Multiarray Sensors with Pattern Recognition for the Detection

Nov 5, 2005 - Multiarray Sensors with Pattern Recognition for the Detection, Classification, and Differentiation of Bacteria at Subspecies and Strain ...
2 downloads 0 Views 309KB Size
Download PDF
Anal. Chem. 2005, 77, 7941-7949

Multiarray Sensors with Pattern Recognition for the Detection, Classification, and Differentiation of Bacteria at Subspecies and Strain Levels Jason Karasinski, Silvana Andreescu,† and Omowunmi A. Sadik*

Department of Chemistry, State University of New YorksBinghamton, P.O. Box 6000, Binghamton, New York 13902 Barry Lavine and Mehul N. Vora

Department of Chemistry, Oklahoma State University, Stillwater, Oklahoma 74078-3071

This work describes the integration of a fully autonomous electrochemical biosensor with pattern recognition techniques for the detection and classification of bacteria at subspecies and strain level. The system provides a continuous, real-time monitoring of bacteria activity upon exposure to antibiotics. The system utilizes 96-well-type electrodes array (DOX-dissolved oxygen sensor) with principal component analysis (PCA) for rapid and routine classification of different classes of bacteria and related strains. A representative sample of a section of the bacteria kingdom has been analyzed and classified using the proposed DOX-PCA system, including the following: Corynebacterium glutamicum, Microcuccus luteus, Staphylococcus epidermidis, Yersinia ruckeri, Escherichia adecarboxylata, Comamonas acidovorans, Alcaligenes odorans, Bacillus globigii, and three strains of Escherichia coli (K12, SM10, ATCC 25922). The new classification scheme is based on the hypothesis that, under identical experimental conditions, various bacteria consume oxygen at different rates and are affected in different ways by selected antibiotics. Thus, the response of the individual electrode in the array is indirectly altered, compared to that of cells growing on medium, by the addition of the antibiotic. By using three different antibiotics in separate wells, a unique fingerprint can be created for a specific bacterium. With the proposed DOX-PCA system, classification of bacteria was achieved at subspecies and strain level in real time. This study represents a basic research tool that may allow researchers to rapidly detect, quantify, and classify bacteria type at subspecies and strain levels. Correct identification and differentiation of bacterial contaminants is an important aspect in many practical applications ranging from medical diagnosis and environmental studies to the agricultural and food industry and homeland security. The availability of a reliable and fast method for early and proper identification of * Corresponding author. Fax: (607) 777-4478. E-mail: [email protected]. † Current address: Department of Chemistry, Clarkson University, Potsdam, NY 13699-5810. 10.1021/ac0512150 CCC: $30.25 Published on Web 11/05/2005

© 2005 American Chemical Society

these contaminants is required due to their rapid growth and toxic effects.1 The ideal method should be able to (i) distinguish individual bacteria from a wide spectrum of toxic agents belonging to different classes, (ii) provide high sensitivity and reproducibility, and (iii) be simple to use and portable. Although current technologies (morphological identification, simple culture/counting, enzyme-linked immunosorbent methods) used for bacteria monitoring allow sensitive detection,2 they are not able to perfectly differentiate and classify bacteria at strains or even at species level. A general approach to developing a reliable method for simple visual classification and identification of bacteria as a member of a bacterial class is to use pattern recognition techniques.3-5 Typically, in a classification study, the characteristics of unknown samples must be compared and matched against a large database stored in a library containing complete and known classification information of well-characterized bacteria from different species and stains. If such a library is available, one bacteria species can be distinguished from other related ones by matching them against the database. In this case, the classification strategy may involve the use of a detection method that will provide chemical information for creating a distinct pattern for each component belonging to the class of compounds under consideration. These data are then transferred to a data processing and reduction system that will extract the relevant information and create a systematic classification of compounds possessing similar characteristics. Such an example was recently described by Lopez-Diez and Goodacre, who used UV resonance Raman spectroscopy combined with chemometric analysis for obtaining reproducible Raman fingerprints of 27 strains of endospore-forming Baccillus and Brevibacillus, which (1) Sadik, O. A.; Wanekaya, A. K.; Andreescu, S. J. Environ. Monit. 2004, 6, 413-522. (b) Andreescu, S.; Karasinski, J.; Sadik, O. A. Encyclopedia of Sensors; American Scientific Publishers. In press. (2) Straitis-Cullum, D. N.; Griffin, G. D.; Mobley, J.; Vass, A. A.; Vo-Dinh, T. Anal. Chem. 2003, 75, 275-280. (b) Vo-Dinh, T.; Alarie, J. P.; Isola, N.; Landis, D.; Wintenberg, A. L.; Ericson, M. N. Anal. Chem. 1999, 71, 358363. (3) Albert, K. J.; Lewis, N. S.; Schauer, C. L.; Sotzing, G. A.; Stitzel, S. E.; Vaid, T. P.; Walt, D. R. Chem. Rev. 2000, 100, 2595-2626. (b) Albert, K. J.; Walt, D. R. Anal. Chem. 2003, 75, 4161-4167. (4) Jurs, P. C.; Bakken, G. A.; McClelland, H. E. Chem. Rev. 2000, 100, 26492678. (5) Lavine, B.; Workman, J. J. Anal. Chem. 2004, 76, 3365-3372.

Analytical Chemistry, Vol. 77, No. 24, December 15, 2005 7941

were subsequently used to discriminate between these microorganisms at species level.6 Chemical and biological sensors could be used with computational techniques for bacteria monitoring.2,7,8 A typical example of sensors combined with computational techniques used for monitoring microbes is the electronic nose technology.3,9-12 Following an approach similar to that of the electronic nose, the main objective of this work is to produce unique microbial fingerprints as a reliable and rapid method for identifying bacteria to the species level. Our approach is to use a unique combination of an electrochemical sensor with pattern recognition techniques that will subtract, process, and classify the relevant analytical information provided by the sensor. For this specific objective, we propose the use of a high-throughput 96-electrode dissolved oxygen sensor (DOX)13-15 for measuring the difference in the oxygen consumed by different bacteria classes and strains over time. The DOX is connected to a data processing and classification module via a specifically designed principal component analysis (PCA) algorithm (Figure 1). The DOX sensor allows determination of dissolved oxygen simultaneously on 96 channels and was used previously for studying A549 lung cancerous cells and their interaction with analytes15 and for monitoring bacteria.13,14 To our knowledge, the present work is the first report utilizing the combination of an electrochemical oxygen multisensor with pattern recognition techniques for detection, classification, and differentiation of bacterial pathogens and related strains. Sensor Concept and Integration with Chemometrics. The new classification scheme described in this work is based on the hypothesis that, under identical experimental conditions, various bacteria (i) consume oxygen at different rates and (ii) are affected in different ways by selected antibiotics. Therefore, two experimental strategies are employed for bacteria differentiation with the DOX-PCA system: (i) direct detection via oxygen consumption and (ii) evaluation of the effect of antibiotics on growth of bacterial pathogen. The first one is the most straightforward and consists of monitoring the rate of bacteria respiration, quantified (6) Lopez-Diez, E. C.; Goodacre, R. Anal. Chem. 2004, 76, 585-591. (b) Jarvis, R. M.; Brooker, A.; Goodacre, R. Anal. Chem. 2004, 76, 5198-5202. (7) Ertl, P.; Mikkelsen, S. R. Anal. Chem. 2001, 73, 4241-4248. (b) Ertl, P.; Wagner, M.; Corton, E.; Mikkelsen, S. R. Biosens. Bioelectron. 2003, 18, 907-916. (8) Chuang, H.; MAcuch, P.; Tobacco, M. B. Anal. Chem. 2004, 73, 462-466. (b) Wu, C. F.; Valdes, J. J.; Bentley, W. E.; Sekowski, J. W. Biosens. Bioelectron. 2003, 19, 1-8. (9) Ampuero, S.; Bosset, J. O. Sens. Actuators, B 2003, 94, 1-12. (b) Soderstrom, C.; Winquist, F.; Krantz-Rulcker, C. Sens. Actuators, B 2003, 89, 248-255. (10) Magan, N.; Pavlou, A.; Chrysanthakis Sens. Actuators, B 2001, 72, 28-34. (b) Gardner, J. W.; Boilot, P.; Hines, E. L. Sens. Actuators, B 2005, 106, 114-121. (c) Dutta, R.; Morgan, D.; Baker, N.; Gardner, J. W.; Hines, E. L. Sens. Actuators B. In press. (11) Gibson, T. D.; Prosser, O.; Hulbert, J. N.; Marshall, R. W.; Corcoran, P.; Lowery, P.; Ruck-Keene, E. A.; Heron, S. Sens. Actuators, B 1997, 44, 413422. (b) McEntegart, Penrose, W. R.; Strathmann, S.; Stetter, J. R. Sens. Actuators, B 2000, 70, 170-176. (12) Masila, M.; Sadik, O. A. In Chemical and Biological Sensors for Environmental Monitoring; Mulchandani, A., Sadik, O. A., Eds.; ACS Symposium Series 762; Oxford University Press: Washington, DC. 2000; pp 37-59. (13) Amano, Y.; Okomura, C.; Yoshida, M.; Katayama, H.; Unten, S.; Arai, J.; Tagawa, T.; Hoshina, S.; Hashimoto, H.; Ishikawa, H. Hum. Cell 1998, 12 (1), 3-. (14) Kitahara, T.; Koyama, N.; Matsuda, J.; Hirakata, Y.; Kohno, S.; Nakashima, M.; Sasaki, H. Biol. Pharm. Bull. 2003, 26 (9), 1229-1234. (15) Andreescu, S.; Sadik, O. A.; McGee, D. W. Anal. Chem. 2004, 76, 23212330.

7942 Analytical Chemistry, Vol. 77, No. 24, December 15, 2005

Figure 1. Schematic representation of the DOX-PCA concept.

as a measure of oxygen consumed by the cells over time. Monitoring bacteria respiration using oxygen sensors has been already described in several works.7,18 When the bacteria grow, they consume oxygen at different rates, and this can be readily quantified by the DOX electrochemical sensor. The second strategy is based on monitoring changes in cell respiration as a result of the toxic effects induced by the presence of different antibiotics. When specific concentrations of antibiotics are added to a medium containing a microbial culture in the growth phase, a variety of effects can be monitored with the DOX system. Antibiotics may be lethal to the cells, completely stopping cell respiration or division, they can reduce the rate of cell division, or they can have no effect at all. These effects can be seen by comparing the oxygen consumption curves of bacteria with and without an antibiotic present. Integration of Chemometrics with a DOX Multisensor. Recent advances in biosensor technology have been reported to increase the sample throughput in order to respond to current requirements for multidetection and multianalysis. In the present work, the DOX oxygen multisensor generates large amounts of data on 96 channels simultaneously in a very short time, which are difficult to process and interpret without proper data reduction and analysis systems. Therefore, appropriate computational techniques and data reduction software are necessary to reduce the dimensionality of the data set by extracting relevant multivariate structure from chemical information and finally to interpret, process, and analyze the results.5,17 These methods in conjunction with multiarray sensors can provide qualitative or quantitative analytical information or both in a comprehensive graphical representation. Some attempts have been made to incorporate chemometric analysis in a biosensor configuration with application to bacterial differentiation and classification.7,18 In the present work, we employ a dissolved oxygen multisensor connected to (16) O’Riodan, T. C.; Buckley, D.; Ogurtsov, V.; O’Connor, R.; Papkovsky, D. B. Anal. Biochem. 2000, 278, 221-227. (17) Sadik, O. A.; land, W. H.; Wanekaya, A. K.; Uematsu, M.; Embrechts, M. J.; Wong, L.; Leibensperger, D.; Volykin, A. J. Chem. Inf. Comput. Sci. 2004, 44, 944-507. (18) Rowe, C. A.; Tender, L. M.; Feldstein, M. J.; Golden, J. P.; Scruggs, S. B.; MacCraith, B. D.; Cras, J. J.; Ligler, F. S. Anal. Chem. 1999, 71, 38463852. (b) Taitt, C. R.; Anderson, G. P.; Lingerfelt, B. M.; Feldstein, Ligler, F. S. Anal. Chem. 2002, 74, 6114-6120. (c) Delehanty, J. B.; Lingler, F. S. Anal. Chem. 2002, 74, 5681-5687.

an appropriate computational method (PCA) that can rapidly search and identify a microbial species in a database of wellcharacterized bacteria. The optimized DOX-PCA method could (i) significantly improve the overall performance of the electrochemical sensor by providing a better and more useful interpretation and understanding of the process being investigated and (ii) speed up the investigation process. Similar to an electronic nose instrument, the system does not necessarily need to be designed for particular bacteria, but to learn specific patterns of well-known bacteria that can then be associated with unknown samples for further identification and recognition (Figure 1). The dimensionality of the DOX data makes this system ideal for the formulation of chemometric methods since a distinct pattern (or fingerprint) is generated, which can be used for further classification and differentiation of bacteria pathogens in a given sample. EXPERIMENTAL SECTION DOX 96-Multisensor. The DOX-96 multiarray sensor is an automated multipotentiostat prototype produced by Daikin Industries, Ltd. that is designed to measure cells’ respiratory activity via the consumption of dissolved oxygen in bacteria cultures. The principle of operation and a schematic representation of the DOX sensor were described in detail in a previous work.15 Basically the DOX device functions as a classical electrochemical oxygen sensor. The uniqueness of this system is that it operates via a multichannel system that enables simultaneous measurement of 96 samples. The device is composed of a multipotentiostat and a 96-electrode array with 96 × 3 metallic pins that ensure the electrical contact with the upper part of the electrodes. Each array contains 96 interconnected sensors, each consisting of three disposable electrodes, working, reference, and auxiliary with the following dimensions: 0.19 mm2 × 5 mm2 × 5 mm2. The protocol for measuring bacterial populations involves the use of a conventional 96-well plate containing 200 µL of growth medium with or without bacteria in the presence or absence of antibiotics (chloramphenicol, ampicillin, tetracycline). Then, 96 × 3 goldplated electrodes (Daikin Corp.) were fitted onto the top of the 96-well plates, and the oxygen reduction current was monitored at a fixed potential (-700 mV versus gold), where the oxygen reduction current is maximum. This potential ensures the highest sensitivity of the system and should provide a better differentiation and classification of bacteria. The electrodes were used as disposable devices, and each plate was used for a single measurement. Experiments with the DOX system for pattern recognition experiments were carried out at ambient temperature, and all the electrochemical measurements were performed at -700 mV versus gold-plated pseudoreference electrode. Materials. The following bacteria were purchased from American Type Culture Collection (ATCC; Manassas, VA): Corynebacterium glutamicum (ATCC 13032), Microcuccus luteus (ATCC 10240), Staphylococcus epidermidis (ATCC 12228), Yersinia ruckeri (ATCC 29473), Escherichia adecarboxylata (ATCC 23216), Comamonas acidovorans (ATCC 51340), Alcaligenes odorans (ATCC 33585), Escherichia coli (ATCC 25922), and Bacillus globigii (ATCC 9372). E. coli K12 and SM10 were donated from the Biological Sciences Department at SUNYsBinghamton. Chloramphenicol, ampicillin, and tetracycline were purchased from Sigma (St. Louis, MO). The bacteria were cultured on agar plates and incubated for 18-24 h at 37 °C. Then, the cells were harvested and

suspended in 10 mL of Mueller-Hinton broth (Sigma). The concentration of the fresh bacteria solutions were checked by measuring the absorbance at 600 nm using a Hewlett-Packard diode array spectrophotometer model HP-8553. The corresponding number of bacteria was calculated using a calibration curve obtained with standard solutions each containing a known number of bacteria. This number was estimated using the conventional “colony forming units” (cfu) method.19 Finally, different concentrations of bacterial solutions were prepared and placed in a 96-well plate. The same experimental protocol was used for all bacterial pathogens and antibiotics selected for this study. Appropriate safety measures were used when handling bacteria preparation. To avoid contamination, all glassware, equipment, pipets, pipets tips, and the benchtop were disinfected with 80% ethanol before and after performing the experiments. Preparation of Plates for Pattern Recognition Experiments. In preparing the plates for pattern recognition experiments, serial dilutions of bacterial pathogens were prepared from a stock solution of bacteria and 200 µL was placed in the 96-well plate. After the insertion of the electrodes in the wells, the plate was placed in the measurement chamber of the DOX sensor. The level of oxygen was measured continuously for more than 8 h by applying a potential of -700 mV versus a gold pseudoreference electrode, without any additional pretreatment or incubation step.15 For experiments involving the study of the effect of antibiotics, stock 1 mg/mL solutions of each antibiotic were prepared in Nanopure water, obtained using a Barnstead water purification system (model D4641) with a minimum resistivity of 17.5 MΩ/ cm2. Tetracycline was diluted to 1 µg/mL in Mueller-Hinton broth. Ampicillin and chloramphenicol were diluted to 5 µg/mL. Then, 100 µL of each diluted antibiotic was added to the appropriate wells of a 96-well tissue culture plate. Additionally, 100-µL aliquots of the broth, in the absence of antibiotics, were added to the positive control wells of the plate. Negative control wells of the plate were loaded with 200 µL of only broth. For each microbe tested, four wells of the plate were prepared with each of the above solutions; thus, each experiment was simultaneously repeated four times. The plates were stored at -20 °C until used. Overnight cultures of each bacterium were prepared in Mueller-Hinton broth at 37 °C. Immediately before each experiment, each culture was diluted to an absorbance at 600 nm of 0.2 in nutrient broth. A 100-µL sample of each dilution was then added to the appropriate wells of the prepared antibiotic plate. The final concentration of the cells in the plate was A600 ) 0.1, 2.5 µg/mL ampicillin, 0.5 µg/mL tetracycline, and 2.5 µg/mL chloramphenicol. The DOX system was allowed to equilibrate for at least 1 h before each experiment so that the temperature inside the chamber essentially remained constant. The current at -700 mV was recorded every minute for 800 min. Antibiotic Susceptibility Using the Tube Dilution Method. Each bacterium was tested for its susceptibility toward tetracycline, chloramphenicol, and ampicillin using the standard tube dilution method.19 Five milliliters of tetracycline was serial diluted in Mueller-Hinton broth, 2-fold, from 8 to 0.5 µg/mL in 15-mL sterile centrifuge tubes. Chloramphenicol and ampicillin were diluted, 2-fold, from 40 to 2.5 µg/mL. Fresh overnight cultures of (19) Prescott, L. M.; Harley, J. P.; Klein, D. A. Microbiology, 4th ed.; Boston: WCB/McGraw-Hill: Boston, MA, 1999,

Analytical Chemistry, Vol. 77, No. 24, December 15, 2005

7943

Figure 2. Taxonomic tree illustrating bacteria selected for this study. Under the bacteria domain, three kingdoms were selected that are fundamentally different: Gram - proteobacteria, high G+C (guanine content by percentage in DNA) Gram +, and the low G+C Gram +. Under the kingdom, the tree further branches into section, genus, and species. For E. coli, the tree further branches into species.

each cell were diluted to an absorbance of 0.2 at 600 nm in Mueller-Hinton broth. Five milliliters of the cell suspension was added to each antibiotic-prepped tube. Tubes were incubated overnight at 37 °C. UV-Visible spectrometry. UV-visible experiments were performed on a Hewlett-Packard 8533 diode array instrument. Overnight cultures of E. coli K12 and B. globigii were diluted to a starting absorbance of 0.2 at 600 nm in Mueller-Hinton broth. Cells were diluted 2-fold into 50-mL sterile centrifuge tubes containing either broth only or broth solutions containing 1.0 µg/ mL tetracycline, 5.0 µg/mL chloramphenicol, or 5.0 µg/mL ampicillin. The solutions were incubated at room temperature to simulate the temperature inside the DOX. The absorbance at 600 nm of each cell solution was measured every 30 min for 300 min by pipetting 1.0 mL of the solution into a 1.0-mL cuvette. Data Acquisition and Processing. Each bacteria sample was represented as a data vector, x ) (x1, x2, x3, ... xj, ... x480) with the components of the vector denoting the electrode current measured; e.g., x1 is the electrode current measured at 1 min and x480 is the electrode current measured at 480 min. For pattern recognition analysis, the data were autoscaled so that each variable has a mean of zero and a variance of one. Autoscaling removes any inadvertent weighting of the data that would otherwise arise due to differences in magnitude among the measurements. After autoscaling, all of the measurements have equal weight and therefore an equal effect in the analysis. RESULTS AND DISCUSSION The main objective of this work is to evaluate whether the DOX multielectrode sensor combined with pattern recognition techniques (PCA) could be used to identify and differentiate different species and strains of bacteria. Recognition of different bacteria with the proposed DOX-PCA system could find numerous applications in various fields such as environmental and food quality control, homeland security, and clinical analysis. Bacteria and Antibiotics Selection. For this study, 11 microbes were chosen that are representative samples of a small section of the bacteria domain. A partial taxonomic tree of the bacteria selected for this study is shown in Figure 2. Each type 7944 Analytical Chemistry, Vol. 77, No. 24, December 15, 2005

of cell selected is an aerobic or facultative anaerobic microorganism that grows optimally in nutrient broth at 37 °C. The samples were selected to ensure that fundamental differences between individual bacteria ranged from very similar (e.g., three different strains of E. coli) to very different (e.g., the Gram-negative, nonsporulating E. coli compared to the Gram-positive, sporeforming B. globigii). Three commonly used antibiotics were selected: tetracycline, ampicillin, and chloramphenicol. These broad-spectrum antibiotics are known to be effective on a wide variety of Gram-positive and Gram-negative bacteria.19 Both tetracycline and chloramphenicol have a static primary effect on their target, meaning that the drug reversibly inhibits growth. The mechanism of both drugs is the inhibition of protein synthesis by binding either the 30S (tetracycline) or 50S (chloramphenicol) ribosomal subunit. In contrast, ampicillin is lethal (cidal) to its target. Its mechanism is the inhibition of enzymes involved in the synthesis of the bacterial wall peptidoglycan. By using broad-spectrum antibiotics with different mechanisms of action and different chemical structures, we postulated that differences in the rate of cellular uptake and the minimal inhibitory concentration between the selected bacteria could produce unique responses in the DOX system. Tube Dilution and UV-Visible Experiments. Antibiotic susceptibility and UV-visible experiments were performed as validation experiments for the DOX-antibiotic protocol. The tube dilution method is commonly used in microbiology to determine the minimal inhibitory concentration (MIC). The MIC is the lowest concentration of a drug that prevents growth of a particular organism. In the procedure, cells are grown overnight in broth containing serial dilutions of antibiotics. After the incubation period, tubes are visually inspected for growth. The MIC is the tube containing the concentration of antibiotic that shows no growth after 16-20 h. For the DOX protocol, the objective was to use low concentrations of the antibiotics, falling below the MIC of most bacteria, to induce changes in oxygen consumption without totally inhibiting cell growth. The concentration of the antibiotics used in the experiments was selected using the lowest MICs as determined by the tube dilution method. With the exception of B. globigii,

Figure 3. Comparison of the raw data from the DOX (A) vs absorbance measurements (B).

which showed an MIC for ampicillin of 1.25 µg/mL, E. coli 25922 proved to have the lowest MIC for all three antibiotics. This was expected because this strain of E. coli is commonly used for quality control and standardization of antibiotic susceptibility testing.20 The minimal inhibitory concentrations for E. coli 25922 were 2.5 µg/mL for ampicillin, 0.5 µg/mL for tetracycline, and 2.5 µg/mL for chloramphenicol. UV-visible experiments were used to examine the absorbance of cell suspension growing with or without the presence of antibiotics. These results were compared to the raw data from the DOX. The UV-visible results for E. coli K12 and B. globigii correlate well with the DOX experiments, although it should be noted that the UV-visible experiments monitor an increase in cell density as a result of cell multiplication, whereas the DOX measures oxygen consumption, which can occur without cell replication. Figure 3 shows the UV-visible and DOX raw data for B, globigii. The dark blue line representing cells in medium only shows the rapid replication of B. globigii. The line slopes downward in Figure 3A, indicating rapid oxygen consumption as compared to the upward slope in Figure 3B indicating increased absorbance resulting from cell replication. As noted, the tube dilution method for B. globigii gave an MIC of 1.25 µg/mL. This corresponds to the light blue lines in Figure 3. After ∼100 min, the oxygen consumption in Figure 3A approaches the slope of the medium only, indicating cell respiration has stopped. The light blue ampicillin line in Figure 3B remains near an absorbance of 0.1, indicating no cell multiplication. Monitoring Cell Growth Using the DOX System. The DOX system was applied to monitor the growth of 11 types of bacteria: C. glutamicum, M. luteus, S. epidermidis, Y. ruckeri, E. adecarboxylata, C. acidovorans, A. odorans, E. coli, B. globigii, and E. coli K12 and SM10 strains. (20) Jorgensen, J. H.; et al. Methods for dilution antimicrobial susceptibility tests for bacteria that grow aerobically, 3rd ed.; National Committee for Clinical Laboratory Standards: Wayne, PA, 1999.

Bacteria Differentiation with the DOX-PCA System via Direct Monitoring of Oxygen. When the bacteria grow, the oxygen in the medium is consumed over time as a result of their respiration, and consequently, the measured oxygen reduction current decreases to a threshold value. The time it takes to achieve this threshold current is proportional to the rate of oxygen consumption and with the concentration of bacteria present in each well.14 By comparison, a theoretically constant level of oxygen in a sample containing only the growth medium in the absence of bacteria is observed, and subsequently, the current value for this blank sample never attains the threshold value. Typical DOX responses for a series of four concentrations of three different bacteria pathogens, Bacillus cereus, Alcaligenes faecalis and Bacillus subtilis, are presented in Figure 4. B. cereus is a cause of food poisoning usually found in rice. A. faecalis is widespread in soil and water, is used in pharmaceutical industry to synthesize pipecolic acid, and is capable of anaerobic respiration. B. subtilis is a harmless microbe usually present in soil. A typical DOX response shows a decrease in the current suggesting a reduction in the concentration of dissolved oxygen in a well as a result of cell respiration. As can be seen in Figure 4, a decrease in the current was also observed in the case of control samples with no cells. This suggests an artificial consumption of the oxygen and can be due to a possible electrode fouling or adsorption phenomenon that appears as a consequence of the complexity of the broth medium. To minimize these unwanted events, the electrodes were used for a single measurement, as disposable devices, and all measurements were performed by comparison with blank samples in the absence of cells. The current decreases until it reaches a constant value or until the entire oxygen present in the medium is depleted. DOX responses in the case of the three bacteria tested (Figure 4) show a variation in the oxygen reduction current that is proportional to the amount of bacteria in each well. These results also indicate a different level of oxygen in the three cases, suggesting that the three bacteria consumed oxygen at different Analytical Chemistry, Vol. 77, No. 24, December 15, 2005

7945

Classification of Bacteria. The basic premise underlying the approach to classification taken in this study is that all multivariate analysis methods will work well when the problem is simple. By identifying the appropriate features, a “hard” problem is reduced to a “simple” one. The focus of the data analysis is, therefore, feature selection, to increase the signal-to-noise ratio of the data by discarding measurements that are not characteristic of the source profile of the classes in the data set. To ensure identification of all relevant features, it is best that our approach to feature selection is multivariate and takes into account the existence of redundancies in the data. In this study, a genetic algorithm (GA) for pattern recognition was used to identify features for principal component analysis.22-27 The pattern recognition GA identifies features that optimize the separation of the classes in a plot of the two or three largest principal components of the data. Since principal components maximize variance, the bulk of the information encoded by the selected features is about differences between the classes in the data set. The principal component plot used by the fitness function of the GA acts as an embedded information filter. Sets of features are selected based on their principal component plots, with a good principal component plot generated by features whose variance or information is primarily about differences between the classes. This limits the search to these types of feature subsets, thereby significantly reducing the size of the search space. In addition, the GA can focus on those classes or samples that are difficult to classify by boosting their weights over successive generations using a perceptron to learn the class and sample weights. Samples that are more difficult to classify are assigned larger weights. The pattern recognition GA integrates aspects of artificial intelligence and evolutionary computations to yield a “smart” one-pass procedure for feature selection and classification. The efficacy and efficiency of the pattern recognition GA has been demonstrated in previous studies.28-33 The first step in the study was to apply principal component analysis to the data. The pattern recognition GA was used to uncover features characteristic of bacterial concentration. Sampling key feature subsets, scoring their principal component plots, and tracking samples that were most difficult to classify identified

Figure 4. Typical DOX responses at -700 mV vs Au pseudoreference electrode during 8 h continuous monitoring of four different concentrations of B. cereus (A), A. faecalis (B), and B. subtilis (C). Results are the average of 6 identical electrodes.

rates. This observation could be used to differentiate between varieties of bacterial pathogens based on their oxygen consumption. To further investigate this hypothesis; to enhance the performance of the sensor for molecular recognition, and to obtain a more accurate (quantitative, qualitative, or both) classification of bacteria, we used PCA to process the data generated by the DOX system. PCA has been successfully used as an effective method for processing responses generated with an E-nose system for discriminating between simple and complex odors of two species of Staphylococcus aureus bacteria.10c In the present work, PCA was used to identify features that (i) differentiate between bacteria based on the oxygen consumed by the cells over time and (ii) recognize the different amounts of bacteria in a sample. 7946 Analytical Chemistry, Vol. 77, No. 24, December 15, 2005

(21) Chevalier, J.; Mallea, M.; Pages, J. M. Biochem. J. 2000, 348, 223-227. (22) Lavine, B. K.; Moores, A. J.; Helfend, L. K. J. Anal. Appl. Pyrolysis 1999, 50, 47-62. (23) Lavine, B. K.; Moores, A. J.; Mayfield, H. T.; Faruque, A. Microchem. J. 1999, 61, 69-78. (24) Lavine, B. K.; Ritter, J.; Moores, A. J.; Wilson, M.; Faruque, A.; Mayfield, H. T. Anal. Chem. 2000, 72 (2), 423-431 (25) Lavine, B. K.; Brzozowski, D.; Moores, A. J.; Davidson, C. E.; Mayfield, H. T. Anal. Chim. Acta 2001, 437 (2), 233-246. (26) Lavine, B. K.; Davidson, C. E.; Moores, A. J.; Griffiths, P. R. Appl. Spectrosc. 2001, 55 (8), 960-966. (27) Lavine, B. K.; Vesanen, A.; Brzozowski, D. M.; Mayfield, H. T. Anal. Lett. 2001, 34 (2), 281-294. (28) Lavine, B. K.; Davidson, C. E.; Moores, A. J. Chemom. Intell. Lab. Instrum. 2002, 60 (1), 161-171. (29) Lavine, B. K.; Davidson, C. E.; Moores, A. J. Vib. Spectrosc. 2002, 28 (1), 83-95. (30) Lavine, B. K.; Davidson, C. E.; Vander Meer, R. K.; Lahav, S.; Soroker, V.; Hefetz, A. Chemom. Intell. Lab. Instrum. 2003, 66 (1), 51-62. (31) Lavine, B. K.; Davidson, C. E.; Breneman, C.; Katt, W. J. Chem. Inf. Sci. 2003, 43, 1890-1905. (32) Lavine, B. K.; Davidson, C. E.; Rayens, W. T. Comb. Chem. High Throughput Screening 2004, 7, 115-131. (33) Lavine, B. K.; Davidson, C. E.; Westover, D. J. J. Chem. Inf. Comput. Sci. 2004, 44 (3), 1056-1064.

Figure 5. PCA plot for the DOX data set for a blank sample containing culture medium in the absence and presence of four different concentrations of A. faecalis: (5) 1.2 × 105, (6) 2.5 × 105, (7) 6.3 × 105, (8) 18 × 105, and (13) 0 cells/mL.

time windows in the data. The boosting routine used this information to steer the population to an optimal solution. After 100 generations, the pattern recognition GA identified 7 features characteristic of bacterial concentration. Figure 5 shows the results obtained when applying PCA to the seven time domain features from the DOX data set for A. faecalis. As can be seen, the blank (culture medium without bacteria) is clearly differentiated as a separate class from the samples containing the bacteria on the largest principal component of the data. Moreover, results obtained with the DOX-PCA system for four different concentrations of A. faecalis, B. cereus, and B. subtillis bacteria show that the method is able to differentiate between different concentrations of all three bacteria (see Figure 6), showing separate classes for each of the configurations tested. These results support our hypothesis that there is a significant difference in the consumption of oxygen by different bacteria. Therefore, classification based on different concentrations of bacteria is possible with the DOXPCA system. When the same concentration of the three bacteria was tested, the DOX-PCA results demonstrate that Alcaligenes (group 2) can be differentiated from Bacillus (groups 1 and 3) groups reasonably well (see Figure 7). However, samples from group 1 (B. cereus) and from group 3 (B. subtillis) overlap. Evidently, the DOX-PCA system is classifying these two Bacilli as members of the same class. Bacteria Differentiation with the DOX-PCA System via Antibiotics Susceptibility. Most types of bacteria, when placed in growth medium at an optical density (600 nm) of 0.1 (typically corresponding to the 106 cfu/mL range), will consume all of the dissolved oxygen within 6 h resulting in zero current at -700 mV. When the selected antibiotic partially inhibits cell growth, the oxygen consumption curve produced is retarded, having a smaller slope and taking a longer time to reach zero current versus a control culture growing in antibiotic-free medium. If the antibiotic induces cell death, no oxygen is consumed and the curve approaches a shape similar to that produced by a control consisting of growth medium containing no cells. Additionally, the time it takes for an antibiotic to diffuse across the cell

Figure 6. PCA plot for the DOX data of four concentrations of A. faecalis, B. cereus, and B. subtillis. (a) (A. faecali): (5) ) 1.2 × 105, (6) 2.5 × 105, (7) 6.3 × 105, and (8) 18 × 105 cells/mL. (b) (B. cereus): (1) 4.3 × 105, (2) 8.6 × 105, (3) 21 × 105, and (4) 65 × 105 cells/mL (c) (B. subtilis): (9) 1.6 × 105, (10) 3.2 × 105, (11) 8.0 × 105, and (12) 24 × 105 cells/mL.

membrane and begin to affect cellular function varies for different types of cells and antibiotic combinations.20 This can be seen on the DOX oxygen consumption curve when the electrode is Analytical Chemistry, Vol. 77, No. 24, December 15, 2005

7947

Figure 7. PCA plot showing differentiation between Alcaligenes and Bacillus at concentration of 2 × 106 cells/mL: (1) B. cereus, (2) A. faecalis, and (3) B. subtilis.

immersed in a well containing a solution of cell and antibiotic begins to deviate from an electrode in a control well containing only cells. We have used this information to create DOX “fingerprints” for a variety of bacteria based on oxygen consumption patterns in the presence of small concentrations of various antibiotics. The setup can be viewed as analogous to an electronic nose, where an array of different polymer electrodes produces a specific response for a gaseous analyte; the data comprising the response are then analyzed by a computational method.17 The electronic nose has been used to identify bacteria growing in solution based upon the composition of the volatile chemicals in the headspace directly above the liquid culture.11 In contrast, the DOX-96 system described here utilizes uniform electrodes throughout the array. The response of the individual electrodes in the array is indirectly altered, compared to that of cells growing in medium, by the addition of an antibiotic to a well in the 96-well plate. By using three different antibiotics in separate wells, a unique fingerprint can be created for a specific bacterium. The best results have been obtained at low concentrations (0.5 µg/mL tetracycline, 2.5 µg/mL for ampicillin and chloramphenicol) where growth/respiration of most microorganisms was not completely inhibited. At these concentrations, the oxygen consumption patterns in the antibiotic-prepped wells fall between the controls containing cells only or medium only. When antibiotics are added to the wells, three types of signals can be produced; (i) The antibiotic may have no effect on cellular oxygen consumption. In this case, the antibiotic concentration is too low to have an effect on cell growth and the response produced by the wells with the antibiotic added is nearly identical to the control wells containing only cells. (ii) The antibiotic slows the growth of the cells, but all oxygen is eventually consumed. In this case, the antibiotic has the effect of slowing cell growth/respiration. The cells continue to grow and eventually consume all of the oxygen in the well, but at a longer time when compared to the control without antibiotic. In all cases, at the beginning of the experiment, the response between the control and the wells containing the antibiotic are nearly identical for a short time because some time 7948 Analytical Chemistry, Vol. 77, No. 24, December 15, 2005

Figure 8. (A) PCA plot showing discrimination of the 11 bacterial pathogens in the presence of ampicilin: (1) A. odorans, (2) B. globigii, (3) C. acidovorans, (4) C. glutamicum, (5) E. adecarboxylata, (6) E. coli (ATCC 25922), (7) E. coli (K12), (8) E. coli (SM10), (9) M. luteus, (10) S. epidermidis, and (11) Y. ruckeri. (B) PCA plot showing discrimination of the 11 bacterial pathogens in the presence of chloramphenicol: (1) A. odorans, (2) B. globigii, (3) C. acidovorans, (4) C. glutamicum, (5) E. adecarboxylata, (6) E. coli (ATCC 25922), (7) E. coli (K12), (8) E. coli (SM10), (9) M. luteus, (10) S. epidermidis, and (11) Y. ruckeri. (C) PCA plot showing discrimination of the 11 bacterial pathogens in the presence of tetracycline: (1) A. odorans, (2) B. globigii, (3) C. acidovorans, (4) C. glutamicum, (5) E. adecarboxylata, (6) E. coli (ATCC 25922), (7) E. coli (K12), (8) E. coli (SM10), (9) M. luteus, (10) S. epidermidis, and (11) Y. ruckeri.

the appropriate set of antibiotics, which would be applied individually to growth medium prior to measuring the bacterial response using the DOX system. Furthermore, this system can be used to quantify the resistance of different strains of bacteria to particular drugs.

Figure 9. PCA plot of the 11 pathogens in the absence of any antibiotics in the growth medium: (1) A. odorans, (2) B. globigii, (3) C. acidovorans, (4) C. glutamicum, (5) E. adecarboxylata, (6) E. coli (ATCC 25922), (7) E. coli (K12), (8) E. coli (SM10), (9) M. luteus, (10) S. epidermidis, and (11) Y. ruckeri.

is needed for the drug to permeate the cell membrane and take effect. (iii) The antibiotic completely inhibits cellular oxygen consumption. Again, the signal from the wells containing the antibiotic starts similar to the control. However, in this case, the cell growth/respiration is reduced to a point where the signal takes the shape similar to that of medium only, indicating that oxygen consumption is approaching zero, and therefore, the signal never reaches zero current. When three different antibiotics are used, these fingerprints become unique to a specific cell and the data can be analyzed with pattern recognition techniques. Pattern Recognition for Bacteria + Antibiotics. Figure 8A shows a principal component plot of the 20 time domain features identified by the pattern recognition GA for bacterial response to ampicilin in the DOX system. Each bacterial pathogen clusters in a distinct region of the principal component map developed from the 20 measurement variables. Figure 8B shows a principal component plot of nine time domain features identified by the pattern recognition GA for bacterial response to chloramphenicol. With the exception of B. globigii, all of the pathogens are well separated in the plot. Figure 8C shows a principal component plot of 13 time domain features identified by the pattern recognition GA for bacterial response to tetracycline. Again, all bacterial classes are well separated in the principal component map. When the bacterial response to the DOX system without any drugs (see Figure 9) is compared to the response with a drug present in the growth medium, it is apparent that addition of an antibiotic to the growth medium can improve the capability of the DOX system to discriminate among pathogens. We believe this is a significant result. For example, it may be possible to develop unique fingerprints for hard-to-classify pathogens by using

CONCLUSION In the present work, pattern recognition techniques have been used in combination with an electrochemical oxygen multisensor array for rapid and routine classification of different classes of bacterial pathogens and strains. This unique combination between a multisensor system and pattern recognition enables classification and differentiation of bacteria at species and stain level. Since the method does not require culturing bacteria, additional reagent, or incubation time, it can be used for screening of samples in which potential bacterial pathogens might be present. The proposed DOX-PCA system allows bacteria classification and differentiation by measuring the difference in the oxygen consumption using the electrochemical sensor, while the PCA provides a simple visual classification and recognition of the results. In addition, due to the unique experimental setup and procedure that involves indirect use of antibiotics, the same test could be used for obtaining practical information on the type, resistance, and dose of antibiotic necessary to establish optimum diagnosis, treatment, and decontamination strategies. The data show that adding small concentrations of broad spectrum antibiotics to growth medium creates unique changes in oxygen consumption for different types of bacteria. For the antibiotics selected in this study, the results were not dependent on the taxonomy of the cells selected. However, the antibiotics can be selected to build a computational library based on the researcher’s interest. For example, if Gram-negative bacteria are of interest, a narrow spectrum antibiotic, selective for Grampositive bacteria, can be added to inhibit Gram-positive bacteria and eliminate possible misidentifications from these types of cells. For this study, three antibiotics proved to be sufficient to discriminate between the 11 types of bacteria. However, if a large library is going to be constructed, it may be desirable to use more antibiotics to produce more confident results. The only limitation on the number of antibiotics that can be used is the 96-well plate and the number of replicates the researcher wishes to run. ACKNOWLEDGMENTS We acknowledge the following agencies for funding: Environmental Protection Agency through the STAR program (RD83090601), National Science Foundation CHE-0513470, and the NYS Center for Advanced Technology (IEEC). The authors also acknowledge Leslie White (BRIDGES student) and Yachao Zhang for help with the DOX. Received for review July 8, 2005. Accepted September 28, 2005. AC0512150

Analytical Chemistry, Vol. 77, No. 24, December 15, 2005

7949