| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
|
|
|
|||||||
|
|||||||||
Technical Briefs |
Departments of1 Microbiology and Molecular Cell Biology and2 Urology, Eastern Virginia Medical School, Norfolk, VA;3 Virginia Prostate Center, Norfolk, VA;4 Department of Urology, Laikon Hospital, Athens, Greece
aaddress correspondence to this author at: Foundation for Biomedical Research, Academy of Athens, Soranou Efesiou 4, Athens, Greece 11527; fax 30-210-6597-545, e-mail vlahoua{at}evms.edu or vlahoua{at}bioacademy.gr
At present, the most reliable means of diagnosis and surveillance of bladder cancer are cystoscopic examination and bladder biopsy for histologic confirmation. The invasive and labor-intensive nature of this procedure underscores the need to develop better, less costly, and nonsurgical diagnostic tools (1)(2). Use of surface-enhanced laser desorption/ionization (SELDI) time-of-flight mass spectrometry has been successful in facilitating protein profiling of complex biological mixtures. This technology uses chemical affinity platforms to capture protein molecules from various biological sources. Retained proteins are subsequently analyzed by mass spectrometry [reviewed in Ref. ((3))]. Recent reports provide evidence that analysis of SELDI data by "learning" algorithms can lead to the identification of serum protein "fingerprints" for prostate, ovarian, and breast cancers (4)(5)(6)(7)(8)(9) and urinary fingerprints for kidney cancer (10). We recently reported the application of the SELDI system for detection of potential bladder cancer-associated biomarkers in urine (11). In this earlier study, we showed that combination of five transitional cell carcinoma (TCC)-associated protein peaks by simple statistical methods provided 87% sensitivity and 66% specificity in disease detection. The objectives of the current study were (a) to evaluate a commercial available data-mining classification algorithm for the analysis of the SELDI mass spectral data, and (b) assessing the clinical utility of this assay in detecting bladder cancer from a geographically and clinically mixed population.
Fresh spot-voided urine specimens from 230 individuals were included in the study. Specimens were collected from patients seen in the Departments of Urology at Eastern Virginia Medical School (Norfolk, VA) and Laikon Hospital (Athens Greece). In all cases, patients gave consent according to the regulations for human subject protection of each institution. TCC (n = 105 patients) was confirmed histologically at the time of specimen collection. Controls included patients with other diseases of the urogenital tract (n = 92) and individuals with no evidence of disease (n = 33). The same protocols for urine collection were followed at both institutions.
Before application to the weak cation-exchange protein ionization chips, urine samples were centrifuged briefly to remove exfoliated cells. The protein concentrations of the supernatants were then estimated by use of the Microprotein-PR assay (Sigma). Samples of the study and control groups as well as duplicates of the same samples were assigned random positions on the chips by an "in-house", Excel-based randomization program. Mass spectrometry was performed with a PBS2 SELDI time-of-flight mass spectrometer (Ciphergen Biosystems Inc.).
We used the SELDI software (Ver. 3.1) to label spectral peaks and normalize their intensities to the total ion current (mass range, 2.5–30 kDa) to account for variation in ionization efficiencies. Peak masses were aligned, and clustering was performed with the Biomarker Wizard software (Ciphergen Biosystems). Pattern recognition and sample classification were performed with the Biomarker Pattern Software (Ciphergen Biosystems), which is based on classification and regression tree analysis (12).
The urinary bladder cancer (UBC; IDL Biotech) and BTAstat (Bion Diagnostic Sciences) tests (13)(14)(15) were performed according to the manufacturers’ instructions. For the UBC, a cutoff value of 12 µg/L was selected based on ROC curve analysis (13) and the recommendations of the test manufacturer (for experimental details and additional patient information, see the Materials-Methods file and Tables 1–3 in the Data Supplement that accompanies the online version of this Technical Brief at http://www.clinchem.org/content/vol50/issue8/).
One urine sample designated as quality control was included on every chip array to monitor instrument reproducibility. From the analysis of a total of 89 quality-control spectra, the CV for mass designation for each quality-control calibration peak (total of four peaks in the range 2.8–12 kDa) was 0.05–0.2%, and the variation for intensity of the same peaks was 40–70% (data not shown).
We used 191 samples for the learning set for the algorithm, and 39 samples were given to the SELDI user in a blind fashion and comprised the test set. The Biomarker Pattern Software algorithm splits the training dataset into two bins based on decision rules (Fig. 1
, squares). The rules are formed by the peak intensities for a designated mass value. Samples that follow the rule (i.e., peak intensity is equal to or less than the cutoff intensity value) go to the left daughter node. When splitting can no longer be performed, terminal nodes are generated and classified according to the samples in the majority (Fig. 1
, circles). Seven protein peaks generated a decision tree that appeared to provide optimal differentiation between the bladder cancer and control groups during the algorithm evaluation (Fig. 1
). With the exception of the peaks at 7.07 and 9.1 kDa, the rest of the main splitters had significantly different intensities between the cancer and control groups (P <0.05).
|
In cross-validation, bladder cancer was detected with 84% sensitivity and 74% specificity. In the blinded test set, a sensitivity of 83% (95% confidence interval, 61–94%) and a specificity of 67% (45–83%) were obtained. In comparison, the BTAstat and the UBC tests (12)(13)(14) detected TCC with 78% and 72% sensitivity and 67% and 71% specificity, respectively. Interestingly, the SELDI decision tree captured more of the "early" (grade I and II) cancers (5 of 6 compared with 2 of 6 for BTAstat and 4 of 6 for the UBC test; Table 1
). Nevertheless, the responses of the three tests were found to be independent of each other (P >0.05); therefore, their combination did not improve the overall diagnostic rates (for details on the splitters and classification rates, see Figs. 1 and 2 and Tables 4 and 5 in the online Data Supplement).
|
Protein profiling of bladder cancer in urine by use of proteomic technologies in combination with bioinformatic tools has not been reported previously. The detection rates we observed in the current study were similar to the ones reported in our earlier study involving a smaller cohort of patients. It should be noted, however, that the two studies are not directly comparable because we used different chip chemistries because of improved experimental protocols. Concerted efforts to explore novel proteomic and genomic approaches as well as bioinformatic tools for combinatorial analysis are of paramount importance to improve the management of bladder cancer. In the case of SELDI profiling, areas for improvement include increasing the volume of the proteome being sampled by combining different types of chip chemistries; use of microfluidics for automated fractionation; and use of new, more accurate, sensitive "peak-picking" computational approaches. Concurrent with these refinements will be the need for validation studies that specifically address the ability to detect early-stage and -grade tumors and to do so from among patients with inflammatory conditions more likely to be encountered in a clinical setting.
Acknowledgments
This work was supported by the Elsa U. Pardee Research Foundation, the National Cancer Institute Early Detection Research Network (Grant CA85067), and the Virginia Prostate Center.
References
The following articles in journals at HighWire Press have cited this article:
|
|
 |
|
  H. Bensmail, J. Golek, M. M. Moody, J. O. Semmes, and A. Haoudi A novel approach for clustering proteomics data using Bayesian fast Fourier transform Bioinformatics, May 15, 2005; 21(10): 2210 - 2224. [Abstract] [Full Text] [PDF]
|
|
|
|
 |
|
 K. R. Coombes Analysis of Mass Spectrometry Profiles of the Serum Proteome Clin. Chem., January 1, 2005; 51(1): 1 - 2. [Full Text] [PDF]
|
|
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
