscopy Mario Krause,† Petra Ro¨sch,† Benno Radt,‡ and Ju¨rgen Popp*,†,§ Institute for Physical Chemistry, University of Jena, Helmholtzweg 4, 07743 Jena, Germany, Carl Zeiss AG, Zentralbereich Forschung and Technologie, Carl-Zeiss-Promenade 10, 07745 Jena, Germany, and Institute of Photonic Technology, Albert Einstein Strasse 9, 07745 Jena, Germany A fast, easy, and reliable identification of microorganisms is indispensable in many fields such as medicine, food production, or the pharmaceutical industry. However, in native samples, biotic particles often appear together with abiotic particles. Therefore, it is a prerequisite that biotic particles can be differentiated from abiotic particles appearing in the identification setup. In addition, for many applications, not all microorganisms are of interest but only the living ones. Therefore, in this contribution, different bacteria species were stained with a live/dead staining kit (SYTO 9 and propidium iodide) prior to Raman spectroscopic identification. Since only living and dead microorganisms are getting stained by SYTO 9 or PI, biotic particles can easily be spotted and localized inbetween abiotic particles. By using a Raman laser excitation wavelength outside the absorption band of the dye, fluorescence-free Raman spectra were obtained. The living cells were identified by means of Raman spectroscopy in combination with a support vector machine. Furthermore, the localization of bacterial cells in a mix of abiotic particles is demonstrated. The identification of microorganisms has become indispensable in many fields. In medicine, for example, a fast diagnosis of microbial infections combined with adequate medical treatment will reduce spreading of (epidemic) diseases and enhance new chances of healing.1 In food production,2 drinking water,3,4 or wastewater treatment,5 the microbiological contamination has to be controlled to achieve sufficient customer protection. A further application is the air monitoring of a clean room environment to * To whom correspondence should be addressed. E-mail: juergen.popp@ uni-jena.de. † University of Jena. ‡ Carl Zeiss AG. § Institute of Photonic Technology. (1) Maquelin, K.; Kirschner, C.; Choo-Smith, L. P.; van den Braak, N.; Endtz, H. P.; Naumann, D.; Puppels, G. J. J. Microbiol. Methods 2002, 51, 255– 271. (2) de Boer, E.; Beumer, R. R. Int. J. Food Microbiol. 1999, 50, 119–130. (3) Baudart, J.; Coallier, J.; Laurent, P.; Prevost, M. Appl. Environm. Microbiol. 2002, 68, 5057–5063. (4) Lemarchand, K.; Masson, L.; Brousseau, R. Crit. Rev. Microbiol. 2004, 30, 145–172. (5) Zwiener, C.; Frimmel, F. H. Anal. Bioanal. Chem. 2004, 378, 862–874.

8568

Analytical Chemistry, Vol. 80, No. 22, November 15, 2008

ensure and control the quality of products, e.g., silicon wafers or pharmaceuticals.6,7 Although the application of conventional microbiological methods for monitoring single cells in native samples is extremely time-consuming, it is still the gold standard for microbiological identification. The cultivation of prokaryotic cells usually requires more than 48 h.8 Nowadays, a large variety of analysis methods and techniques for a faster microorganism detection are pushed on the market. For rapid total cell counting, use of solid-phase cytometry9 and flow cytometry10,11 technology is widespread. The fast identification of microorganisms by real-time PCR12,13 or mass spectrometry14 is also well established. The PCR analysis is constrained by the availability of special primers for a specific DNA sequence amplification, while to apply mass spectrometry, the samples are destroyed by ionization and further analysis by other analysis methods becomes impossible. An alternative, nondestructive approach to identify microorganisms without adding auxiliary substances is the application of vibrational spectroscopic techniques (IR and Raman spectroscopy), because vibrational spectra yield fingerprint-like information about all chemical components within one cell.15-18 IR and Raman spectroscopy are two methods that yield complementary results. However, Raman spectroscopy is in comparison to IR absorption not influenced by water in the spectroscopic range of interest, which favors Raman spectros(6) Ljungqvist, B.; Reinmuller, B. VTT Symp. 2006, 421–428. (7) Ro ¨sch, P.; Harz, M.; Peschke, K.-D.; Ronneberger, O.; Burkhardt, H.; Schuele, A.; Schmauz, G.; Lankers, M.; Hofer, S.; Thiele, H.; Motzkus, H.W.; Popp, J. Anal. Chem. 2006, 78, 2163–2170. (8) Al-Khaldi, S. F.; Mossoba, M. M. Nutrition 2004, 20, 32–38. (9) Broadaway, S. C.; Barton, S. A.; Pyle, B. H. Appl. Environm. Microbiol. 2003, 69, 4272–4273. (10) Bunthof, C. J.; Bloemen, K.; Breeuwer, P.; Rombouts, F. M.; Abee, T. Appl. Environ. Microbiol. 2001, 67, 2326–2335. (11) Yang, R.; Zou, M. Fenxi Ceshi Xuebao 2004, 23, 124–128. (12) Belgrader, P.; Benett, W.; Hadley, D.; Richards, J.; Stratton, P.; Mariella, R., Jr.; Milanovich, F. Science 1999, 284, 449–450. (13) Hanna, S. E.; Connor, C. J.; Wang, H. H. J. Food Sci. 2005, 70, R49-R53. (14) Richardson, S. D. Anal. Chem. 2000, 72, 4477–4496. (15) Naumann, D. In Encyclopedia of Analytical Chemistry; Meyers, R. A., Ed.; John Wiley & Sons Ltd.: Chichester, 2000; Vol. 10, pp 2-131. (16) Neugebauer, U.; Ro ¨sch, P.; Schmitt, M.; Popp, J.; Julien, C.; Rasmussen, A.; Budich, C.; Deckert, V. ChemPhysChem 2006, 7, 1428–1430. (17) Neugebauer, U.; Schmid, U.; Baumann, K.; Ziebuhr, W.; Kozitskaya, S.; Deckert, V.; Schmitt, M.; Popp, J. ChemPhysChem 2007, 8, 124–137. (18) Neugebauer, U.; Schmid, U.; Baumann, K.; Simon, H.; Schmitt, M.; Popp, J. J. Raman Spectrosc. 2007, 38, 1246–1258. 10.1021/ac8014559 CCC: $40.75  2008 American Chemical Society Published on Web 10/11/2008

copy for the analysis of biological cells.15 Furthermore, Raman spectroscopy becomes even more interesting for the analysis of microbiological samples when combining the spectrometer with a microscope (micro-Raman spectroscopy). By doing so, a spatial resolution down to the single-cell level becomes possible.7,19 Thus a micro-Raman setup allows investigations of single cells to identify potentially pathogen bacteria already at the beginning of bacteria reproduction.17 However, localization problems occur, when sampling is performed in native environments. Beside microorganisms or in general biotic particles, abiotic particles like organic polymer particles, inorganic salts, or small dust particles can be found on the sample carrier. The amount of abiotic particles exhibiting the same diameter fraction than microorganisms is expected to be orders of magnitudes higher than those of biotic cells. The difficulty to localize a biotic cell in an abiotic sample environment can be solved by fluorescence staining methods. Adequate fluorescence dyes can selectively highlight bacteria due to DNA, surface proteins, or esterase activity.20,21 However, a further demand in microbiology is the differentiation between living and dead cells. Here, cell-interacting fluorescence multistaining methods provide a suitable solution. The different fluorescent behavior of particles can be chosen as a localization marker to visualize the relevant particles of interest. In the case of disease prevention, the interest in dead cells is minimal since they are no longer hazardous for health and therefore do not have to be identified in detail, which decreases the analysis time. Finally, the identification and differentiation of bacteria is carried out by Raman spectroscopy in combination with chemometrical methods.22-25 Since the laser excitation beam is focused by the microscope objectives to the size range of microorganisms, the spectroscopic information can be recorded from single cells without intense background signals of the sample carrier. In the case of high background signals, a confocal Raman approach can be used.33 The different fingerprint regions of bacteria spectra are characteristic for each species. Fluorescence staining methods and Raman spectroscopy cannot always be combined because these are two competitive processes in which the fluorescence emission can mask the much weaker Raman scattering process. However, a suitable combination is possible, if the laser excitation wavelength is chosen to be outside the fluorophore absorption band.26 In this paper, we demonstrate the application of a live/ dead staining kit to highlight living as well as dead cells in an (19) Schuster, K. C.; Reese, I.; Urlaub, E.; Gapes, J. R.; Lendl, B. Anal. Chem. 2000, 72, 5529–5534. (20) Ziegler, G. B.; Ziegler, E.; Witzenhausen, R. ZBL Bakteriol. 1975, 230, 252–264. (21) Shapiro, H. M. J. Microbiol. Methods 2000, 42, 3–16. (22) Kirschner, C.; Maquelin, K.; Pina, P.; Thi, N. A. N.; Choo-Smith, L. P.; Sockalingum, G. D.; Sandt, C.; Ami, D.; Orsini, F.; Doglia, S. M.; Allouch, P.; Mainfait, M.; Puppels, G. J.; Naumann, D. J. Clinical Microbiol. 2001, 39, 1763–1770. (23) Ro ¨sch, P.; Schmitt, M.; Kiefer, W.; Popp, J. J. Mol. Struct. 2003, 661662, 363–369. (24) Ro ¨sch, P.; Harz, M.; Schmitt, M.; Peschke, K.-D.; Ronneberger, O.; Burkhardt, H.; Motzkus, H.-W.; Lankers, M.; Hofer, S.; Thiele, H.; Popp, J. Appl. Environm. Microbiol. 2005, 71, 1626–1637. (25) Ro ¨sch, P.; Harz, M.; Peschke, K. D.; Ronneberger, O.; Burkhardt, H.; Popp, J. Biopolymers 2006, 82, 312–316. (26) Krause, M.; Radt, B.; Rösch, P.; Popp, J. J. Raman Spectrosc. 2007, 38, 369–372.

abiotic environment. Furthermore, we were able to identify different bacteria species by means of Raman spectroscopy in combination with chemometrical methods, even if cells are stained. This unique approach can be transferred to many other fields, where localization, living and dead differentiation, and identification of the bacterial cells are required. EXPERIMENTAL SECTION Fluorescence Observation. The fluorescence excitation is carried out with a mercury lamp. Different excitation and emission wavelengths are realized by exchanging the filter cubes (Olympus) in the light path (UMWIBA3: blue exaction from 460 to 495 nm and green emission from 510 to 550 nm; UMWIGA3 excitation from 530 to 550 nm and emission from 575 to 625 nm). The CCD camera (CC12, Olympus) is adapted to an inverse microscope (BX 41, Olympus), and the fluorescence images are taken with integration times of 500 ms and digital contrast amplification. Raman Instrumentation. A micro-Raman setup (HR LabRam inverse, Horiba Jobin Yvon) is used for collecting the Raman spectra. The spectrometer has an entrance slit of 100 µm, a focal length of 800 mm, and is equipped with a 300 lines/ mm grating. As excitation wavelength the 532-nm line of a frequency-doubled Nd:YAG laser (Coherent Compass) with a laser power of 3-5 mW incident on the sample is used. The Raman scattered light is detected by a CCD camera operating at 220 K. A Leica PLFluoar objective (100× magnification) focuses the laser light onto the samples (∼1.2 µm focus diameter). An integration time of 30 s is needed to record single bacteria Raman spectra. The Raman measurement is preceded by a photobleaching step of the samples with an irradiation period of 30 s. Sample Preparation. Three different Bacillus species (B. subtilis DSM 10, B. pumilus DSM 354, B. sphaericus DSM 488) and three different Staphylococcus species (S. cohnii DSM 20260, S. warneri DSM 20316, S. epidermidis ATCC 12218) were cultured with the same liquid nutrient medium at the temperature of 32.5 °C. After 2 days of growing, an aliquot of cells is stained with the LIVE/DEAD BacLight Bacterial Viability and Counting Kit (Invitrogen). This kit consists of a mixture of two nucleic acid stains s the green fluorescent SYTO 9 dye and the red fluorescent propidium iodide s for viability differentiation. Depending on the bacterium species the strain-specific DNA concentration differs. For similar staining results, the dye concentration has to be adapted to the DNA concentration. Therefore, the dye concentrations vary in the range of 10-120 nM for SYTO 9 and 10-85 nM for propidium iodide. The incubation time is 20-30 min. Afterward, the sample is centrifuged 3 times at 10000g for 3 min and washed in each case with sterile filtrated water. Five microliters of the suspension were dropped onto a fused-silica surface and dried afterward on air. Data Preprocessing and Multivariate Analysis. The analysis of Raman spectra is performed in two steps: First, in a preprocessing step the second derivatives of Raman spectra are calculated by the Savitzky-Golay algorithm combined with a smoothing step (filter width 9) with subsequent vector normalization. Second, the classification of Raman spectra is carried out by using both a hierarchical cluster analysis (HCA, OPUS IDENT) and a support vector machine (SVM) analysis. Analytical Chemistry, Vol. 80, No. 22, November 15, 2008

8569

Figure 1. Bright-field and fluorescence images of B. subtilis, B. pumilus, B. sphaericus, S. cohnii, S. warneri, and S. epidermidis stained by the live/dead staining kit. The green and red fluorescence images are overlaid to one picture for better comparison of living and dead cells.

Figure 2. Influence of fluorescence dyes on bacterial Raman spectra. Microscopic pictures present a living (position 1) and a dead (position 2) cell of S. warneri in bright-field (A) and fluorescence (B). In panel C, the corresponding Raman spectra of the marked cells are depicted.

For HCA two spectral regions (from 3140 to 2750 and 1800 to 960 cm-1) are applied for classification. The spectral distances are calculated by Euclidian distance and Ward’s algorithm is 8570

Analytical Chemistry, Vol. 80, No. 22, November 15, 2008

used for classification. For SVM analysis, the full range of 1223 data points per spectrum is used for calculation. The main parameters for the calculations are the radial basis function

Figure 3. Photobleaching effect of stained B. pumilus cells. (A) Microscopic bright-field image, (B) stained bacteria before Raman measurements, and (C) bacteria after the Raman measurements. For better comparison, the measurement positions were marked with cycles.

kernel algorithm, the SVM parameter cost value of 1.1, and the polynominal kernel parameter γ value of 8.5 × 10-4. RESULTS AND DISCUSSION Microscopic Differentiation of Living and Dead Bacteria Cells. The localization of the cells by means of fluorescence staining can be realized by different cell-dye interaction mechanisms depending on the dye’s nature. Three interaction mechanisms are common: Either the dye complexes with DNA, marks surface proteins when bound to an adequate antibody, or becomes fluorescent when is decomposed by enzymes. The staining kit applied within this study to differentiate between living and dead cells is based on DNA coordinating dyes and their different ability to overcome the positively charged cell potential of intact cells. The live/dead staining kit contains the green fluorescent dye SYTO 9 and the red fluorescent dye propidium iodide (PI). Their distinct different emission maxima enable the differentiation. Furthermore, the positively charged fluorescent label PI can only stain dead bacterial cells since it cannot pass through the positive charged cell membranes of living cells. In contrast to the PI molecules, the neutral SYTO 9 molecules can pass through the cell membrane of both living and dead cells. However, in the case of labeling dead cells with a mix of SYTO 9 and PI, the green fluorescence caused by SYTO 9 will be suppressed by a FRET effect and dead cells fluoresce in the red. It is well-known that the combination of SYTO 9 and PI is well suited for staining eukaryotic cells.27,28 However, these labels can also be applied to stain the prokaryotic bacteria species as described in this contribution. In Figure 1, the staining results for all analyzed species are presented. On the left hand the bright-field images and on the right hand the corresponding fluorescence images are shown. They were generated separately by using different fluorescence filters and combined afterward to visualize living and dead cells in one (27) Haugland, R. P. Handbook of Fluorescent Probes and Research Chemicals, 10 ed.; Molecular Probes, 2005. (28) Berney, M.; Hammes, F.; Bosshard, F.; Weilenmann, H.-U.; Egli, T. Appl. Environ. Microbiol. 2007, 73, 3283–3290.

image. Figure 1 demonstrates the great capabilities of fluorescence staining since the rod-shaped Bacilli species as well as the spherical Staphylococci species are easier visualized in the fluorescence image compared to the bright-field image. Furthermore, the variable ratios of dead cells are visible in Figure 1 and help to focus only on the particles of interest. Spectroscopic Differentiation of Living and Dead Bacteria Cells. We have recently shown that the application of a 532nm laser excitation line yields Raman spectra with an appropriate S/N ratio for Raman identification of nonstained single bacteria.29,30 The combination of fluorescence staining with Raman spectroscopy is not applicable in every case. Since fluorescence emission is up to 1010-fold more intensive than Raman scattering, only a small amount of fluorophores is needed to mask a Raman spectrum.31 The dyes used to mark living and dead cells affect the bacteria Raman spectra differently. Figure 2 shows the microscopic bright-field (A) and fluorescence (B) pictures of a green and a red fluorescing S. warneri cell as well as the corresponding Raman spectra in panel C. It is obvious that PI used for dead staining affects a Raman spectrum drastically while the living cell marker SYTO 9 causes almost no fluorescence background after a photobleaching step. The reason is that the excitation wavelength of 532 nm is at the edge of the SYTO 9 absorption profile. On the other hand, PI exhibits an absorption maximum at 528 nm leading to fluorescence emission masking the Raman information.26 This leads to the fact that Raman spectra of dead marked cells are overlaid by fluorescence and meaningful spectra cannot be registered. In consequence, unknown dead bacteria cannot be identified by using the 532-nm excitation laser line. However, in many cases, dead cells are not of interest to further investigations because they cannot replicate anymore and (29) Ro ¨sch, P.; Harz, M.; Schmitt, M.; Popp, J. J. Raman Spectrosc. 2005, 36, 377–379. (30) Gaus, K.; Ro ¨sch, P.; Petry, R.; Peschke, K. D.; Ronneberger, O.; Burkhardt, H.; Baumann, K.; Popp, J. Biopolymers 2006, 82, 286–290. (31) Schrader, B. Infrared and Raman Spectroscopy, Methods and Applications; VCH: Weinheim, 1995.

Analytical Chemistry, Vol. 80, No. 22, November 15, 2008

8571

Figure 4. Sample mix containing abiotic particles (titanium dioxide, quartz powder, and PMMA) and biotic particles (S. warneri) stained with the live/dead staining kit. (A) Bright-field image; the analyzed particles are marked by a cycle. (B) Fluorescence image of the wavelengths 510-550 and 575-625 nm. The same sample positions are posted. (C) Raman spectra 1-5 corresponding to the listing in the bright-field and fluorescence images.

therefore do not cause undesired effects. In this case, the live/ dead staining kit is an appropriate tool to highlight viable bacteria cells for further Raman measurements. If Raman investigations of dead cells are of interest, measurements will be possible by changing the laser excitation wavelength or using dead staining dyes with different absorption profiles and afterward creating a separate data set for further identification procedures. Furthermore, the reason for the suitability of SYTO 9 to Raman spectroscopic investigations is its photobleaching behavior when excited by the laser. Figure 3 represents one example of the photobleaching effect, where image A displays the microscopic bright-field of stained B. pumilus cells, while image B presents 8572

Analytical Chemistry, Vol. 80, No. 22, November 15, 2008

the stained bacteria before, and image C the bacteria after the Raman measurements. For better comparison, the measured cells are marked with cycles. Because the excitation wavelength is on the verge of the dye excitation profile, a weak fluorescence background is detected. Furthermore, the background varies depending on the DNA concentration and the dye accumulation inside the cell. The photobleaching process minimizes the influence of the background to the identification procedure. Thirty seconds of a laser radiation is sufficient to measure afterward Raman spectra with similar spectral background. Differentiation of Biotic and Abiotic Particles. A major advantage of the active staining method is the highlighting of particles of interest against other particles that are in the same

Figure 5. Analysis of a data set containing 102 Raman spectra of six bacteria species. (A) Representative average Raman spectra for each species out of the data set: (a) B. subtilis, (b) B. sphaericus, (c) B. pumilus, (d) S. epidermidis, (e) S. cohnii, and (f) S. warneri. (B) Dendrogram of the HCA classification using second derivative, Euclidian distance, and Ward’s algorithm. Table 1. SVM Classification Results by the Leave-One-Out Method of a Cross-Validation Calculation for B. subtilis, B. pumilus, B. sphaericus, S. cohnii, S. warneri, and S. epidermidis data set

classification results

sample

number of spectra

B. subtilis B. pumilus B. sphaericus S. epidermidis S. cohnii S. warneri

17 16 17 19 15 18

B. subtilis

B. pumilus

B. sphaericus

S. epidermidis

estimation/%

S. cohnii

S. warneri

17 11

5 17

diameter range of ∼1-3 µm but not in the focus of analysis. So a localization procedure, for example, of living cells based on image processing reduces the measurement time drastically. As shown in the prior sections, the bacteria can be stained with different DNA coordinating dyes, living cells can be highlighted by green fluorescence, and Raman spectroscopy with 532-nm excitation of these cells is possible. At the same time, particles containing no DNA should not be visible by using these dyes. To prove this assumption, S. warneri cells were mixed with abiotic particles (titanium dioxide, quartz powder, poly(methyl methacrylate) (PMMA)) of similar diameter size. The mixed sample was stained with the live/dead staining kit. The results of bright-field and fluorescence microscopy (A, B) as well as

19 15 2

16

correct

error type I

100 68.8 100 100 100 88.9 93.1

0 31.2 0 0 0 11.1 6.9

the Raman spectroscopy analysis (C) are shown in Figure 4. The bright-field image shows ∼30 single particles with similar shape, whereas the fluorescence image delivers detailed information of the biotic particle localization. Only three particles of interest are present: two green-stained living cells and one red-stained dead cell. Exemplary one weak red fluorescent (1), the two strong green fluorescent (3, 4), and two nonfluorescent (2, 5) particles are analyzed. Particle 1 is identified as titanium dioxide. The two nonfluorescent particles correspond to quartz powder (2) and PMMA (5). The two green fluorescent particles are identified as bacteria cells. The diffuse fluorescence of titanium dioxide is likely the result of surface adsorption effects and is under further investigation. However, Analytical Chemistry, Vol. 80, No. 22, November 15, 2008

8573

Table 2. SVM Classification Results by the Test Set Methoda classification results

data set

sample

Number of spectra

B. subtilis B. pumilus B. sphaericus S. epidermidis S. cohnii S. warneri

7 6 7 8 5 9

a

B. subtilis

B. pumilus

B. sphaericus

estimation/%

S. epidermidis

S. cohnii

S. warneri

7 4

2 7 8 5 2

7

correct

error type I

100 66.7 100 100 100 77.8 90.5

0 33.3 0 0 0 22.2 9.5

The preprocessed data are divided randomly into a data set for calibration containing 60 spectra and a data set for validation of 42 spectra.

the green fluorescence restricts highlighting of biological cells and can be applied as a marker for the particles of interest. The resulting reduction of analysis time can be estimated in this figure, too. Without fluorescence staining, all 30 particles visible in the bright-field image would have to be analyzed to find the two bacteria cells of interest. Applying fluorescence staining, 28 particles are excluded for a Raman analysis in this sample. So only the two green-stained particles have to be measured, and the analysis can be accelerated in this example to a 15-fold faster measurement time. Classification and Identification of Stained Bacteria Species. Beside the advantage of reducing the measurement time, the final aim of this analytical approach is the identification of unknown bacterial species by means of Raman spectroscopy. Therefore, we analyzed the Raman spectra of 102 spectra resulting from six stained bacterial species. In principle, the identification can be realized by comparison of measured Raman spectra with a data set including defined spectra of various bacterial species. A profound data set has to contain as many different variables as possible such as the cell age, the available nutrients, and the growing temperature to ensure a successful prediction.32 Although the influence of the dye SYTO 9 after photobleaching is marginal to Raman spectra excited at 532 nm, a new variable occurs and so a new data set is required. The 102 stained bacteria Raman spectra were collected to form such a data set. To get an impression of the similarity of bacteria Raman spectra, the average spectrum for each species of the data set is shown in Figure 5A. The spectra are shifted vertically for better visualization. The Raman spectra exhibit similar signal profiles concerning intensive signals around 2900 cm-1 due to CH stretching vibrations, a fingerprint region between 1100 and 1700 cm-1, and a fused-silica signal around 1000-1100 cm-1 resulting from the substrate. For an objective spectra comparison, the spectral differences revealed from the background are eliminated by calculating secondderivative spectra. These transformations are the basis for a further analysis. Due to the fact that small and suitable differences in the fingerprint region are not clearly visible by eye, chemometrical methods can be used for a graphic visualization. By applying the HCA, the similarity of spectra without any further spectral (32) Harz, M.; Ro ¨sch, P.; Peschke, K. D.; Ronneberger, O.; Burkhardt, H.; Popp, J. Analyst 2005, 130, 1543–1550. (33) Baia, L.; Gigant, K.; Posset, U.; Schottner, G.; Kiefer, W.; Popp, J. Appl. Spectrosc. 2002, 56, 536–540.

8574

Analytical Chemistry, Vol. 80, No. 22, November 15, 2008

specifications are calculated (unsupervised method). The spectral distances between all spectra in the data set are calculated by the Euclidian distance formula. According to the spectral similarities, the derived spectra are classified into clusters by the Ward algorithm to form the most homogeneous clusters. The result of a HCA analysis is presented in a dendrogram (Figure 5B) by plotting the spectra names against the heterogeneity to each other. The illustration shows first a classification of two large clusters. These clusters are assigned to the genus Bacillus (a-c) and the genus Staphylococcus (d-f). Second, an optimal classification for each species is visible. Eight spectra are misclassified on a species level resulting in a recovery rate of 92%. On the genus level, only three spectra are not classified correctly. Furthermore, the spectra can be analyzed by a second chemometrical method to support the results of HCA. The SVM analysis is a supervised method and needs preliminary information concerning the membership of spectra to the bacterial species. In Table 1, the results of a cross-validation (leave-oneout-method) is shown. The recovery rate can be enhanced to a mean value of 93% with this method. Misclassifications only occur on species level and can rise up to an error of ∼30%, since B. pumilus and B. sphaericus are spectroscopic too similar. On a genus level, Bacillus and Staphylococcus species can be separated clearly. Furthermore, the data set undergoes an independent test to verify the identification performance. In this contribution, the test set method is applied. The data are divided randomly into a calibration data set of 60 spectra and a validation data set of 42 spectra. The first mentioned data are used to train a model. This model is the basis for the classification of the independent validation data. In Table 2, the results are presented. A recovery rate mean value of 90% is possible by using the SVM algorithm. Like for the other methods, all spectra are classified correctly on genus level. These calculations demonstrate that it is possible to differentiate even stained bacterial species by means of Raman spectroscopy in combination with the chemometrical methods HCA and SVM with accuracy higher than 90%. CONCLUSION AND OUTLOOK With this contribution, we present a proof-of-principle experiment pointing out that a Raman spectroscopic analysis is capable of identifying different bacterial species stained with

SYTO 9 and PI. The presence of these DNA coordinating fluorescence dyes allows the differentiation of biotic and abiotic particles. This feature is necessary to reduce the analysis time significantly when applying the Raman analysis method selectively to the detection of microorganism cells. Furthermore, multistaining methods offer the advantage of focusing the Raman analysis only on living cells. This task is performed by the dye SYTO 9, which can be used in combination with the laser excitation wavelength of 532 nm. Finally, it could be shown by the chemometric methods HCA and SVM that it is possible to classify all species and to identify unknown bacteria spectra with accuracy higher than 90%. In the future, this prelocalization method can be one step of an automated analysis process to identify microorganisms. For best recognition results, the influence of different growing conditions has to be taken into account as well as the age of stained cells. Therefore, the stained bacteria data set has to

be extended to these parameters as well as to other bacterial species of interest. In combination with a suitable sample preparation, the evaluation of native samples becomes possible and provides a fast microorganism analysis method in a large application range as in in food analysis, air monitoring, or blood diagnosis. ACKNOWLEDGMENT We highly acknowledge the Carl Zeiss fund, the Deutsche Forschungsgemeinschaft (Graduiertenkolleg 1257, DE 307/7-1 and PO 563/7-1), and the funding of the research project FKZ 13N9549 from the Federal Ministry of Education and Research, Germany (BMBF) for financial support. Received for review July 14, 2008. Accepted September 10, 2008. AC8014559

Analytical Chemistry, Vol. 80, No. 22, November 15, 2008

8575