© The Author 2007. Published by Oxford University Press.
ARTICLES |
A 25-Signal Proteomic Signature and Outcome for Patients With Resected Non–Small-Cell Lung Cancer
Affiliations of authors: Institute for Advanced Research, Nagoya University, Nagoya, Aichi, Japan (KY); Division of Molecular Carcinogenesis, Center for Neurological Diseases and Cancer, Nagoya University Graduate School of Medicine, Nagoya, Aichi, Japan (KY, ST, YS, TT); Departments of Pathology and Molecular Diagnostics (YY) and Thoracic Surgery (TM), Aichi Cancer Center Hospital, Nagoya, Aichi, Japan
Correspondence to: Kiyoshi Yanagisawa, MD, PhD, Institute for Advanced Research, Nagoya University, Furo-cho, Chikusa-ku, Nagoya, Aichi 464-8601, Japan (e-mail: kyana{at}med.nagoya-u.ac.jp).
 |
ABSTRACT |
|---|
 Top Abstract Context and CaveatsSubjects and MethodsResultsDiscussionReferencesNotes |
|---|
Background: Among patients with non–small-cell lung cancer (NSCLC), those with poor prognosis cannot be distinguished from those with good prognosis.
Methods: Matrix-assisted laser desorption–ionization mass spectrometry was used to analyze protein profiles of 174 specimens from NSCLC tumors and 27 specimens from normal lung tissue and to derive a prognosis-associated proteomic signature. Frozen resected tissue specimens were randomly divided into a training set (116 NSCLC and 20 normal lung specimens) and an independent, blinded validation set (58 NSCLC and seven normal lung specimens). Mass spectrometry signals from training set specimens that were differentially associated with specimens from patients with a high risk of recurrence (i.e., who died within 5 years of surgical treatment because of relapse) compared with those from patients with a low risk of recurrence (i.e., alive with no symptoms of relapse after a median follow-up of 89 months) were selected by use of the Fisher's exact test, the Kruskal–Wallis test, and the significance analysis of microarray test. These signals were used to build an individualized, weighted voting–based prognostic signature. The signature was then validated in the independent dataset. Survival was assessed by multivariable Cox regression analysis. Proteins corresponding to individual signals were identified by ion-trap mass spectrometry coupled with high-performance liquid chromatography. All statistical tests were two-sided.
Results: From 2630 mass spectrometry signals from specimens in the training cohort, we derived a signature of 25 signals that was associated with both relapse-free survival and overall survival. Among stage I NSCLC patients in the validation set, the signature was statistically significantly associated with both overall survival (hazard ratio [HR] of death for patients in the high-risk group compared with those in the low-risk group = 61.1, 95% confidence interval [CI] = 8.9 to 419.2, P<.001) and relapse-free survival (HR of relapse = 11.7, 95% CI = 3.1 to 44.8, P<.001). Proteins corresponding to signals in the signature were identified that had various cellular functions, including ribosomal protein L26-like 1, acylphosphatase, and phosphoprotein enriched in astrocytes 15.
Conclusions: We defined a mass spectrometry signature that was associated with survival among NSCLC patients and appeared to distinguish those with poor prognosis from those with good prognosis.
Prior knowledge Patients with non–small-cell lung cancer (NSCLC) who have poor prognosis cannot be distinguished from those with good prognosis. Study design A training set–validation set design was used to derive a proteomic signature that was associated with prognosis. Tissues were tumor and normal lung tissues from NSCLC patients with a high risk of recurrence or with a low risk of recurrence. Contribution A signature of 25 mass spectrometry signals was derived that could statistically significantly distinguish NSCLC tumors with good prognosis from those with poor prognosis. Implications The 25-signal proteomic signature may be useful to distinguish NSCLC patients with good prognosis from those with poor prognosis. Determination of the roles of the proteins represented in the proteomic signature in NSCLC tumorigenesis and progression may lead to improved treatment for NSCLC. Limitations Approximately half of proteins contributing to the proteomic signature have been identified. Additional resources are needed for the identification of proteins in matrix-assisted laser desorption–ionization mass spectrometry signals. Mass spectrometry signals are limited to relatively highly abundant proteins with a low molecular weight.
|
Lung cancer is the leading cause of cancer death in Japan (1), the United States (2), and many other developed countries (2). Approximately 80%–85% of the cases of lung cancer are classified as non–small-cell lung cancer (NSCLC) (1), and approximately 25% of patients diagnosed with NSCLC are eligible for surgical resection, which gives the best hope of a cure for patients. However, the long-term survival of NSCLC patients is still unsatisfactory, with no more than 50% surviving for more than 5 years after surgical resection and the rest succumbing to widespread metastases or local recurrence (3). Results of recent analyses indicate that systemic platinum-based adjuvant chemotherapy might substantially improve both disease-free survival and overall survival of NSCLC patients after surgical resection (4,5). However, the prognosis for individual patients with NSCLC cannot be determined accurately, even though determining prognosis is important if treatment is to be optimized and overtreatment avoided.
To identify new prognostic factors for NSCLC patients, various biomarkers, including serum and tissue parameters and genetic and epigenetic factors, have been investigated, but none has been adopted for use in general clinical practice (6). DNA microarray technology has had limited success in identifying gene expression profiles and biomarkers that are associated with survival in NSCLC patients (7–10). However, mRNA expression is often not tightly associated with the level of protein expression and cannot identify posttranslational modification of proteins. Although two-dimensional gel electrophoresis technology has also been used to search for novel protein biomarkers of human lung cancer (11–13), such technology is a technically challenging and time-intensive process that is probably not adaptable for the rapid throughput assays that are needed for clinical application. Treatment decisions thus continue to rely largely on information from conventional clinical and pathologic examinations. New technologies for protein profiling that are applicable to clinical practice need to be developed that have high accuracy, sensitivity, and throughput.
Mass spectrometry is a new technology that has allowed rapid progress in the field of proteomics. With mass spectrometry, protein expression profiles can be obtained from biologic materials and evaluated for molecular markers of human cancers. Such profiles, which can be obtained with surface-enhanced laser desorption–ionization mass spectrometry (14–16) and matrix-assisted laser desorption–ionization (MALDI) mass spectrometry (17), can distinguish serum of patients with various malignancies from that of normal healthy control subjects. In a previous study (18), we used MALDI mass spectrometry to distinguish between NSCLC tissue and normal lung tissue and between two groups of NSCLC patients with different median overall survivals, one of 6 months and the other of 33 months. However, that study was limited by a short postoperative follow-up (range = 1–42 months) that did not allow us to develop a model to assess long-term survival and had a small sample size (i.e., the total number of NSCLC patients = 66), which did not allow validation with an independent dataset. In this study, which was conducted at a different institution with a different group of NSCLC patients than those in our previous study, we used 201 specimens (174 of NSCLC tissue and 27 of normal lung tissue). We obtained protein expression profiles with MALDI mass spectrometry from these 201 specimens, separated into training and validation sets, to identify a molecular signature that was associated with relapse-free survival and overall survival.
|
Subjects and Methods |
|---|
|
TopAbstractContext and CaveatsSubjects and MethodsResultsDiscussionReferencesNotes |
|---|
Study Population
Specimens from 174 NSCLC patients (111 with adenocarcinoma, 36 with squamous cell carcinoma, 19 with large-cell carcinoma, and 8 with adenosquamous carcinoma) who had undergone resection with curative intent between December 1, 1995, and December 31, 1999, were obtained from the Department of Pathology and Molecular Diagnostics, Aichi Cancer Center, Nagoya, Japan, on the basis of tissue availability. The median follow-up for the patients who survived beyond 5 years without evidence of relapse (n = 75) was 89 months (range = 64–108 months). Two-thirds (n = 116) of the patients were randomly assigned to the training set, and one-third (n = 58) were randomly assigned to the testing set. One tumor specimen per patient was analyzed. Twenty-seven normal lung specimens were obtained from 27 NSCLC patients in this study (20 specimens from patients in the training cohort and seven specimens from those in the test cohort). Analysis of characteristics of the patients in both training and test cohorts (see Table 1) showed no statistically significant differences in clinicopathologic features between the cohorts. Normal lung tissues were resected from NSCLC patients in both the training and test cohorts. Primary tumor material in excess of that necessary for diagnostic purposes and normal lung tissue were obtained less than 15 minutes after removal from the patients and stored at –80 °C for about 10 years until use in this study. There was no thawing of any tissue specimens until they were used for this study. The requisite approval from the institutional review board and patients’ written informed consent had been obtained.
|
Proteomic Analysis
Sections (12-µm thick) from frozen lung tumor and normal lung tissue were cut, transferred to MALDI mass spectrometry sample plates (PE Biosystems, Foster City, CA), and dried at 4 °C for 1 hour in a desiccator. Adjacent sections were stained with Giemsa as a guide to determine which part of the tissue should be analyzed, and areas of normal and tumor tissues were precisely marked by a surgical pathologist (Y. Yatabe) at Aichi Cancer Center Hospital. We deposited 100-nL drops of an energy-absorbing matrix solution (saturated sinapinic acid in water/acetonitrile/trifluoroacetic acid, 500:500:1, by volume; this solution allows molecules to be protonated and desorbed from tissue surface) on as much of each tissue slice as possible. Heterogeneous areas with tumor and normal tissue components were avoided. Spectra were acquired over the tumor surface that was covered by each drop with a Voyager STR Instrument (Applied Biosystems, Foster City, CA), essentially as described previously (18,19) (Fig. 1). In this analysis, signals in the range of 2000–20000 m/z were considered to identify low–molecular-weight markers, which are difficult to analyze by two-dimensional gel electrophoresis technology. Each spectrum underwent smoothening to reduce the electrical noise and internal calibration with the peaks for the hemoglobin 
chain (molecular weight + one proton [M + H]+ = 15868.2) and the hemoglobin 
chain ([M + 2H]2+ = 7564.2) by use of Data Explorer software version 4.5 (Applied Biosystems) (18). Additional processes for peak detection and alignment to compare spectra obtained from different patients were conducted with MarkerView software version 1.0 (Applied Biosystems). Spectra were normalized by summing the ion counts across all peaks of each sample and choosing the largest sum from all samples as a global maximum. The sum of total ion counts of each spectrum was divided by the global maximum to obtain the normalization factor, which was then used to multiply each data point for each spectrum (18). Previous studies (17–21) have demonstrated the accuracy and reproducibility of MALDI mass spectrometry analysis of human materials (e.g., tumor tissues and serum). We confirmed the accuracy and reproducibility of our proteomic analysis by examining two spots, on which 100-nL drops of matrix solution were deposited, from the same tissue section in 20 randomly selected NSCLC specimens from the training cohort. The top 500 peaks of the two spectra from one tissue specimen showed strong correlations (average R2 = .99) across all 20 samples.
|
Identification of Individual Proteins in the Proteomic Signature
For four of the specimens that were used for MALDI mass spectrometry analysis, we also carried out detailed analysis of individual proteins. The frozen NSCLC specimens (350 mg) were immersed in 500 µL of extraction buffer (0.25 M sucrose, 0.01 M Tris–HCl at pH 7.36, and 0.1 mM phenylmethylsulfonyl fluoride), sonicated for three 30-second periods at 4 °C, and centrifuged three times as follows: once for 10 minutes at 680g, once for 10 minutes at 10000g, and once for 1 hour at 55000g. After each centrifugation, the soluble fraction was transferred to a new tube, and pellets were dissolved in 250 µL of an extraction buffer (0.3 M NaCl and 0.01 M Tris–HCl at pH 7.36 in water for pellet after centrifugation at 680g; 2 M NaCl, 5 M urea, 0.01 M Tris–HCl at pH 7.36 in water for pellet after centrifugation at 10000g; and 80% acetonitrile and 0.1% trifluoroacetic acid in water for pellet after centrifugation at 55000g). These extracts were analyzed by MALDI mass spectrometry to find those containing proteins corresponding to peaks in the signature, and then the selected extracts were separated over a polymeric column (Amersham Biosciences, Upsalaa, Sweden) with a high-performance liquid chromatography (HPLC) pump (Toso, Tokyo, Japan). Solvent A was 0.1% trifluoroacetic acid in water, and solvent B was 0.1% trifluoroacetic acid in acetonitrile. A flow rate of 1 mL/min was used for 5 minutes with a gradient from solvent A to 5% solvent B, for 75 minutes with a gradient from 5% solvent B to 60% solvent B, and then for 20 minutes with a gradient from 60% solvent B to 95% solvent B as a wash in which no protein was eluted. HPLC fractions were collected every minute for 80 minutes. Each fraction was lyophilized, reconstituted with 50% acetonitrile in water containing 0.1% trifluoroacetic acid, and analyzed by MALDI mass spectrometry to identify the HPLC fractions that contained proteins corresponding to the peaks in the signature with molecular weights that were selected by bioinformatic analysis as candidate molecular markers for the NSCLC and prognostic signature. The selected fractions were lyophilized, material was dissolved in a mixture of 10 µL of 0.4 M ammonium hydrogen carbonate and 5 µL of 45 mM dithiothreitol, and then 10 µL of 100 mM iodoacetamide was added. This mixture was incubated for 4 hours at 37 °C with 5 µL of 200 nM mass-grade trypsin (Promega, Madison, WI) to obtain peptides. The peptides were separated and sequenced by a microcapillary reversed-phase column (AMR, Tokyo, Japan) with an HPLC pump (AMR) that was coupled on-line to an ion-trap mass spectrometer (Thermo Electron, San Jose, CA). Tandem mass spectra were acquired in the data-dependent scanning mode, with one full mass spectrometry scan followed by one tandem mass spectrometric scan of the most abundant precursor ion. These spectra were compared with those in the human databases of the National Center for Biotechnology Information (nonredundant) by use of Mascot version 2.1.0 (Matrix Science Inc, Boston, MA). A minimum of two peptide matches and a positive association between the m/z values detected with MALDI mass spectrometry and the molecular weight of the intact protein (including posttranslational modifications) were required for protein identification.
Statistical Methods
Protein profiles obtained by MALDI mass spectrometry from NSCLC tumors and normal tissues were analyzed with the significance analysis of microarray (SAM) test (22) to investigate mass spectrometry signals that appeared to differentiate NSCLC tissues from normal lung tissues in the training set. The cut point for this analysis was a false discovery rate of less than 1%. The agglomerative hierarchical clustering algorithm was applied to investigate the pattern among the statistically significant discriminator proteins as well as the biologic status with Eisen's software (23).
Three distinct statistical methods, the Fisher's exact test, the Kruskal–Wallis test, and SAM, were used to select informative signals for discriminating patients with a relapse-related death in the first 5 years of follow-up (i.e., high-risk patients) from those who were alive with no symptoms of relapse after a median of 89 months of follow-up (i.e., low-risk patients) in the training cohort of 116 patients. The Fisher's exact test was used to detect informative mass spectrometry signals that were statistically significantly different between patient groups at low and high risk, when each peak was treated as either present or absent. The total number of present and absent calls for mass spectrometry peaks in each subgroup of patients (i.e., high- and low-risk groups) were calculated, and mass spectrometry peaks were selected as informative when nonrandom associations were found. The Kruskal–Wallis test was the other nonparametric statistical method used to detect informative mass spectrometry signals that were statistically significantly different between patient groups at low and high risk. All values of mass spectrometry signals were ranked from low to high, by disregarding the group to which each value belonged. The smallest number was assigned the rank of 1, and the largest number was assigned a rank of N, where N is the total number of values in all the groups. The rank values were then summed in each group, and the mass spectrometry peaks of interest were identified as those with statistically significant differences in the sums of the ranks between the subgroups (high- and low-risk groups). These two statistical tests were executed by use of the R statistical programming language (24,25). SAM, which is a widely used analytic method that is based on the t statistic with permutations, was also used to test each peak value as continuous data, by use of the program SAM version 1.21, developed by Tibshirani et al. (26). The cutoff points were a P value of less than .001 for the Fisher's exact test, a P value of less than .001 for the Kruskal–Wallis test, and a false discovery rate of less than 1% for the SAM test. Those mass spectrometry signals that met at least one of the three selection criteria were analyzed further.
To construct an individualized prognosis prediction classifier, we used a well-established technique for supervised classification, the weighted voting algorithm, in which each weight value is calculated as the signal-to-noise ratio (27). To obtain a generally applicable classifier without specifically overfitting it to the training cohort, we used a 10-fold cross-validation strategy with 10000 randomized iterations when we selected mass spectrometry peaks for constructing the prediction classifier (28,29).
The average number of misclassified patients for N-signal model was calculated as follows. With 10-fold cross-validation, the training set was divided into 10 nonoverlapping subsets of essentially equal sizes. Then, the weight (signal-to-noise metric) of each mass spectrometry signal was calculated by use of data from nine of these subsets, and at the same time, each mass spectrometry signal was ranked on the basis of the absolute value of the signal-to-noise metric. The top N-ranked mass spectrometry signals, by rank, were used to construct a prognosis classifier. An N-signal model was used on the remaining one subset, and the number of misclassified patients (Nerror N,P,M, where N is the number of signals used for classifier, P is the Pth permutation, and M is the Mth subset) was calculated. This process was repeated for each of the 10 subsets, and the 10-fold cross-validated learning error was calculated as the sum of the number of misclassified patients of each process.
 |
Finally, the 10000 times average number of misclassified patients for N-signal model was calculated as
 |
The top 25–ranked mass spectrometry peaks, with the accumulated rank in each cross-validation procedure, are presented in Supplementary Table 1 (available online).
The prognosis prediction classifier was assessed as follows. If we define patients with a high and a low risk of recurrence as class 0 and 1, respectively, the signal-to-noise statistic (Sx; x = gene x) is calculated as Sx = (µclass0 – µclass1/
class0 + class1), where µclass0 is the mean value and class0 is the standard deviation for that profile in all samples in class 0. We selected the top 25–ranked mass spectrometry profiles on the basis of the absolute values for Sx of each gene. A weighted voting classification algorithm was used to predict outcome with data from the mass spectrometry signals selected as described above, and the resulting outcome classifiers were tested by use of 10-fold cross-validation. In this scheme, the algorithm can also be used to find the decision boundaries between the class means as bx = (µclass0 + µclass1)/2 for each gene, in addition to computing Sx. To predict the class of a test sample 
, each profile x in the predictive mass spectrometry profile set has a vote (Vx) that is based on the expression in this sample (gx) and bx [i.e., Vx = Sx(gx – bx)] and the final vote for class 0 or 1 is sign (>xVx). The votes were summed to determine the winning class (Supplementary Fig. 1, available online).
Because it is possible that unintended biased resubstitution or partial cross-validation can result in an underestimate of the error rate after cross-validation, the performance of any class prediction rule is best assessed by applying the rule created by use of one dataset (the training set) to an independent dataset (the validation or test set) (30). The individualized prognosis prediction classifier constructed with the training dataset of 116 patients was then validated by use of a completely independent validation of 58 patients. Kaplan–Meier survival analyses and Cox proportional hazards model analyses (Stata version 7.0) were used to investigate the relationship between the proteomic signature and the overall survival and/or relapse-free survival in the validation set patients. Hazard ratios (HRs) of death and relapse are presented, respectively. The proportional hazards assumptions were verified graphically (as implemented in stphplot function in STATA package). The P values presented for the multivariable Cox model were obtained with a likelihood ratio test. All statistical tests were two-sided.
|
Results |
|---|
|
TopAbstractContext and CaveatsSubjects and MethodsResultsDiscussionReferencesNotes |
|---|
Protein Expression Profiling and the Training Set
We first obtained protein expression profiles for the 136 frozen human lung tissue specimens in the training set, including 116 specimens from NSCLC patients (Table 1) and 20 specimens of normal lung tissue from patients in the training set, by use of MALDI mass spectrometry. These expression profiles contained 2630 distinct proteomic signals. Representative spectra from a specimen with regions of normal human lung tissue and NSCLC tissue are shown in Fig. 1. Among these 2630 signals, we found 694 proteomic signals that were differentially expressed between NSCLC tissues and normal lung tissues with a false discovery rate of less than 1%. These 694 signals constituted the tissue classification signature.
To construct a prognosis signature, we compared proteomic patterns from the 46 patients in the training set who died within 5 years of surgical resection because of relapse (i.e., the group at high risk of recurrence) and from the 52 patients also from the training set who survived more than 5 years and showed no symptom of relapse during the entire follow-up period (i.e., the group at low risk of recurrence). Of the remaining 18 patients in the training set, 15 survived for more than 5 years but had some symptoms of relapse during follow-up, two died within 5 years because of pneumonia, and the recurrence status of one patient could not be ascertained. Therefore, these 18 patients were not included in the development of the proteomic signature. By use of our statistical selection criteria (i.e., signals met any one of the following criteria: P<.001, for the Fisher's exact test and the Kruskal–Wallis test, and false discovery rate < 1%, for SAM), 178 proteomic signals from training set tumor tissues were found to be statistically significantly differentially expressed between groups at high and low risk of recurrence. These proteomic signals were further ranked according to the signal-to-noise metric and used to construct the signature that could best distinguish patients at high risk of recurrence (i.e., those in the poor prognosis group) from those at low risk (i.e., those in the good prognosis group). The learning errors, to which increasing numbers of the differentially expressed proteomic signals were applied, were calculated by use of a 10-fold cross-validation analysis and repeated with newly assigned sets 10000 times. From this cross-validation analysis, 25 proteomic signals were used to construct a weighted voting prognosis signature that was associated with outcome for NSCLC patients with the highest accuracy (Supplementary Fig. 2 and Supplementary Table 1, available online).
Protein Expression Profiling and the Validation Set
To examine the robustness of the proteomic signatures obtained with MALDI mass spectrometry, we evaluated the association between the proteomic signatures and clinicopathologic variables with an independent test dataset that included 58 NSCLC tissue specimens and seven normal lung tissue specimens (Table 1). We selected two signatures in the training set, the tissue classification signature and the prognosis signature, and both were validated by use of the validation cohort. We determined first that the proteomic signature could distinguish NSCLC tissue from normal lung tissue in the validation set by use of agglomerative hierarchical clustering analysis. Clustering patterns of expression of proteomic signals obtained from MALDI mass spectrometry analysis reflected the strong association between proteomic profiles and tissue classification and could correctly distinguish all NSCLC tissue specimens from normal lung specimens (Supplementary Fig. 3, available online).
Next, we estimated the discriminatory power of the proteomic prognosis signature and compared it with that of prognostic factors currently in clinical use, including histologic examination and the tumor–node–metastasis classification. Results with the validation set indicated that the proteomic signature could distinguish between patients at high and low risks of recurrence better than either of these conventional prognostic factors (for overall survival, HR of death in the high-risk group compared with the low-risk group = 18.6, 95% confidence interval [CI] = 5.9 to 58.3, P<.001; and for relapse-free survival, HR of relapse in the high-risk group compared with the low-risk group = 6.4, 95% CI = 2.5 to 16.0, P<.001). Among patients in the group at high risk of recurrence, the overall survival rate after surgical resection was statistically significantly lower than that among patients in the group at low risk of recurrence (at 2 years, 50% versus 97%, respectively, difference = 47%, 95% CI = 26% to 69%, P<.001; and at 5 years, 14% versus 89%, respectively, difference = 75%, 95% CI of the difference = 58% to 93%, P<.001, by log-rank test) (Fig. 2, A). Overall survival was statistically significantly different between the two groups, with a median survival of 24 months for the 22 patients in the group at high risk of recurrence and a median survival that had not yet been reached by 108 months for the 36 patients in the group at low-risk of recurrence (P<.001 by log-rank test, Fig. 2, A). Among patients in the group at high risk of recurrence, relapse-free survival was statistically significantly lower than that among patients in the group at low risk of recurrence (at 2 years, 35% versus 86%, respectively, difference = 51%, 95% CI = 26% to 76%, P<.001; at 5 years, 5% versus 72%, respectively, difference = 67%, 95% CI = 48% to 86%, P<.001, by log-rank test; Fig. 2, B).
|
In a univariate Cox regression analysis using patients in the validation set, both the proteomic profile (P<.001 for both overall survival and relapse-free survival) and the pathologic disease stage (P = .003 and .001 for overall survival and relapse-free survival, respectively) were statistically significantly associated with clinical outcome among NSCLC patients in the validation set (Table 2). In a multivariable Cox proportional hazards model, the proteomic signature was the only prognostic factor that was statistically significantly associated with overall survival, whereas both proteomic profile (HR = 6.4, 95% CI = 2.5 to 16.0, P<.001) and pathologic disease stage (HR = 2.6, 95% CI = 1.1 to 6.2, P = .026) were statistically significantly associated with relapse-free survival, after adjusting for patient age, sex, smoking history, histology, and pathologic disease stage (Table 2).
|
All 22 patients with the high-risk proteomic signature died during the follow-up period. The relapse status of 20 of 22 patients could be ascertained. Among these 20 patients, 19 experienced relapse during the follow-up period (one patient was censored). Thus, results from the validation set support the robustness of the proteomic signature.
We further evaluated the discriminatory power of the proteomic signature by stratifying NSCLC patients in the validation set by disease stage (i.e., into an early-stage [stage I] group and a more advanced-stage [stage II–III] group). All eight stage I patients with a high-risk signature died during the follow-up period (survival rates in stage I patients with high-risk and low-risk signatures at 2 years, 63% alive versus 100% alive, respectively, difference = 37%, 95% CI = 4% to 71%, P = .014; and survival rates at 5 years, 13% alive versus 96% alive, respectively, difference = 84%, 95% CI = 60% to 100%, P<.001, by log-rank test; Fig. 3, A). Among seven of the eight stage I patients with a high-risk signature (the relapse status for one patient could not be ascertained), six patients (one patient was censored) experienced relapse during the follow-up period (percentage free of relapse in stage I patients with high-risk and low-risk signatures at 2 years, 33% alive versus 92% alive, respectively, difference = 59%, 95% CI = 19% to 98%, P =.002; and percentage free of relapse at 5 years, 17% alive versus 84% alive, respectively, difference = 67%, 95% CI = 34% to 100%, P<.001, by log-rank test; Fig. 3, B). Among stage II–III patients, overall survival and relapse-free survival were also statistically significantly different between the groups with high- and low-risk signatures. Among the 14 stage II–III patients with high-risk signatures, the median overall survival was 20 months from surgical resection (interquartile range = 7–49 months; P = .003, by log-rank test). Among 13 of the 14 stage II–III patients with high-risk signatures (the relapse status for one patient could not be ascertained), the median relapse-free survival was 22 months (interquartile range = 7–31 months; P = .046, by log-rank test). Among the 10 such patients with low-risk signatures, the median overall survival was not reached, and median relapse-free survival was 29 months (interquartile range = 11 months to value not achieved during experimental period) (Fig. 3, C and D).
|
In a multivariable Cox proportional hazards model, we also found that the proteomic signature was statistically significantly associated with prognosis (overall survival, HR = 61.1, 95% CI = 8.9 to 419.2, P<.001; and relapse-free survival, HR = 11.7, 95% CI = 3.1 to 44.8, P<.001) among patients with pathologic disease stage I lung cancer, who are considered to be the most curable.
We also investigated whether the stage provided more prognostic information than the proteomic signature. Characteristics of the 58 patients in the validation set (36 of whom had the low-risk signature and 22 of whom had the high-risk signature) are presented in Table 1. These groups had statistically significant differences in pathologic lymph node stage (pathologic lymph node stage 0, n = 26 versus 10; pathologic lymph node stage 1, n = 6 versus 3; pathologic lymph node stage 2–3, n = 4 versus 9 in low-risk versus high-risk groups, respectively; P = .041) and in pathologic disease stage (pathologic disease stage I, n = 26 versus 8; pathologic disease stage II, n = 4 versus 4; pathologic disease stage III, n = 6 versus 10 in low-risk versus high-risk groups, respectively; P = .019). Among NSCLC patients in the group at low risk of recurrence, overall survival was related to pathologic disease stage, but not statistically significantly so (P = .066, by log-rank test, Fig. 4, A). Among patients in high-risk group, no statistically significant difference was found in overall survival between the stage I and stage II–III patients (P = .800, by log-rank test, Fig. 4, B). Thus, even among NSCLC patients with stage I disease, the proteomic signature was statistically significantly associated with clinical outcome.
|
As an initial step toward elucidating the biologic mechanism of the association between the proteomic signature and relapse, we identified several proteins that correspond to the mass spectrometry signals in the proteomic signature from NSCLC tissues. Extracts from four surgically resected human NSCLC specimens (two from the group at low risk and two from the group at high risk) were fractionated by reversed-phase HPLC and analyzed by MALDI mass spectrometry to identify the HPLC fractions that contained proteins corresponding to peaks in the proteomic signature. These selected fractions were lyophilized, reconstituted with ammonium hydrogen carbonate, and digested with trypsin to obtain peptides. Each fraction was subjected to sequence analysis of tryptic peptides by use of an ion-trap mass spectrometry coupled with HPLC. By this method, we identified the following proteins as part of the proteomic signature: thymosin
4 (which can sequester cytoplasmic monomeric actin); proteins that may be specifically expressed at a higher level in NSCLC tissue than in normal lung tissue, including thymosin 10 ([M + H]+ = 4937.5), ribosomal protein L39 ([M + H]+ = 6276.5), ribosomal protein S30 ([M + H]+ = 6647.8), calcyclin ([M + H]+ = 10091.6), and histone H2A.2 ([M + H]+ = 14004.3) (Table 3); and proteins associated with a high risk of recurrence, including ribosomal protein L26-like 1 ([M + H]+ = 17257.2), acylphosphatase ([M + H]+ = 11172.8), and phosphoprotein enriched in astrocytes 15 ([M + H]+ = 17218.5) (Table 3).
|
|
Discussion |
|---|
|
TopAbstractContext and CaveatsSubjects and MethodsResultsDiscussionReferencesNotes |
|---|
In this study, we analyzed surgically resected human NSCLC specimens by protein expression profiling with MALDI mass spectrometry. By use of sophisticated bioinformatic analysis, we derived a 25-signal proteomic signature that was associated with a high risk of recurrence in a training dataset and showed that it was also associated with a high risk of recurrence in an independent validation dataset. These results support the robustness of the direct protein expression profiling of tissue specimens with MALDI mass spectrometry.
Proteomic techniques have been developed to analyze the proteome in serum, tumor tissues, and other clinical materials and to obtain information about which proteins are expressed, their expression levels, and whether the proteins are posttranslationally modified, and such analyses can provide useful information about tumor characteristics and cancer biology. However, some limitations of our proteomic study need to be addressed. First, mass spectrometry signals obtained from human materials with MALDI mass spectrometry can contain thousands of data points that are derived from real proteins expressed in tumor cells. However, these signals can be contaminated with electronic and chemical noise from the variability in instrumentation. Second, mass spectrometry signals that can be obtained are limited to relatively highly abundant proteins with a low molecular weight. Third, further steps for identification of proteins that consist of proteomic signatures are needed. Therefore, it is necessary to evaluate mass spectrometry signals carefully. Accordingly, we used sophisticated statistical methods combining the most differentially expressed proteins associated with the biologic variable of interest from many analytical methods. Thus, we have described a 25-signal proteomic signature that appears to distinguish NSCLC patients at high risk of recurrence from those at low risk. Interestingly, disease stage was not statistically significantly associated with prognosis in the multivariable analysis of overall survival, although it was statistically significantly associated with it in a univariate Cox regression analysis. This result indicates that the proteomic signature may include information about disease stage. Results of another study (18) that used tissue-based proteomic profiling with MALDI mass spectrometry found that a mass spectrometry profile could divide NSCLC patients into two groups, one at low risk for survival and the other at high risk. In our study, however, the follow-up was longer. We followed patients who survived for at least 5 years (median follow-up = 89 months; range = 64–108 months). We also obtained comprehensive, detailed clinicopathologic information, including the relapse status of 168 of a total of 174 patients. With these data, we constructed a model to estimate the long-term prognosis of NSCLC patients in our training set and obtain a proteomic signature associated with prognosis. We then tested this signature with validation dataset. By this method, we found that the strength of the association between the proteomic signature and overall survival in the validation set (HR = 18.6, 95% CI = 5.9 to 58.3, P <.001) was approximately that of the training dataset. Among patients with pathologic disease stage I lung cancer (who are considered to be the most curable), we also found that the proteomic signature was associated with prognosis (overall survival, HR = 61.1, 95% CI = 8.9 to 419.2, P<.001; and relapse-free survival, HR = 11.7, 95% CI = 3.1 to 44.8, P<.001). Thus, the 25-signal proteomic signature is, to our knowledge, the first proteomic signature described that is strongly associated with lung cancer prognosis.
Recent evidence (4,5) indicates that the poor prognosis of NSCLC patients after surgical resection may be improved by early adjuvant therapy intensification before relapse is diagnosed. Because 70%–80% of stage I patients will not relapse, such early-stage treatment leads to overtreatment. Thus, it is important to identify those stage I patients who are likely to relapse from those who are not, and our 25-signal proteomic signature appears to do that. Among stage I patients in the validation cohort, lower overall survival rates after operation was statistically significantly associated with the high-risk group, as defined by our 25-signal proteomic signature, than the low-risk group (63% alive versus 100% alive at 2 years and 13% alive versus 96% alive at 5 years). Thus, proteomic analysis focusing on the 25-signal proteomic signature appears to identify patients at high risk of relapse because of residual disease, which cannot be detected with current diagnostic approaches for NSCLC (e.g., computed tomography scanning, positron emission tomography scanning, and tissue-based diagnostic tests) and which would benefit from systemic adjuvant treatment with cytotoxic and molecularly targeted drugs and radiotherapy. Consequently, use of the proteomic signature to identify high-risk patients could reduce rates of both overtreatment and undertreatment and improve survival for NSCLC patients. Because MALDI mass spectrometry analysis can be applied to very small tissue specimens, it may also be useful for treatment decisions for patients with small stage I tumors, which are being found more frequently because of sensitive imaging techniques, such as helical computerized tomography scanning.
Although relapse is the principal cause of death in patients with any cancer, the underlying molecular mechanism leading to relapse are still poorly understood. In addition to our primary goal of exploration of proteomic profiles that can distinguish patients at high risk of recurrence from those at low risk, we set out to identify molecules constituting the 25-signal proteomic signature, which could provide a basis for improved diagnosis and treatment. By advanced proteomic analyses with HPLC fractionation followed by peptide sequencing with tandem mass spectrometry, we identified several signals that corresponded to various proteins involved in cell migration, cell death, the cell cycle, protein metabolism, and transcription. The higher expression of thymosin 4 and 10 in the tumors than in normal tissues is in accord with previous reports from different institutes (18,21,31) and appears to support the reproducibility and precision of MALDI mass spectrometry for comprehensive protein expression profiling of human tissue specimens. In NSCLC specimens, we also identified the overexpression of calcyclin, which may regulate cell cycle progression and cell differentiation and which has been found to be expressed in many different types of tumors, especially at the margins of invasive cancers (19,32). We also identified phosphoprotein enriched in astrocytes 15, which has been shown to be highly expressed in tumors and to inhibit apoptosis (33,34), and acylphosphatase, which regulates activity of the Na+/K(+)-ATPase pump and may be involved in metastasis (35). More than half of the molecules in the proteomic signature remain to be identified. Determination of their functional and biologic roles of the proteins represented in the proteomic signature in NSCLC tumorigenesis and progression may eventually lead to improved treatments for NSCLC.
|
NOTES |
|---|
|
TopAbstractContext and CaveatsSubjects and MethodsResultsDiscussionReferencesNotes |
|---|
Drs K. Yanagisawa and S. Tomida contributed equally to this article.
None of the authors of this study have a conflict of interest.
This work was supported in part by a Grant-in-Aid for Scientific Research on Priority Areas, a Grant-in-Aid for Scientific Research (B), and a Grant-in-Aid for Exploratory Research and Program for Improvement of Research Environment for Young Researchers from Special Coordination Funds for Promoting Science and Technology commissioned by the Ministry of Education, Culture, Sports, Science and Technology of Japan.
The sponsors of the study had no role in the design of the study, the collection of the data, the analysis and interpretation of the data, the decision to submit the manuscript for publication, and the writing of the manuscript.
|
REFERENCES |
|---|
|
TopAbstractContext and CaveatsSubjects and MethodsResultsDiscussionReferencesNotes |
|---|
(1) Statistics and Information Department, Minister's Secretariat. Vital statistics of Japan 2001. (2003) Vol. 3. Tokyo (Japan): Ministry of Health, Labor and Welfare. 384–411.
(2) Minna JD. Neoplasms of the lung. In: Harrison's principles of internal medicine—Braunwald E, Fauci AS, Isselbacher KJ, Kasper DL, Hauser SL, Longo DL, et al, eds. (2001) 15th ed. New York (NY): McGraw-Hill. 562–71.
(3) Giaccone G. Clinical impact of novel treatment strategies. Oncogene (2002) 21:6970–81.[CrossRef][ISI][Medline]
(4) Winton T, Livingston R, Johnson D, Rigas J, Johnston M, Butts C, et al. Vinorelbine plus cisplatin vs. observation in resected non-small-cell lung cancer. N Engl J Med (2005) 352:2589–97.
(5) Pisters KM, Le Chevalier T. Adjuvant chemotherapy in completely resected non-small-cell lung cancer. J Clin Oncol (2005) 23:3270–8.
(6) Takahashi T, Sidransky D. Biology of lung cancer. In: Textbook of respiratory medicine. 4th ed—Mason R, Broaddus V, Murray J, Nadel J, eds. (2005) Philadelphia (PA): Elsevier Science. 1311–27.
(7) Beer DG, Kardia SL, Huang CC, Giordano TJ, Levin AM, Misek DE, et al. Gene-expression profiles predict survival of patients with lung adenocarcinoma. Nat Med (2002) 8:816–24.[ISI][Medline]
(8) Bhattacharjee A, Richards WG, Staunton J, Li C, Monti S, Vasa P, et al. Classification of human lung carcinomas by mRNA expression profiling reveals distinct adenocarcinoma subclasses. Proc Natl Acad Sci U S A (2001) 98:13790–5.
(9) Garber ME, Troyanskaya OG, Schluens K, Petersen S, Thaesler Z, Pacyna-Gengelbach M, et al. Diversity of gene expression in adenocarcinoma of the lung. Proc Natl Acad Sci U S A (2001) 98:13784–9.
(10) Tomida S, Koshikawa K, Yatabe Y, Harano T, Ogura N, Mitsudomi T, et al. Gene expression-based, individualized outcome prediction for surgically treated lung cancer patients. Oncogene (2004) 23:5360–70.[CrossRef][ISI][Medline]
(11) Chen G, Gharib TG, Huang CC, Thomas DG, Shedden KA, Taylor JM, et al. Proteomic analysis of lung adenocarcinoma: identification of a highly expressed set of proteins in tumors. Clin Cancer Res (2002) 8:2298–305.
(12) Hanash S, Brichory F, Beer D. A proteomic approach to the identification of lung cancer markers. Dis Markers (2001) 17:295–300.[ISI][Medline]
(13) Oh JM, Brichory F, Puravs E, Kuick R, Wood C, Rouillard JM, et al. A database of protein expression in lung cancer. Proteomics (2001) 1:1303–19.[CrossRef][ISI][Medline]
(14) Petricoin EF, Ardekani AM, Hitt BA, Levine PJ, Fusaro VA, Steinberg SM, et al. Use of proteomic patterns in serum to identify ovarian cancer. Lancet (2002) 359:572–7.[CrossRef][ISI][Medline]
(15) Petricoin EF 3rd, Ornstein DK, Paweletz CP, Ardekani A, Hackett PS, Hitt BA, et al. Serum proteomic patterns for detection of prostate cancer. J Natl Cancer Inst (2002) 94:1576–8.
(16) Adam BL, Qu Y, Davis JW, Ward MD, Clements MA, Cazares LH, et al. Serum protein fingerprinting coupled with a pattern-matching algorithm distinguishes prostate cancer from benign prostate hyperplasia and healthy men. Cancer Res (2002) 62:3609–14.
(17) Sidransky D, Irizarry R, Califano JA, Li X, Ren H, Benoit N, et al. Serum protein MALDI profiling to distinguish upper aerodigestive tract cancer patients from control subjects. J Natl Cancer Inst (2003) 95:1711–7.
(18) Yanagisawa K, Shyr Y, Xu BJ, Massion PP, Larsen PH, White BC, et al. Proteomic patterns of tumour subsets in non-small-cell lung cancer. Lancet (2003) 362:433–9.[CrossRef][ISI][Medline]
(19) Schwartz SA, Weil RJ, Thompson RC, Shyr Y, Moore JH, Toms SA, et al. Proteomic-based prognosis of brain tumor patients using direct-tissue matrix-assisted laser desorption ionization mass spectrometry. Cancer Res (2005) 65:7674–81.
(20) Caprioli RM, Farmer TB, Gile J. Molecular imaging of biological samples: localization of peptides and proteins using MALDI-TOF MS. Anal Chem (1997) 69:4751–60.[Medline]
(21) Stoeckli M, Chaurand P, Hallahan DE, Caprioli RM. Imaging mass spectrometry: a new technology for the analysis of protein expression in mammalian tissues. Nat Med (2001) 7:493–6.[CrossRef][ISI][Medline]
(22) Tusher VG, Tibshirani R, Chu G. Significance analysis of microarrays applied to the ionizing radiation response. Proc Natl Acad Sci U S A (2001) 98:5116–21.
(23) Eisen MB, Spellman PT, Brown PO, Botstein D. Cluster analysis and display of genome-wide expression patterns. Proc Natl Acad Sci U S A (1998) 95:14863–8.
(24) Ihaka R, Gentleman RR. A language for data analysis and graphics. J Computat Graph Stat (1996) 5:299–314.[CrossRef]
(25) R statistical software. Available at: http://www.r-project.org/index.html.
(26) SAM software, version 1.21. Available at: http://www-stat.stanford.edu/
tibs/SAM. [Last accessed: May 2007.
(27) Golub TR, Slonim DK, Tamayo P, Huard C, Gaasenbeek M, Mesirov JP, et al. Molecular classification of cancer: class discovery and class prediction by gene expression monitoring. Science (1999) 286:531–7.
(28) Ambroise C, McLachlan GJ. Selection bias in gene extraction on the basis of microarray gene-expression data. Proc Natl Acad Sci U S A (2002) 99:6562–6.
(29) Molinaro AM, Simon R, Pfeiffer RM. Prediction error estimation: a comparison of resampling methods. Bioinformatics (2005) 21:3301–7.
(30) Simon R, Radmacher MD, Dobbin K, McShane LM. Pitfalls in the use of DNA microarray data for diagnostic and prognostic classification. J Natl Cancer Inst (2003) 95:14–8.
(31) Hall AK. Differential expression of thymosin genes in human tumors and in the developing human kidney. Int J Cancer (1991) 48:672–7.[ISI][Medline]
(32) Komatsu K, Kobune-Fujiwara Y, Andoh A, Ishiguro S, Hunai H, Suzuki N, et al. Increased expression of S100A6 at the invading fronts of the primary lesion and liver metastasis in patients with colorectal adenocarcinoma. Br J Cancer (2000) 83:769–74.[CrossRef][ISI][Medline]
(33) Xiao C, Yang BF, Asadi N, Beguinot F, Hao C. Tumor necrosis factor-related apoptosis-inducing ligand-induced death-inducing signaling complex and its modulation by c-FLIP and PED/PEA-15 in glioma cells. J Biol Chem (2002) 277:25020–5.
(34) Hao C, Beguinot F, Condorelli G, Trencia A, Van Meir EG, Yong VW, et al. Induction and intracellular regulation of tumor necrosis factor-related apoptosis-inducing ligand (TRAIL) mediated apotosis in human malignant glioma cells. Cancer Res (2001) 61:1162–70.
(35) Riley HD, Macnab J, Farrell TJ, Cohn K. The expression of acylphosphatase is associated with the metastatic phenotype in human colorectal tumors. Carcinogenesis (1997) 18:2453–5.
(36) Mountain CF. Revisions in the international system for staging lung cancer. Chest (1997) 111:1710–7.[ISI][Medline]
Manuscript received August 23, 2006; revised March 20, 2007; accepted April 24, 2007.
Related Articles in JNCI
CiteULike 
Connotea 
Del.icio.us What's this?
J Natl Cancer Inst 2007 99: 825.
J Natl Cancer Inst 2007 99: 825.
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||