Purpose: Our study was undertaken to determine the utility of plasma proteomic profiling using surface-enhanced laser desorption/ionization time-of-flight (SELDI-TOF) mass spectrometry for the detection of head and neck squamous cell carcinomas (HNSCCs).
Experimental Design: Pretreatment plasma samples from HNSCC patients or controls without known neoplastic disease were analyzed on the Protein Biology System IIc SELDI-TOF mass spectrometer (Ciphergen Biosystems, Fremont, CA). Proteomic spectra of mass:charge ratio (m/z) were generated by the application of plasma to immobilized metal-affinity-capture (IMAC) ProteinChip arrays activated with copper. A total of 37,356 data points were generated for each sample. A training set of spectra from 56 cancer patients and 52 controls were applied to the “Lasso” technique to identify protein profiles that can distinguish cancer from noncancer, and cross-validation was used to determine test errors in this training set. The discovery pattern was then used to classify a separate masked test set of 57 cancer and 52 controls. In total, we analyzed the proteomic spectra of 113 cancer patients and 104 controls.
Results: The Lasso approach identified 65 significant data points for the discrimination of normal from cancer profiles. The discriminatory pattern correctly identified 39 of 57 HNSCC patients and 40 of 52 noncancer controls in the masked test set. These results yielded a sensitivity of 68% and specificity of 73%. Subgroup analyses in the test set of four different demographic factors (age, gender, and cigarette and alcohol use) that can potentially confound the interpretation of the results suggest that this model tended to overpredict cancer in control smokers.
Conclusions: Plasma proteomic profiling with SELDI-TOF mass spectrometry provides moderate sensitivity and specificity in discriminating HNSCC. Further improvement and validation of this approach is needed to determine its usefulness in screening for this disease.
Head and neck squamous cell carcinoma (HNSCC) is the fifth most common malignancy worldwide (1) . Despite modern intervention, the 5-year survival rate for this disease has improved only marginally over the past decade, and recurrent disease is observed in ∼50% of all patients (2) . Patients with early-stage cancer often manifest minimal/nonspecific physical symptoms or findings, resulting in delayed diagnosis, which translates into poor tumor control and survival. In addition, HNSCC patients often have widespread field changes in their upper-aerodigestive track mucosa and, therefore, are susceptible for developing secondary cancers after successful treatment of the primary malignancy (3) . Therefore, any innovation that facilitates early detection of these tumors has the potential to improve survival and quality of life in this patient group. Recent studies have focused on the use of biomarkers and gene array technology for detection, surveillance, and prognostication in these patients (4) . Several of these studies require availability of tumor and normal tissues and significant expertise in tissue processing, immunohistochemical studies, and gene array analysis. The time required to perform many of these studies may not be practical for routine clinical use for patient screening and prognostication. There is a real need for a relatively simple, high-throughput, noninvasive, and universally available method to identify patients with early-stage tumors.
The proteome is the complete repertoire of proteins that contribute to a cell’s physiological phenotype. Recent advancements in proteomic analysis have yielded novel techniques to aid in biomarker identification (5 , 6) . One such advancement is the development of surface enhanced laser desorption/ionization time-of-flight mass spectrometry (SELDI-TOF MS), a high throughput and extremely sensitive proteomic approach that allows protein expression profiling of large sample sets (7 , 8) . Briefly, in SELDI, proteins of interest from biologically complex samples bind selectively to chemically modified affinity surfaces on the ProteinChip (Ciphergen Biosystems, Fremont CA), with nonspecifically bound impurities washed away. The retained sample is complexed with an energy-absorbing molecule and is analyzed by SELDI-TOF MS, producing spectra of mass:charge ratio (m/z). In combination with pattern-recognition statistical tools, this technology was successfully applied for the early detection of ovarian, breast, and prostate cancers from patient sera (9, 10, 11, 12, 13) .
In this study, we used the SELDI-TOF MS technique to generate multiple proteomic spectra from plasma samples of patients with HNSCC and those without known neoplastic disease. We used a bioinformatics method known as the “Lasso” algorithm developed by Tibshirani (14) and Efron et al. (15) to extrapolate proteomic patterns that can best discriminate HNSCC patients from noncancer controls in a training set of 109 subjects. We then applied it to a separate masked test set of 107 patients. We show that this approach can be used to separate HNSCC patients from noncancer controls with moderate sensitivity and specificity.
PATIENTS AND METHODS
From March 1999 to December 2002, as part of a Stanford University Institutional Review Board-approved trial, pretreatment citrated plasma samples were collected from 143 consenting patients with a histologically confirmed diagnosis of HNSCC. Of these, 116 individual samples were available for SELDI-TOF study. Because the proteomic spectra from three samples were uninterpretable because of gross hemolysis, 113 HNSCC patients formed the cohort of this study. Plasma from 104 controls without known neoplastic disease was also obtained. All of the study subjects (HNSCC and healthy volunteers) signed an Institutional Review Board-approved informed consent form, and all guidelines for human subject investigation were followed.
Staging evaluation for HNSCC patients consisted of history and physical examination, panendoscopy with examination under anesthesia, complete blood count, liver function tests, and computed tomography or magnetic resonance imaging head and neck imaging. All of the patients had biopsy-proven squamous cell carcinoma involving one or multiple head and neck mucosal subsites, and all were staged according to the 1988 American Joint Committee on Cancer staging (16) .
Noncancer controls were recruited from the pulmonary function laboratory, the pulmonary clinic (for management of chronic obstructive pulmonary disease), the cardiac clinic (for management of cardiovascular disease), the phlebotomy laboratory (for routine blood drawn), or the radiation oncology department (as accompanying family members of actively treated patients). All of the controls were free of cancer based on clinical history; however, no additional imaging approaches or routine marker assays were performed on the controls. Attempts were made to match for known risk factors for HNSCC (e.g., age, and extent of cigarette and alcohol use) between cancer patients and noncancerous volunteers as much as possible.
For all patient and control specimens, a 3.5-ml blood sample was obtained by venipuncture into a vacutainer with 3.2% sodium citrate, centrifuged at 3000 rpm at 4°C for 10 min within 30 min of collection. The plasma was aliquoted and stored at −80°C until assayed. Only samples that had undergone less than two freeze/thaw cycles were used for proteomic analysis.
SELDI ProteinChip Array Binding.
Ten μl of PBS (pH 7.4), was added to 10 μl of thawed plasma sample. Eight-spot immobilized metal-affinity capture arrays (IMAC-3) were activated with 100 mm CuSO4 for 15 min, followed by a distilled water rinse. The activated array surfaces were equilibrated with 10 μl of PBS for 15 min. Five μl of diluted plasma sample were applied to the array surface and were allowed to bind at room temperature for 60 min. Plasma samples from patients with HNSCC and from healthy volunteers were applied in alternating fashion on the same chip. All of the samples were run in duplicate on a separate chip. After removal of the sample, each array was washed in PBS for 5 min, followed by a distilled water rinse.
SELDI-TOF MS Analysis.
After the spots on the IMAC-3 arrays were air-dried, 0.5 μl of a saturated solution of sinapinic acid in 0.5% (v/v) trifluoroacetic acid and 50% (v/v) acetonitrile was applied to each bait surface, allowed to air-dry, and reapplied. Mass/charge (m/z) spectra of proteins with affinity to the chelated metal surface (containing tryptophan, cysteine, histidine, or phosphorylated amino acid residues) were generated in a Ciphergen Protein Biology System IIc Time-of-Flight mass spectrometer. Laser shots (150) were averaged, with laser intensity of 170, detector sensitivity of 9, high mass to acquire 150 kDa, with optimization range of 3–30 kDa. External calibration was performed using β-endorphin (3465.0 Da), bovine ubiquitin (8564.8 Da; Ciphergen), and bovine insulin (5734.51 Da; Sigma-Aldrich, St. Louis, MO) as standards. All array binding and SELDI-TOF MS were performed on the same day. A subset of cancer and control samples were assayed on separate days also, to determine the reproducibility of the spectra.
Statistical Analysis of SELDI-TOF Mass Spectra.
All of the duplicated spectra were compiled, and the protein peak intensities were normalized to the total ion current of m/z values from 1.5 to 150 kDa using Ciphergen ProteinChip Software 3.1.1. For each sample, 37,356 m/z values were exported. m/z values of <1.5 kDa, corresponding to the signal from the sinapinic acid matrix, were omitted, yielding 32,055 points/sample for statistical analysis. We then applied the Lasso method (14 , 15) to find m/z values that discriminate normal profiles from cancer profiles. This method finds the linear combination of intensity values that best separates the two classes, with a constraint on the absolute norm of the weight vector. Use of this constraint prevents overfitting and automatically selects the sites that are most important for discriminating the classes. Ten-fold cross-validation was used to estimate the optimal constraint bound and to assess the performance of the resulting classifier. The LARS algorithm3 was used to fit the Lasso. The method identified the m/z values with the highest weighting for discriminating cancer versus control in the training set. This key subset of m/z values was then applied to an unrelated masked test set for classification of unknown samples. Sensitivity, specificity, and their standard errors were calculated for SELDI-TOF classification of cancer versus noncancer according to the true disease state in the test set.
Fig. 1⇓ shows examples of three SELDI-TOF spectra obtained either from the same run or from different runs for the plasma sample of the same subject. Intra-assay coefficient of variation (CV) for intensity based on duplicate sample testing was 22% and interassay CV for intensity was 23%. Both intra- and interassay CVs for m/z were less than 0.07%. These values were all consistent with the reproducibility data for the PBS IIc TOF mass spectrometer reported by the manufacturer (Ciphergen Biosystems, Fremont CA).
Detection of HNSCC.
Table 1⇓ lists the patient, tumor, and treatment characteristics of the HNSCC patients and noncancer controls. The majority of patients had stage III/IV disease. In the control group, there were fewer men, fewer smokers, and slightly higher median age.
Plasma samples of 56 HNSCC patients and 52 noncancerous controls formed the training set. Using the Lasso method, the 65 most significant data points were identified and applied to the masked test set for group classification. The m/z location of these points and their weights are shown in Table 2⇓ . The weights that are different from 0 signify a significant m/z value for this model. A weight >0 means a higher average intensity in cancer patients when compared with noncancer controls and vice versa. Interestingly, most of these data points had very low intensity and could have been mistaken for background “noise” if the spectra had not been magnified and if all of the m/z points had not been considered. Only 11 m/z values were associated with an obvious peak. Fig. 2⇓ shows a magnified version of representative SELDI-TOF mass spectra from two cancer patients and two controls. The data points located at 8,003 and 10,338 Da, identified by the statistical algorithm, could be seen. One was higher in cancer patients and one was higher in noncancer controls.
First, a 10-fold cross validation approach was applied to the training set to determine the predictive accuracy of the proteomic pattern generated by Lasso. The pattern was able to correctly identify cancer in 35 of 56 HNSCC patients (sensitivity of 63%) and noncancer in 40 of 52 healthy controls (specificity of 77%). The same model was then applied to a masked unrelated test set consisting of 57 HNSCC cancer patients and 52 noncancer controls. For this subject group, the plasma proteomic patterns could correctly identify 39 of 57 HNSCC patients and 38 of 52 volunteers with no known malignant disease, yielding sensitivity and a specificity value of 68% (SE ± 6%) and 73% (SE ± 7%), respectively (Table 3)⇓ . Fig. 3⇓ shows the receiver-operating characteristic curve for the training set.
Because some of the demographic factors may have influenced our ability to classify patient group correctly, we performed subset analysis by applying the same Lasso complexity parameters to subgroups in the test set stratified by gender, age group, cigarette use, alcohol use, and cancer stage. The results are shown in Table 3⇓ . There was a higher false-negative rate for cancer prediction in the female group; however, the sample size was rather small, and the difference was not statistically significant. There was also a higher false-positive rate in the control group with a history of cigarette or alcohol consumption. These differences were statistically significant only for cigarette consumption based on binomial testing (difference, 0.39; SE, 0.13).
The reproducibility of our model over time was tested by rerunning 11 samples ∼3 months after the initial experiment. When applying the model to these “rerun” samples as a new test set, class prediction was reproducible for 10 of 11 samples.
Although HNSCC is less likely than other solid neoplasms to disseminate distantly, it often presents as a locally advanced tumor at diagnosis that requires aggressive local and regional therapy. These aggressive treatments often result in significant functional morbidities that can adversely impact the quality of life and survival of these patients. For example, many functions involved in the activity of daily living such as speech, chewing, and swallowing are often compromised by surgery and radiation therapy as part of the management of these tumors. Therefore, early detection is critical in improving the care of HNSCC patients.
Unfortunately, at present, there exists no effective screening tool for these cancers. A number of novel molecular approaches to detect clonal genetic alterations in the blood and saliva of HNSCC patients have been developed for the early detection of these tumors. These include the detection of specific patterns of microsatellite markers (17) , promoter hypermethylation (18 , 19) , and mitochondrial DNA mutations (20) in patient saliva and serum. Although promising, these techniques are time consuming and require significant expertise in specimen processing. Proteomic analysis via SELDI-TOF MS technology, therefore, provides a new and exciting approach that can be used for high throughput detection of HNSCC.
SELDI-TOF MS has been shown by several groups to detect solid cancers with a high sensitivity and specificity (9, 10, 11, 12, 13 , 21, 22, 23) . In our study, we found that this technique can be used to detect HNSCC with moderate sensitivity of 68% and a specificity of 73%. Our results are inferior to other published data, specifically that reported by Sidransky et al. (24) . The discrepancy may be related to the intrinsic characteristics of the HNSCC patient population that we tested and the different technical proteomic approaches used by the two groups. Sidransky et al. (24) used matrix-assisted laser desorption and ionization (MALDI), which is similar to SELDI-TOF but without the surface preselection or enrichment steps used in SELDI. MALDI results are, therefore, less likely to be affected by technical artifacts such as consistency in surface coating and washing as found in SELDI. In addition, another inherent limitation of SELDI, which is not observed in MALDI, rests in the selective nature of specific ProteinChip surfaces because proteins and/or peptides that do not bind to the chip are washed away with buffer. A recent study evaluating the utility of SELDI-TOF in detecting prostate cancer showed classification accuracy of 42 and 67% for individual IMAC-3 and weak cation exchange (WCX) arrays, respectively (25) . However, when the spectral data of both types of arrays were combined, both sensitivity and specificity were improved to 85%. Therefore, the differences in results noted between the data of Sidransky et al (24) and our data may be related to these technical differences, and our results may be improved when multiple array surfaces are used and spectral data from these surfaces are combined. Overall, the discrepancies between the results of the two groups suggest that MALDI may be better than single-surface SELDI for detection of head and neck cancers.
Similar to most currently available data sets, our sample sizes are relatively small compared with the total number of detected m/z peaks. There is a real danger of selecting spurious m/z peaks purely by chance because of the artifacts in the data rather than because of the disease process. Therefore, a rigorous statistical approach was necessary to overcome this potential dilemma. We used the Lasso statistical method for this study because it can efficiently and accurately identify the most significant data points in a very large data set (32,055 data points/plasma sample) for the discrimination of the two patient groups, with built-in constraints on the weight vector to avoid overfitting. The main advantage of Lasso over other statistical methods such as support vector machines and linear discriminant analysis is that it provides automatic feature selection that is not offered by linear discriminant analysis or support vector machines (26) . In addition, Lasso is a least squares regression method, and when there are only two outcome classes to consider, Lasso is essentially equivalent to linear discriminant analysis. Lasso has been used successfully to correctly predict several prognostic factors that correlate with prostate-specific antigen levels in prostate cancer patients (27) . A 10-fold cross-validation approach was applied to the training set to substantiate the results found by the Lasso analysis. Then, a separate masked test data set was used to determine the sensitivity and specificity for group classification. These measures minimize the risk that the results can be obtained by chance alone. We have also compared Lasso with contingent valuation methods, linear discriminant analysis, and prediction analysis for microarray (PAM; data not shown), but we found that Lasso provided the least error rate for both training and test sets among all statistical approaches evaluated for this patient population.
A drawback of our study is the type of patients enrolled. Although we tried to match the control group with HNSCC patients as much as possible, the control group included a younger population, more women, and more nonsmokers. In addition, we did not systematically enroll patients with benign inflammatory conditions as controls. We did have one patient with myelodysplastic syndrome in the control test set, and he was classified as noncancer by our algorithm. To systematically determine the effect that some of these demographic parameters may have on our model, we applied the Lasso results to individual patient subsets in the test set, stratified by age, gender, and history of cigarette and alcohol use. We found that the model significantly overpredicted for cancer in the control group with a history of cigarette use. There was also a trend for overprediction of cancer in the control group with a history of alcohol use but the difference was not statistically significant. An overfit for heavy drinkers was also noted by Sidransky et al. (24) . A possible cause of this overprediction could be due to the presence of unknown circulating proteins or peptides that are related to heavy cigarette and/or alcohol use. The presence of these peptides may confound the prediction algorithms in subjects.
Another weakness of our study is that most HNSCC patients had stage III-IV tumors at diagnosis because this study was part of a larger study of tumor hypoxia markers and all of the patients were required to have tumors that were accessible for tumor oxygenation measurement. Therefore, the majority of cancer patients who qualified for this study had involved neck nodes at diagnosis. Although we found no statistically significant difference between in-group classification for stage I-III versus stage IV, the number in the early-stage group was too small to make any definitive statement. Therefore validating our results with a large group of patients with earlier stages is an important future goal.
The relatively low incidence of head and neck cancers, the rarity of a hereditary form of disease, and the lack of an easily available, patient-accepted diagnostic aid make population-based screening presently untenable for head and neck cancers, except in certain high-risk areas and high-risk populations. However, screening programs for these tumors may be more cost-effective if efforts were to focus on a high-risk population such as men over age 40 years who are tobacco and alcohol users. Dower et al. showed that two to three times the number of lives would be saved with oral cancer screening if only a selective high-risk population (a priori identified from a decision analysis simulation model) were targeted (28) . In India, periodic screening of high-risk groups resulted in more patients with early-stage tumors and less case-fatality rate in the screened group compared with the control group (29) . Unfortunately, the only presently widely available screening tool for head and neck cancers is visual screening, which can be highly subjective and operator dependent, and is useful only for oral cavity subsites. Other available tools such as fluorescence spectrometry or detection of clonal and genetic changes in blood and saliva, although promising, are available only at limited academic institutions with the required special expertise for analysis and do not allow for high-throughput screening. SELDI-TOF analysis has the promise of being a versatile and high-throughput technique for screening of head and neck cancers when combined with visual screening. However, its sensitivity and specificity for detecting these tumors need to be improved over what is achieved in our study, and the robustness of the studies needs to be reproducible and maintained over time before it can be considered useful clinically.
Although the identity of the dominant peaks used for group classification is not necessary for diagnosis, knowing the identity of the biomarkers is essential for understanding the biological role these proteins have in the formation of HNSCC and for developing future therapies. The major drawback of SELDI-TOF MS is the difficulty in determining the accurate identity of these dominant or most predictive peptide peaks. Similar to Petricoin et al. (10) , we found that most of the dominant data points have very low intensity and can easily be mistaken for background noise. This may be related to the rarity of these proteins in relation to the more abundant, yet less discriminatory, housekeeping proteins in the plasma. The low abundance of these proteins makes it very difficult to obtain an adequate quantity for protein identification and characterization. Other proteomic approaches such as serum or tissue MALDI may be helpful in determining the identity of these markers (24 ; 30 ).
In summary, we have found that SELDI-TOF MS in combination with sophisticated bioinformatics tools can distinguish HNSCC patients from noncancer control with moderate sensitivity and specificity. Further improvement and validation of this approach is necessary to determine its utility in the detection of head and neck cancers.
Grant support: Q-T. Le was supported by the USPHS Grant CA 67166. A. C. Koong was supported by the Damon Runyon–Lilly Clinical Investigator Award.
The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.
Note: S. G. Soltys and Q-T. Le contributed equally to this work. Presented at the Annual Meeting of American Society of Therapeutic-Radiology and Oncology, 2003.
Requests for reprints: Quynh-Thu Le. Department of Radiation Oncology Stanford University, 875 Blake Wilbur Drive, MC 5847, Stanford, CA 94305-5847, Tel: 650-498-5032, Fax: 650-725-3865, Email:
↵3 The LARS algorithm is available at http://www-stat.stanford.edu/∼hastie/Papers/LARS/.
- Received October 24, 2003.
- Revision received February 3, 2004.
- Accepted February 18, 2004.