- Email: [email protected]
Diagnosis of cerebral cryptococcoma using a computerized analysis of 1H NMR spectra in an animal model
Diagnostic Microbiology and Infectious Disease 52 (2005) 101 – 105 www.elsevier.com/locate/diagmicrobio
Mycology
Diagnosis of cerebral cryptococcoma using a computerized analysis of 1H NMR spectra in an animal model Theresa E. Dzendrowskyja,b, Brion Dolenkoc, Tania C. Sorrelld, Rajmund L. Somorjaic, Richard Malike, Carolyn E. Mountforda,b, Uwe Himmelreicha,b,d,T a Institute for Magnetic Resonance Research, P.O. Box 148, New South Wales 2065, Australia Department of Magnetic Resonance in Medicine, University of Sydney, Sydney, New South Wales 2006, Australia c Institute for Biodiagnostics, National Research Council of Canada, Winnipeg, Manitoba, Canada R3B 1Y6 d Centre for Infectious Diseases and Microbiology, ICPMR, University of Sydney at Westmead Hospital, Westmead, New South Wales 2145, Australia e Faculty of Veterinarian Sciences, University of Sydney, Sydney, New South Wales 2006, Australia Received 12 July 2004; accepted 5 February 2005 b
Abstract Viable cryptococci load in biopsy material from an animal model of cerebral cryptococcoma were correlated with 1H NMR spectra and metabolite profiles. A statistical classification strategy was applied to distinguish among high-resolution 1H NMR spectra acquired from cryptococcomas, glioblastomas, and normal brain tissue. The overall classification accuracy was 100% when a genetic-algorithm-based optimal region selection preceded the development of linear discriminant analysis-based classifiers. The method remained robust despite differences in the microbial load of the cryptococcoma group when harvested at different time points. These results indicate the feasibility of the method for diagnosis without isolation of the pathogenic microorganism and its potential for in vivo diagnosis based on computerized analysis of magnetic resonance spectra. D 2005 Elsevier Inc. All rights reserved. Keywords: Proton NMR; Statistical classification strategy; Cryptococcoma; Glioblastoma; Animal model; Rapid methods
1. Introduction Potentially life-threatening neurological infections caused by the yeast Cryptoccoccus neoformans most commonly present as meningitis, but in up to 14% of cases, clinical manifestations result from circumscribed lesions known as cryptococcomas, in the brain parenchyma (Casadevall and Perfect, 1998; Chen et al., 2000). Brain biopsy is required for diagnosis when lesions are confined to the brain or in the absence of other diagnostic material (Mitchell et al., 1995), because the pathological characteristics of infective lesions cannot be reliably distinguished by modalities such as computed tomography or magnetic resonance imaging (Andreula et al., 1993; Garg et al., 2004). Nuclear magnetic resonance (NMR) spectroscopy
T Corresponding author. Max-Planck-Institute for Neurological Research, D-50931 Cologne, Germany. Tel.: +49-221-4726321; fax: +49-221-4726337. E-mail address: [email protected] (U. Himmelreich). 0732-8893/$ – see front matter D 2005 Elsevier Inc. All rights reserved. doi:10.1016/j.diagmicrobio.2005.02.004
has been applied successfully to the noninvasive clinical diagnosis of some acute bacterial abscesses based on its ability to detect low molecular weight molecules from microbial cells and/or cells recruited during the host response to infection (Garg et al., 2004; Lai et al., 2002; Himmelreich et al., 2005). Interpretation of clinical spectra is subjective and is problematic in cases where typical marker metabolites are not detectable. We identified 2 marker metabolites of C. neoformans, a,a-trehalose and mannitol, in NMR spectra of biopsy samples from an animal model of acute cerebral and pulmonary cryptococcoma (Himmelreich et al., 2001; Himmelreich et al., 2003a). Resonances of these compounds were readily observed in 1-dimensional (1D) 1H NMR spectra. However, the resonances from trehalose and mannitol were reduced or not clearly distinguished in spectra from some animals at a later stage of infection. The application of computerized data analysis using pattern recognition techniques to determine the pathological status of tissue based on NMR spectra would provide an
102
T.E. Dzendrowskyj et al. / Diagnostic Microbiology and Infectious Disease 52 (2005) 101–105
objective and robust means for diagnosis. Pattern recognition techniques have been applied to in vivo NMR spectra for tumor diagnosis and grading using cluster analysis (Hagberg, 1998), artificial neural networks (Somorjai et al., 1996; Poptani et al., 1999), or linear discriminant analysis (LDA) (Somorjai et al., 1996; Preul et al., 1996). A statistical classification strategy (SCS) was specifically developed for the analysis of spectroscopic data from biofluids and tissue biopsies, where data sets contain many fewer spectra than data points (attributes) present in each spectrum (Somorjai et al., 2002; Nikulin et al., 1998). This method has resulted in the development of highly accurate and reliable classifiers for a variety of clinical and diagnostic problems, including the identification of pathogenic yeasts in cell suspensions (Himmelreich et al., 2003b; Baumgartner et al., 2004). It was our aim to develop a robust and objective method to distinguish between tumors and cryptococcoma, collected at different stages of evolution, and to correlate the load of viable cryptococci with trehalose content. This approach was used to allow for variations in the time between natural acquisition of human infection and clinical presentation and thus provide a robust, rapid, and objective method to distinguish between cryptococcomas and the major differential diagnosis, namely, cerebral tumors.
2. Materials and methods
Guidelines and with ethical approval from the University of Sydney Animal Ethics Committee. Viable cryptococci in the cryptococcomas were quantified for 12 animals. 2.2. NMR spectroscopy 1
H NMR spectra were obtained at 37 8C on a Bruker Avance 360 MHz NMR spectrometer using a 5-mm [1H, 13 C] inverse-detection dual-frequency probe. 1H NMR spectra were acquired with acquisition parameters as follows: frequency 360.13 MHz, pulse angle 908 (6 –7 As), repetition time 2.3 s, 4096 data points, 128 transients, spectral width 3600 Hz. Deuterium lock was used to optimize magnetic homogeneity. Water suppression was performed by selective excitation field gradients. Spectra were processed using Bruker xwinnmr 3.1 software. Chemical shift calibration was performed by setting the center of the spectrum to 4.65 ppm (the nominal position of the water resonance with respect to tetramethylsilane in PBS/D2O at 37 8C). Two-dimensional (2D) homo- and heteronuclear correlation spectra were acquired from cryptococcomas harvested at selected time points to assign 1H NMR resonances to specific compounds according to Himmelreich et al. (2001). The trehalose concentration was estimated using calibrated cross peak volumes from 1H, 1H COSY spectra compared with p-amino benzoic acid as a concentration standard in 12 animals with cerebral cryptococcoma (Himmelreich et al., 2001).
2.1. Animal model
2.3. Microbiology
A virulent strain of C. gattii serotype B (McBride), isolated from a cat, was obtained from the Westmead Hospital culture collection and used to establish cerebral cryptococcomas as described previously (Himmelreich et al., 2001). In brief, male Wistar-Furth rats (body weight 200–250 g) were anesthetized and the head immobilized in a stereotactic head frame. Coordinates for cerebral microinjections were 2.2 mm below dura; lateral: 3 mm; anteriorposterior: +2.4 mm relative to ear bar zero. Five microliters of phosphate-buffered saline (PBS) containing 5 104 CFU C. neoformans (n = 64), 5 AL containing 1 106 CFU C6 tumor cells (n = 50), or 5 AL of PBS (n = 45) were injected slowly. Infected and control animals were killed 7, 14, 21, and 28 days after injection or when they developed signs of illness. All other parameters were as stated in the work of Himmelreich et al. (2001). Pathology was confirmed by microscopy of tissue sections stained with hematoxylin-eosin or periodic acidSchiff reagent. Brain samples (maximum diameter of 5 mm) obtained from each animal were suspended in PBScontaining deuterated water (PBS/D2O, Australian Nuclear Science and Technology Organization, Lucas Heights, Australia), snap-frozen in liquid nitrogen, and stored at 70 8C for up to 4 months before NMR spectroscopy. Animal experimentation was carried out according to the Australian National Health and Medical Research Council
Cryptococcomas from 12 animals (3 at each time point) were weighed, dispersed, and spread on plates made from serial dilutions of cryptococcal suspensions. The number of colony-forming units per gram was determined after 48 h of incubation at 35 8C. Cryptococci were biotyped and serotyped (Crypto Check agglutination test, Iatron Labs, Chiba, Japan). 2.4. Classification of NMR spectra An SCS approach was used for data analysis (Somorjai et al., 2002; Nikulin et al., 1998). NMR data were prepared using software developed in house (Xprep, IBD, NRC, Winnipeg, Canada). Magnitude spectra, consisting of 4096 data points over a spectral width of 10 ppm, were reduced to 1500 points between 0.35 and 4.05 ppm. The spectra were normalized to unit area in this region. The correct alignment of the NMR spectra was verified by simultaneous and sequential display of all NMR spectra using the lipid/lactate resonance at 1.3 ppm. The magnitude NMR spectra were analyzed by a geneticalgorithm-based optimal region selection process (GAORS) to reduce the number of attributes and hence eliminate redundant information (Nikulin et al., 1998). Two maximally discriminatory subregions in the 1D NMR spectra of each class (pathology) were selected for development of LDA-based classifiers. The averages of these subregions
T.E. Dzendrowskyj et al. / Diagnostic Microbiology and Infectious Disease 52 (2005) 101–105
103
were used to develop LDA-based pairwise classifiers for all combinations of the 3 pathologies. These classifiers were made robust by a bootstrap-based cross-validation method developed in-house (IBD, NRC, Winnipeg, Canada) (Somorjai et al., 2002). Specifically, about half the spectra from each class were selected at random and used to train an LDA classifier (the btraining setQ). The remaining spectra were then tested against this classifier. This process was repeated 10 000 times (with random replacement), and the optimized LDA coefficients were saved. The ultimate classifier consisted of the weighed output of the 10 000 different bootstrap classifier coefficient sets. Each classifier yielded probabilities of class assignment for the individual spectra. Class assignment was defined crisp if the probability of belonging to 1 class was larger than 75% (Nikulin et al., 1998). Correct classification refers to assignment of a spectrum to the same class as histopathological examination with a classification probability N75%. Indeterminate or fuzzy classification refers to assignment of a spectrum to a class with classification probability V 75%. 2.5. Validation Classifiers were evaluated using independent validation sets that consisted of NMR data from biopsy samples of animals that were not part of the classifier development.
3. Results 3.1. Animal model Fifty-seven of 64 animals developed histologically confirmed cryptococcomas. Three animals died during or immediately after surgery and 4 animals did not develop brain lesions. 3.2. Microbiology Viable cryptococci (in CFU per gram) in lesions harvested at each time point are summarized in Table 1. Only data from animals with lesions larger than approximately 5 mm3 have been included to minimize errors in determining the weight of smaller lesions. Cryptococcal loads reached a maximum at day 7 and decreased thereafter.
Fig. 1. 1H NMR spectra from lesion material isolated from animals with cerebral cryptococcoma at different time points: (A) 7, (B) 14, (C) 21, and (D) 28 days after initiation of the infection. Abbreviations: AA = amino acid residues, ac = acetate, gln/glu = glutamine/glutamate, lac = lactate, lip = lipid, man = mannose, NCH3 = choline containing metabolites; tre = trehalose.
3.3. NMR spectroscopy Table 1 Weight-specific cell counts of C. neofomans in lesion material from cerebral cryptococcoma and trehalose concentration (mean F standard deviation) Time
n
CFU per gram lesion material
7 days 14 days 21 days 28 days
3 4 4 5
2 7 3 5
F F F F
1 2 1 1
106 105 105 104
Trehalose (Ag) per gram lesion material 150 50 18 2
F F F F
20 10 5 1
Three animals per time point were used. n refers to the number of samples.
NMR spectra were collected from the 57 cases of cryptococcoma (12–15 animals killed at 7, 14, 21, and 28 days after infection). Randomly chosen NMR spectra from cryptococcomas at each time point are shown in Fig. 1. As in previous studies (Himmelreich et al., 2001), metabolites characteristic of normal brain tissue were absent. At days 7 and 14 postinfection, all spectra were dominated by resonances from trehalose, mannitol, and lipids. These resonances have previously been identified as metabolites of C. neoformans (trehalose, mannitol, and lipids) or the host
104
T.E. Dzendrowskyj et al. / Diagnostic Microbiology and Infectious Disease 52 (2005) 101–105
Table 2 Classification of NMR spectra from rat brain biopsy samples using SCS
1. Training set Number of animals Crisp classification (%) Specificity (%) Sensitivity (%) Spectral regions (ppm) selected by GA-ORS 2. Validation set Number of animals Crisp classification (%) Specificity (%) Sensitivity (%)
Cryptococcoma versus glioma
Cryptococcoma versus control
Glioma versus control
49/42 100
49/37 100
42/37 100
100 100 3.43 – 3.47
100 100 2.73–2.75
100 100 1.73–1.83
3.93 – 4.01
3.61–3.71
3.68–3.72
8/8 88
8/8 100
8/8 100
88 100
100 100
100 100
1 cryptococcoma (day 21), which was classified as glioma, and 1 cryptococcoma (day 28), which was indeterminate.
4. Discussion
cellular response (lipids) (Himmelreich et al., 2003a). NMR spectra acquired from cryptococcomas 21 or 28 days after infection contained generally lower metabolite concentrations, as demonstrated by a decrease in the signal-to-noise ratio by a factor of 4 –6 compared with spectra from the same amount of lesion material collected at early time points (Fig. 1). Trehalose and/or mannitol were identifiable in COSY spectra of cryptococcomas from 11 of the 14 animals at day 21 and 7 of the 12 animals at day 28. For the remaining COSY spectra, signal-to-noise ratios were insufficient for unequivocal identification of these marker metabolites. 3.4. Correlation of trehalose concentration with cell counts The concentration of trehalose per gram of lesion material was estimated relative to a concentration standard added to the NMR tube, using 1H, 1H COSY spectra. The correlation between the number of live organisms per gram of lesion material with trehalose concentration is illustrated in Table 1. Both declined over time with a relatively greater decline in the trehalose concentration. 3.5. Statistical classification strategy Forty-nine NMR spectra from the cryptococcomas and 42 from experimental rat cerebral glioblastomas obtained from a previous study (Himmelreich et al., 2001) were used as a btraining setQ to develop the SCS-based classifiers for identification. The results are shown in Table 2. Cryptococcomas and glioblastomas were distinguished with an accuracy of 100% using 2 most discriminatory spectral regions identified by the GA-ORS algorithm. Additional spectra (validation set) from 2 cryptococcomas per time point (n = 8), 8 from gliomas, and 8 from saline-injected animals were tested against the classifiers. All spectra from the validation set were classified correctly, except for
Identification of C. neoformans-specific metabolites in cerebral cryptococcomas by NMR spectroscopy is a potentially attractive diagnostic approach, as it does not require isolation and culture of the microorganism prior diagnosis. It can avoid the complications of neurosurgery and biopsy when applied in vivo in connection with magnetic resonance imaging (Garg et al., 2004; Lai et al., 2002). When applied to ex vivo biopsy samples, the method avoids culture of the microorganisms for their identification. Previous studies in animal models indicated that characteristic metabolites including the disaccharide trehalose and mannitol were reproducibly present in experimental cryptococcomas sampled early in the course of infection, when the host cellular response is limited (Himmelreich et al., 2001; Himmelreich et al., 2003a). The present study indicates that the concentration of these metabolites in experimental cryptococcomas decreases over time, with decreasing numbers of viable cryptococci in these lesions, presumably because of an effective host cell response. The observation that trehalose levels, which were maximal at 14 days, declined more rapidly than the cryptococcal load in mature lesions suggests that cryptococci were less exposed to factors that contribute to the accumulation of trehalose, such as osmotic stress (Wiemken, 1990). The only metabolite that did not decrease in concentration with time was acetate. Acetate was evident in all samples obtained 21 and 28 days after infection and only in some early cryptococcomas. This was most likely because of the decrease in all other metabolite concentrations. However, the acetate concentration was still relatively low and was not a suitable marker metabolite as it is also regularly found in bacterial abscess (Garg et al., 2004; Lai et al., 2002). It is notable that operator-based assignment of resonances of NMR spectra from some cryptococcomas at the later time points lacked discernable resonances due to trehalose. They resembled to some extent NMR spectra obtained from glioblastomas. However, computerized data analysis of NMR spectra from lesion material was able to distinguish between cryptococcoma and glioma where operator-based identification of key metabolites failed. It has been shown previously that SCS and other computerized data analysis methods are more powerful for diagnosis of disease and identification of microorganisms than those based on the recognition of particular bmarkerQ metabolites or the determination of resonance peak ratios (Somorjai et al., 2002; Himmelreich et al., 2003b; Himmelreich et al., 2005). In this study, SCS distinguished reliably between glioblastoma and cryptococcoma despite substantial variability in NMR spectral characteristics within one of the groups (cryptococcoma). Regions in the NMR spectra
T.E. Dzendrowskyj et al. / Diagnostic Microbiology and Infectious Disease 52 (2005) 101–105
selected by GA-ORS refer mainly to carbohydrates/polyols (3.53–4.01 and 3.43–3.47 ppm) for the distinction between cryptococcoma and glioblastoma. This suggests that trehalose and mannitol, although not identified unequivocally by 2-dimensional NMR methods, still contribute to diagnostic resonances when analyzed by NMR/SCS. This is of considerable importance in establishing NMR spectroscopy as a potentially useful clinical tool, because human cases of cryptococcosis typically present with mature rather than very early lesions and lesions in different patients are likely to be of different degrees of chronicity. Robust and objective diagnosis of cryptococcomas will shorten the time before institution of appropriate therapy. Although the classification on the training set of cryptococcoma and glioblastoma spectra resulted in 100% diagnostic accuracy, the ultimate test for robustness of a classification is the verification with an independent validation set. Classifiers developed on relatively small data sets would lead to overly optimistic classification results with attendant poor generalizability, that is, new samples, not used in developing the classifier, will not classify accurately. The fact that in the validation set only 1 sample was classified as indeterminate and 1 was misclassified indicates that the classifiers are robust and that overtraining is unlikely. It remains to be determined if this methodology is applicable to noninvasive diagnosis of cryptococcomas in humans but our unpublished initial results have shown that the acquisition of in vivo NMR spectra from cerebral cryptococcomas is feasible. Acknowledgments The authors thank Ms Susan Dowd for assistance with the animal experiments and Dr Heide-Marie Daniel for assistance with yeast identification and quantification. References Andreula CF, Burdi N, Carella A (1993) CNS cryptococcosis in AIDS: Spectrum of MR findings. J Comput Assist Tomogr 17:438 – 441. Baumgartner R, Somorjai R, Bowman C, Sorrell TC, Mountford CE, Himmelreich U (2004) Unsupervised feature dimension reduction for classification of MR spectra. Magn Reson Imaging 22:251 – 256. Casadevall A, Perfect JE (1998) Cryptococcus neoformans. Washington, DC7 American Society for Microbiology Press. Chen SC, Sorrell TC, Nimmo G, Speed B, Currie B, Ellis D, Marriott D, Pfeiffer T, Parr D, Byth K (2000) Epidemiology and host- and varietydependent characteristics of infection due to Cryptococcus neoformans
105
in Australia and New Zealand. Australasian Cryptococcal Study Group. Clin Infect Dis 31:499 – 508. Garg M, Gupta RK, Husain M, Chawla S, Chawla A, Kumar R, Rao SB, Misra MK, Prasad KN (2004) Brain abscesses: Etiologic categorization with in vivo proton MR spectroscopy. Radiology 230:519 – 527. Hagberg G (1998) From magnetic resonance spectroscopy to classification of tumours. A review of pattern recognition methods. NMR Biomed 11:148 – 156. Himmelreich U, Dzendrowskyj TE, Allen C, Dowd S, Malik R, Shehan PB, Russell P, Mountford CE, Sorrell TC (2001) Cryptococcomas distinguished from gliomas with MR spectroscopy: An experimental rat and cell culture study. Radiology 220:122 – 128. Himmelreich U, Allen C, Dowd S, Malik R, Shehan PB, Russell P, Mountford CE, Sorrell TC (2003a) Identification of metabolites of importance in the pathogenesis of pulmonary cryptococcoma using nuclear magnetic resonance spectroscopy. Microbes Infect 5:285 – 290. Himmelreich U, Somorjai RL, Dolenko B, Lee OC, Daniel HM, Murray R, Mountford CE, Sorrell TC (2003b) Rapid identification of Candida species by using nuclear magnetic resonance spectroscopy and a statistical classification strategy. Appl Environ Microbiol 69:4566 – 4574. Himmelreich U, Accurso R, Malik R, Dolenko B, Somorjai RL, Gupta RK, Gomes L, Mountford CE, Sorrell TC (2005) Identification of Staphylococcus aureus brain abscesses: Rat and human studies using 1H magnetic resonance spectroscopy. Radiology. [ Forthcoming]. Lai PH, Ho JT, Chen WL, Hsu SS, Wang JS, Pan HB, Yang CF (2002) Brain abscess and necrotic brain tumor: Discrimination with proton MR spectroscopy and diffusion-weighted imaging. Am J Neuroradiol 23:1369 – 1377. Mitchell DH, Sorrell TC, Allworth AM, Heath CH, McGregor AR, Papanaoum K, Richards MJ, Gottlieb T (1995) Cryptococcal disease of the CNS in immunocompetent hosts: Influence of cryptococcal variety on clinical manifestations and outcome. Clin Infect Dis 20:611 – 616. Nikulin A, Dolenko B, Bezabeh T, Somorjai RL (1998) Near-optimal region selection for feature space reduction: Novel preprocessing methods for classifying MR spectra. NMR Biomed 11:209 – 217. Poptani H, Kaartinen J, Gupta RK, Niemitz M, Hiltunen Y, Kauppinen RA (1999) Diagnostic assessment of brain tumors and non-neoplastic brain disorders in vivo using proton nuclear magnetic resonance spectroscopy and artificial neural networks. J Cancer Res 125:343 – 349. Preul MC, Caramons Z, Collins D, Villemure J, Leblanc R, Olivier A, Pokrupa R, Arnold DL (1996) Accurate, noninvasive diagnosis of human brain tumour by using proton magnetic resonance spectroscopy. Nat Med 2:323 – 325. Somorjai RL, Dolenko B, Nikulin AK, Pizzi N, Scarth G, Zhilkin P, Halliday W, Fewer D, Hill N, Ross I, West M, Smith ICP, Donnelly SM, Kuesel AC, Briere KM (1996) Classification of 1H MR spectra of human brain neoplasms: The influence of preprocessing and computerized consensus diagnosis on classification accuracy. J Magn Reson Imaging 6:437 – 444. Somorjai RL, Dolenko B, Nikulin A, Nickerson P, Rush D, Shaw A, Glogowski M, Rendell J, Deslauriers R (2002) Distinguishing normal from rejecting renal allografts: Application of a three-stage classification strategy to MR and IR of urine. Vibr Spectrosc 28:97 – 102. Wiemken A (1990) Trehalose in yeast, stress protectant rather than reserve carbohydrate. Antonie Van Leeuwenhoek 58:209 – 217.
Mycology
Diagnosis of cerebral cryptococcoma using a computerized analysis of 1H NMR spectra in an animal model Theresa E. Dzendrowskyja,b, Brion Dolenkoc, Tania C. Sorrelld, Rajmund L. Somorjaic, Richard Malike, Carolyn E. Mountforda,b, Uwe Himmelreicha,b,d,T a Institute for Magnetic Resonance Research, P.O. Box 148, New South Wales 2065, Australia Department of Magnetic Resonance in Medicine, University of Sydney, Sydney, New South Wales 2006, Australia c Institute for Biodiagnostics, National Research Council of Canada, Winnipeg, Manitoba, Canada R3B 1Y6 d Centre for Infectious Diseases and Microbiology, ICPMR, University of Sydney at Westmead Hospital, Westmead, New South Wales 2145, Australia e Faculty of Veterinarian Sciences, University of Sydney, Sydney, New South Wales 2006, Australia Received 12 July 2004; accepted 5 February 2005 b
Abstract Viable cryptococci load in biopsy material from an animal model of cerebral cryptococcoma were correlated with 1H NMR spectra and metabolite profiles. A statistical classification strategy was applied to distinguish among high-resolution 1H NMR spectra acquired from cryptococcomas, glioblastomas, and normal brain tissue. The overall classification accuracy was 100% when a genetic-algorithm-based optimal region selection preceded the development of linear discriminant analysis-based classifiers. The method remained robust despite differences in the microbial load of the cryptococcoma group when harvested at different time points. These results indicate the feasibility of the method for diagnosis without isolation of the pathogenic microorganism and its potential for in vivo diagnosis based on computerized analysis of magnetic resonance spectra. D 2005 Elsevier Inc. All rights reserved. Keywords: Proton NMR; Statistical classification strategy; Cryptococcoma; Glioblastoma; Animal model; Rapid methods
1. Introduction Potentially life-threatening neurological infections caused by the yeast Cryptoccoccus neoformans most commonly present as meningitis, but in up to 14% of cases, clinical manifestations result from circumscribed lesions known as cryptococcomas, in the brain parenchyma (Casadevall and Perfect, 1998; Chen et al., 2000). Brain biopsy is required for diagnosis when lesions are confined to the brain or in the absence of other diagnostic material (Mitchell et al., 1995), because the pathological characteristics of infective lesions cannot be reliably distinguished by modalities such as computed tomography or magnetic resonance imaging (Andreula et al., 1993; Garg et al., 2004). Nuclear magnetic resonance (NMR) spectroscopy
T Corresponding author. Max-Planck-Institute for Neurological Research, D-50931 Cologne, Germany. Tel.: +49-221-4726321; fax: +49-221-4726337. E-mail address: [email protected] (U. Himmelreich). 0732-8893/$ – see front matter D 2005 Elsevier Inc. All rights reserved. doi:10.1016/j.diagmicrobio.2005.02.004
has been applied successfully to the noninvasive clinical diagnosis of some acute bacterial abscesses based on its ability to detect low molecular weight molecules from microbial cells and/or cells recruited during the host response to infection (Garg et al., 2004; Lai et al., 2002; Himmelreich et al., 2005). Interpretation of clinical spectra is subjective and is problematic in cases where typical marker metabolites are not detectable. We identified 2 marker metabolites of C. neoformans, a,a-trehalose and mannitol, in NMR spectra of biopsy samples from an animal model of acute cerebral and pulmonary cryptococcoma (Himmelreich et al., 2001; Himmelreich et al., 2003a). Resonances of these compounds were readily observed in 1-dimensional (1D) 1H NMR spectra. However, the resonances from trehalose and mannitol were reduced or not clearly distinguished in spectra from some animals at a later stage of infection. The application of computerized data analysis using pattern recognition techniques to determine the pathological status of tissue based on NMR spectra would provide an
102
T.E. Dzendrowskyj et al. / Diagnostic Microbiology and Infectious Disease 52 (2005) 101–105
objective and robust means for diagnosis. Pattern recognition techniques have been applied to in vivo NMR spectra for tumor diagnosis and grading using cluster analysis (Hagberg, 1998), artificial neural networks (Somorjai et al., 1996; Poptani et al., 1999), or linear discriminant analysis (LDA) (Somorjai et al., 1996; Preul et al., 1996). A statistical classification strategy (SCS) was specifically developed for the analysis of spectroscopic data from biofluids and tissue biopsies, where data sets contain many fewer spectra than data points (attributes) present in each spectrum (Somorjai et al., 2002; Nikulin et al., 1998). This method has resulted in the development of highly accurate and reliable classifiers for a variety of clinical and diagnostic problems, including the identification of pathogenic yeasts in cell suspensions (Himmelreich et al., 2003b; Baumgartner et al., 2004). It was our aim to develop a robust and objective method to distinguish between tumors and cryptococcoma, collected at different stages of evolution, and to correlate the load of viable cryptococci with trehalose content. This approach was used to allow for variations in the time between natural acquisition of human infection and clinical presentation and thus provide a robust, rapid, and objective method to distinguish between cryptococcomas and the major differential diagnosis, namely, cerebral tumors.
2. Materials and methods
Guidelines and with ethical approval from the University of Sydney Animal Ethics Committee. Viable cryptococci in the cryptococcomas were quantified for 12 animals. 2.2. NMR spectroscopy 1
H NMR spectra were obtained at 37 8C on a Bruker Avance 360 MHz NMR spectrometer using a 5-mm [1H, 13 C] inverse-detection dual-frequency probe. 1H NMR spectra were acquired with acquisition parameters as follows: frequency 360.13 MHz, pulse angle 908 (6 –7 As), repetition time 2.3 s, 4096 data points, 128 transients, spectral width 3600 Hz. Deuterium lock was used to optimize magnetic homogeneity. Water suppression was performed by selective excitation field gradients. Spectra were processed using Bruker xwinnmr 3.1 software. Chemical shift calibration was performed by setting the center of the spectrum to 4.65 ppm (the nominal position of the water resonance with respect to tetramethylsilane in PBS/D2O at 37 8C). Two-dimensional (2D) homo- and heteronuclear correlation spectra were acquired from cryptococcomas harvested at selected time points to assign 1H NMR resonances to specific compounds according to Himmelreich et al. (2001). The trehalose concentration was estimated using calibrated cross peak volumes from 1H, 1H COSY spectra compared with p-amino benzoic acid as a concentration standard in 12 animals with cerebral cryptococcoma (Himmelreich et al., 2001).
2.1. Animal model
2.3. Microbiology
A virulent strain of C. gattii serotype B (McBride), isolated from a cat, was obtained from the Westmead Hospital culture collection and used to establish cerebral cryptococcomas as described previously (Himmelreich et al., 2001). In brief, male Wistar-Furth rats (body weight 200–250 g) were anesthetized and the head immobilized in a stereotactic head frame. Coordinates for cerebral microinjections were 2.2 mm below dura; lateral: 3 mm; anteriorposterior: +2.4 mm relative to ear bar zero. Five microliters of phosphate-buffered saline (PBS) containing 5 104 CFU C. neoformans (n = 64), 5 AL containing 1 106 CFU C6 tumor cells (n = 50), or 5 AL of PBS (n = 45) were injected slowly. Infected and control animals were killed 7, 14, 21, and 28 days after injection or when they developed signs of illness. All other parameters were as stated in the work of Himmelreich et al. (2001). Pathology was confirmed by microscopy of tissue sections stained with hematoxylin-eosin or periodic acidSchiff reagent. Brain samples (maximum diameter of 5 mm) obtained from each animal were suspended in PBScontaining deuterated water (PBS/D2O, Australian Nuclear Science and Technology Organization, Lucas Heights, Australia), snap-frozen in liquid nitrogen, and stored at 70 8C for up to 4 months before NMR spectroscopy. Animal experimentation was carried out according to the Australian National Health and Medical Research Council
Cryptococcomas from 12 animals (3 at each time point) were weighed, dispersed, and spread on plates made from serial dilutions of cryptococcal suspensions. The number of colony-forming units per gram was determined after 48 h of incubation at 35 8C. Cryptococci were biotyped and serotyped (Crypto Check agglutination test, Iatron Labs, Chiba, Japan). 2.4. Classification of NMR spectra An SCS approach was used for data analysis (Somorjai et al., 2002; Nikulin et al., 1998). NMR data were prepared using software developed in house (Xprep, IBD, NRC, Winnipeg, Canada). Magnitude spectra, consisting of 4096 data points over a spectral width of 10 ppm, were reduced to 1500 points between 0.35 and 4.05 ppm. The spectra were normalized to unit area in this region. The correct alignment of the NMR spectra was verified by simultaneous and sequential display of all NMR spectra using the lipid/lactate resonance at 1.3 ppm. The magnitude NMR spectra were analyzed by a geneticalgorithm-based optimal region selection process (GAORS) to reduce the number of attributes and hence eliminate redundant information (Nikulin et al., 1998). Two maximally discriminatory subregions in the 1D NMR spectra of each class (pathology) were selected for development of LDA-based classifiers. The averages of these subregions
T.E. Dzendrowskyj et al. / Diagnostic Microbiology and Infectious Disease 52 (2005) 101–105
103
were used to develop LDA-based pairwise classifiers for all combinations of the 3 pathologies. These classifiers were made robust by a bootstrap-based cross-validation method developed in-house (IBD, NRC, Winnipeg, Canada) (Somorjai et al., 2002). Specifically, about half the spectra from each class were selected at random and used to train an LDA classifier (the btraining setQ). The remaining spectra were then tested against this classifier. This process was repeated 10 000 times (with random replacement), and the optimized LDA coefficients were saved. The ultimate classifier consisted of the weighed output of the 10 000 different bootstrap classifier coefficient sets. Each classifier yielded probabilities of class assignment for the individual spectra. Class assignment was defined crisp if the probability of belonging to 1 class was larger than 75% (Nikulin et al., 1998). Correct classification refers to assignment of a spectrum to the same class as histopathological examination with a classification probability N75%. Indeterminate or fuzzy classification refers to assignment of a spectrum to a class with classification probability V 75%. 2.5. Validation Classifiers were evaluated using independent validation sets that consisted of NMR data from biopsy samples of animals that were not part of the classifier development.
3. Results 3.1. Animal model Fifty-seven of 64 animals developed histologically confirmed cryptococcomas. Three animals died during or immediately after surgery and 4 animals did not develop brain lesions. 3.2. Microbiology Viable cryptococci (in CFU per gram) in lesions harvested at each time point are summarized in Table 1. Only data from animals with lesions larger than approximately 5 mm3 have been included to minimize errors in determining the weight of smaller lesions. Cryptococcal loads reached a maximum at day 7 and decreased thereafter.
Fig. 1. 1H NMR spectra from lesion material isolated from animals with cerebral cryptococcoma at different time points: (A) 7, (B) 14, (C) 21, and (D) 28 days after initiation of the infection. Abbreviations: AA = amino acid residues, ac = acetate, gln/glu = glutamine/glutamate, lac = lactate, lip = lipid, man = mannose, NCH3 = choline containing metabolites; tre = trehalose.
3.3. NMR spectroscopy Table 1 Weight-specific cell counts of C. neofomans in lesion material from cerebral cryptococcoma and trehalose concentration (mean F standard deviation) Time
n
CFU per gram lesion material
7 days 14 days 21 days 28 days
3 4 4 5
2 7 3 5
F F F F
1 2 1 1
106 105 105 104
Trehalose (Ag) per gram lesion material 150 50 18 2
F F F F
20 10 5 1
Three animals per time point were used. n refers to the number of samples.
NMR spectra were collected from the 57 cases of cryptococcoma (12–15 animals killed at 7, 14, 21, and 28 days after infection). Randomly chosen NMR spectra from cryptococcomas at each time point are shown in Fig. 1. As in previous studies (Himmelreich et al., 2001), metabolites characteristic of normal brain tissue were absent. At days 7 and 14 postinfection, all spectra were dominated by resonances from trehalose, mannitol, and lipids. These resonances have previously been identified as metabolites of C. neoformans (trehalose, mannitol, and lipids) or the host
104
T.E. Dzendrowskyj et al. / Diagnostic Microbiology and Infectious Disease 52 (2005) 101–105
Table 2 Classification of NMR spectra from rat brain biopsy samples using SCS
1. Training set Number of animals Crisp classification (%) Specificity (%) Sensitivity (%) Spectral regions (ppm) selected by GA-ORS 2. Validation set Number of animals Crisp classification (%) Specificity (%) Sensitivity (%)
Cryptococcoma versus glioma
Cryptococcoma versus control
Glioma versus control
49/42 100
49/37 100
42/37 100
100 100 3.43 – 3.47
100 100 2.73–2.75
100 100 1.73–1.83
3.93 – 4.01
3.61–3.71
3.68–3.72
8/8 88
8/8 100
8/8 100
88 100
100 100
100 100
1 cryptococcoma (day 21), which was classified as glioma, and 1 cryptococcoma (day 28), which was indeterminate.
4. Discussion
cellular response (lipids) (Himmelreich et al., 2003a). NMR spectra acquired from cryptococcomas 21 or 28 days after infection contained generally lower metabolite concentrations, as demonstrated by a decrease in the signal-to-noise ratio by a factor of 4 –6 compared with spectra from the same amount of lesion material collected at early time points (Fig. 1). Trehalose and/or mannitol were identifiable in COSY spectra of cryptococcomas from 11 of the 14 animals at day 21 and 7 of the 12 animals at day 28. For the remaining COSY spectra, signal-to-noise ratios were insufficient for unequivocal identification of these marker metabolites. 3.4. Correlation of trehalose concentration with cell counts The concentration of trehalose per gram of lesion material was estimated relative to a concentration standard added to the NMR tube, using 1H, 1H COSY spectra. The correlation between the number of live organisms per gram of lesion material with trehalose concentration is illustrated in Table 1. Both declined over time with a relatively greater decline in the trehalose concentration. 3.5. Statistical classification strategy Forty-nine NMR spectra from the cryptococcomas and 42 from experimental rat cerebral glioblastomas obtained from a previous study (Himmelreich et al., 2001) were used as a btraining setQ to develop the SCS-based classifiers for identification. The results are shown in Table 2. Cryptococcomas and glioblastomas were distinguished with an accuracy of 100% using 2 most discriminatory spectral regions identified by the GA-ORS algorithm. Additional spectra (validation set) from 2 cryptococcomas per time point (n = 8), 8 from gliomas, and 8 from saline-injected animals were tested against the classifiers. All spectra from the validation set were classified correctly, except for
Identification of C. neoformans-specific metabolites in cerebral cryptococcomas by NMR spectroscopy is a potentially attractive diagnostic approach, as it does not require isolation and culture of the microorganism prior diagnosis. It can avoid the complications of neurosurgery and biopsy when applied in vivo in connection with magnetic resonance imaging (Garg et al., 2004; Lai et al., 2002). When applied to ex vivo biopsy samples, the method avoids culture of the microorganisms for their identification. Previous studies in animal models indicated that characteristic metabolites including the disaccharide trehalose and mannitol were reproducibly present in experimental cryptococcomas sampled early in the course of infection, when the host cellular response is limited (Himmelreich et al., 2001; Himmelreich et al., 2003a). The present study indicates that the concentration of these metabolites in experimental cryptococcomas decreases over time, with decreasing numbers of viable cryptococci in these lesions, presumably because of an effective host cell response. The observation that trehalose levels, which were maximal at 14 days, declined more rapidly than the cryptococcal load in mature lesions suggests that cryptococci were less exposed to factors that contribute to the accumulation of trehalose, such as osmotic stress (Wiemken, 1990). The only metabolite that did not decrease in concentration with time was acetate. Acetate was evident in all samples obtained 21 and 28 days after infection and only in some early cryptococcomas. This was most likely because of the decrease in all other metabolite concentrations. However, the acetate concentration was still relatively low and was not a suitable marker metabolite as it is also regularly found in bacterial abscess (Garg et al., 2004; Lai et al., 2002). It is notable that operator-based assignment of resonances of NMR spectra from some cryptococcomas at the later time points lacked discernable resonances due to trehalose. They resembled to some extent NMR spectra obtained from glioblastomas. However, computerized data analysis of NMR spectra from lesion material was able to distinguish between cryptococcoma and glioma where operator-based identification of key metabolites failed. It has been shown previously that SCS and other computerized data analysis methods are more powerful for diagnosis of disease and identification of microorganisms than those based on the recognition of particular bmarkerQ metabolites or the determination of resonance peak ratios (Somorjai et al., 2002; Himmelreich et al., 2003b; Himmelreich et al., 2005). In this study, SCS distinguished reliably between glioblastoma and cryptococcoma despite substantial variability in NMR spectral characteristics within one of the groups (cryptococcoma). Regions in the NMR spectra
T.E. Dzendrowskyj et al. / Diagnostic Microbiology and Infectious Disease 52 (2005) 101–105
selected by GA-ORS refer mainly to carbohydrates/polyols (3.53–4.01 and 3.43–3.47 ppm) for the distinction between cryptococcoma and glioblastoma. This suggests that trehalose and mannitol, although not identified unequivocally by 2-dimensional NMR methods, still contribute to diagnostic resonances when analyzed by NMR/SCS. This is of considerable importance in establishing NMR spectroscopy as a potentially useful clinical tool, because human cases of cryptococcosis typically present with mature rather than very early lesions and lesions in different patients are likely to be of different degrees of chronicity. Robust and objective diagnosis of cryptococcomas will shorten the time before institution of appropriate therapy. Although the classification on the training set of cryptococcoma and glioblastoma spectra resulted in 100% diagnostic accuracy, the ultimate test for robustness of a classification is the verification with an independent validation set. Classifiers developed on relatively small data sets would lead to overly optimistic classification results with attendant poor generalizability, that is, new samples, not used in developing the classifier, will not classify accurately. The fact that in the validation set only 1 sample was classified as indeterminate and 1 was misclassified indicates that the classifiers are robust and that overtraining is unlikely. It remains to be determined if this methodology is applicable to noninvasive diagnosis of cryptococcomas in humans but our unpublished initial results have shown that the acquisition of in vivo NMR spectra from cerebral cryptococcomas is feasible. Acknowledgments The authors thank Ms Susan Dowd for assistance with the animal experiments and Dr Heide-Marie Daniel for assistance with yeast identification and quantification. References Andreula CF, Burdi N, Carella A (1993) CNS cryptococcosis in AIDS: Spectrum of MR findings. J Comput Assist Tomogr 17:438 – 441. Baumgartner R, Somorjai R, Bowman C, Sorrell TC, Mountford CE, Himmelreich U (2004) Unsupervised feature dimension reduction for classification of MR spectra. Magn Reson Imaging 22:251 – 256. Casadevall A, Perfect JE (1998) Cryptococcus neoformans. Washington, DC7 American Society for Microbiology Press. Chen SC, Sorrell TC, Nimmo G, Speed B, Currie B, Ellis D, Marriott D, Pfeiffer T, Parr D, Byth K (2000) Epidemiology and host- and varietydependent characteristics of infection due to Cryptococcus neoformans
105
in Australia and New Zealand. Australasian Cryptococcal Study Group. Clin Infect Dis 31:499 – 508. Garg M, Gupta RK, Husain M, Chawla S, Chawla A, Kumar R, Rao SB, Misra MK, Prasad KN (2004) Brain abscesses: Etiologic categorization with in vivo proton MR spectroscopy. Radiology 230:519 – 527. Hagberg G (1998) From magnetic resonance spectroscopy to classification of tumours. A review of pattern recognition methods. NMR Biomed 11:148 – 156. Himmelreich U, Dzendrowskyj TE, Allen C, Dowd S, Malik R, Shehan PB, Russell P, Mountford CE, Sorrell TC (2001) Cryptococcomas distinguished from gliomas with MR spectroscopy: An experimental rat and cell culture study. Radiology 220:122 – 128. Himmelreich U, Allen C, Dowd S, Malik R, Shehan PB, Russell P, Mountford CE, Sorrell TC (2003a) Identification of metabolites of importance in the pathogenesis of pulmonary cryptococcoma using nuclear magnetic resonance spectroscopy. Microbes Infect 5:285 – 290. Himmelreich U, Somorjai RL, Dolenko B, Lee OC, Daniel HM, Murray R, Mountford CE, Sorrell TC (2003b) Rapid identification of Candida species by using nuclear magnetic resonance spectroscopy and a statistical classification strategy. Appl Environ Microbiol 69:4566 – 4574. Himmelreich U, Accurso R, Malik R, Dolenko B, Somorjai RL, Gupta RK, Gomes L, Mountford CE, Sorrell TC (2005) Identification of Staphylococcus aureus brain abscesses: Rat and human studies using 1H magnetic resonance spectroscopy. Radiology. [ Forthcoming]. Lai PH, Ho JT, Chen WL, Hsu SS, Wang JS, Pan HB, Yang CF (2002) Brain abscess and necrotic brain tumor: Discrimination with proton MR spectroscopy and diffusion-weighted imaging. Am J Neuroradiol 23:1369 – 1377. Mitchell DH, Sorrell TC, Allworth AM, Heath CH, McGregor AR, Papanaoum K, Richards MJ, Gottlieb T (1995) Cryptococcal disease of the CNS in immunocompetent hosts: Influence of cryptococcal variety on clinical manifestations and outcome. Clin Infect Dis 20:611 – 616. Nikulin A, Dolenko B, Bezabeh T, Somorjai RL (1998) Near-optimal region selection for feature space reduction: Novel preprocessing methods for classifying MR spectra. NMR Biomed 11:209 – 217. Poptani H, Kaartinen J, Gupta RK, Niemitz M, Hiltunen Y, Kauppinen RA (1999) Diagnostic assessment of brain tumors and non-neoplastic brain disorders in vivo using proton nuclear magnetic resonance spectroscopy and artificial neural networks. J Cancer Res 125:343 – 349. Preul MC, Caramons Z, Collins D, Villemure J, Leblanc R, Olivier A, Pokrupa R, Arnold DL (1996) Accurate, noninvasive diagnosis of human brain tumour by using proton magnetic resonance spectroscopy. Nat Med 2:323 – 325. Somorjai RL, Dolenko B, Nikulin AK, Pizzi N, Scarth G, Zhilkin P, Halliday W, Fewer D, Hill N, Ross I, West M, Smith ICP, Donnelly SM, Kuesel AC, Briere KM (1996) Classification of 1H MR spectra of human brain neoplasms: The influence of preprocessing and computerized consensus diagnosis on classification accuracy. J Magn Reson Imaging 6:437 – 444. Somorjai RL, Dolenko B, Nikulin A, Nickerson P, Rush D, Shaw A, Glogowski M, Rendell J, Deslauriers R (2002) Distinguishing normal from rejecting renal allografts: Application of a three-stage classification strategy to MR and IR of urine. Vibr Spectrosc 28:97 – 102. Wiemken A (1990) Trehalose in yeast, stress protectant rather than reserve carbohydrate. Antonie Van Leeuwenhoek 58:209 – 217.