- © 2004 by American Society of Clinical Oncology
Identification of Novel Prognosticators of Outcome in Squamous Cell Carcinoma of the Head and Neck
- Volkert B. Wreesmann,
- Weiji Shi,
- Howard T. Thaler,
- Ashok Poluri,
- Dennis H. Kraus,
- David Pfister,
- Ashok R. Shaha,
- Jatin P. Shah,
- Pulivarthi H. Rao and
- Bhuvanesh Singh
- From the Laboratory of Epithelial Cancer Biology, Head and Neck Service, Departments of Surgery, Epidemiology and Biostatistics, and Medicine, Memorial Sloan-Kettering Cancer Center, New York, NY; and Department of Pediatrics, Baylor College of Medicine, Houston, TX
- Address reprint requests to Bhuvanesh Singh, MD, Laboratory of Epithelial Cancer Biology, Head and Neck Service, Memorial Sloan-Kettering Cancer Center, 1275 York Ave, New York, NY 10021; e-mail: singhb{at}mskcc.org
Abstract
Purpose The goal of this study was to identify chromosomal aberrations associated with poor outcome in patients with head and neck squamous cell carcinoma (HNSCC).
Patients and Methods We assessed the global genomic composition of 82 HNSCCs from previously untreated patients with comparative genomic hybridization (CGH). The CGH data were subcategorized into individual cytogenetic bands. Only genomic aberrations occurring in more than 5% of cases were analyzed, and redundancies were eliminated. Each aberration was submitted to univariate analysis to assess its relationship with disease-specific survival (DSS). We used Monte Carlo simulations (MCS) to adjust P values for the log-rank approximate χ2 statistics for each abnormality and further applied the Hochberg-Benjamini procedure to adjust the P values for multiple testing of the large number of abnormalities. We then submitted abnormalities whose univariate tests resulted in an adjusted P value of less than .15 together with significant demographic/clinical variables to stepwise Cox proportional hazards regression. We again verified and adjusted P values for the χ2 approximation of the final model by MCS.
Results CGH analysis revealed a recurrent pattern of chromosomal aberrations typical for HNSCC. Univariate analysis revealed 38 abnormalities that were correlated with DSS. After controlling for multiple comparisons and confounding effects of stage, five chromosomal aberrations were significantly associated with outcome, including amplification at 11q13, gain of 12q24, and losses at 5q11, 6q14, and 21q11 (MCS adjusted P = .0009 to P = .01).
Conclusion HNSCC contains a complex pattern of chromosomal aberrations. A sequential approach to control for multiple comparisons and effect of confounding variables allows the identification of clinically relevant aberrations. The significance of each individual abnormality merits further consideration.
INTRODUCTION
Head and neck squamous cell carcinoma (HNSCC) can arise anywhere along the upper aerodigestive tract and is etiologically associated with tobacco and alcohol exposure.1 Overall, HNSCC portend a grave clinical course, with 5-year survival rates less than 50%.2 Selection of optimal treatment for individual patients requires accurate assessment of malignant potential at presentation. Traditionally, prognostic stratification of HNSCC has relied on analysis of clinicopathologic variables, which are inconsistent predictors of clinical behavior.3 Because HNSCC results from a sequential accumulation of genetic abnormalities, recent focus has shifted toward prognostication based on molecular profiling.4,5 The advent of global genomic screening based on comparative genomic hybridization (CGH) has contributed significantly to the molecular characterization of HNSCC.6-9
CGH analysis of HNSCC has identified gains of chromosomal regions 1q21, 3q26.3, 5p15, 7p12, 8q24, 9q34, 11q13, and 20q12 and deletions of 3p, 4q, 5q, 7q22, 8p23, 9p21, 13q12-24, 17p, 18q21, 21q11-21, and 22q.10 Studies investigating the prognostic utility of individual alterations have suggested that gains of 3q26 and 11q13 and deletions of 8p23 and 22q may be valuable markers of aggressive disease.11,12 However, the density of the information resulting from CGH analysis has limited attempts at defining individual abnormalities associated with outcome. The simultaneous analysis of multiple genetic and clinicopathologic factors is limited by several factors, especially issues pertaining to multiplicity of testing, with a significant possibility for inaccurate correlations.13-15 In the present study, we used a novel statistical approach to simultaneously assess the prognostic utility of CGH-detected DNA copy number changes in a large panel of HNSCC while controlling for multiple correlations. This study identified gains of 11q13 and 12q24 and deletions of 5q11-15, 6q14-21 and 21q11-21 as independent prognostic factors in patients with HNSCC. The approach presented here is applicable to the analysis of large-scale molecular-cytogenetic profiling of cancer in general.
PATIENTS AND METHODS
Tissue Acquisition and Clinical Information
Randomly selected tumor samples from patients undergoing curative treatment for HNSCC were obtained at the time of biopsy or surgical resection in the operating room following guidelines established by the institutional review board. Only previously untreated cases with biopsy-proven, primary squamous cell carcinomas involving the oral cavity or laryngopharyngeal complex were included in the study. Clinical information was collected independent of the laboratory data from medical records, and all identifier data were terminally coded to maintain patient anonymity. The median follow-up time was 22.5 months overall and 29 months (range, 2 to 249 months) for survivors. Table 1 lists the clinical information characterizing these cases.
Distributions of Patient Demographic and Clinical Characteristics
Comparative Genomic Hybridization
All tumor specimens were examined histologically and confirmed to contain greater than 70% tumor. Test DNA was extracted from tumor specimens and reference DNA was extracted from normal placenta from a healthy donor as described previously. CGH was performed as described previously.6,10,16 In brief, tumor DNA from each case was labeled with fluorescein-12-dUTP (NEN Dupont, Boston, MA) and reference DNA from normal human placenta was labeled with Texas red 5-dUTP using nick translation (Gibco BRL, Rockville, MD). After competitive hybridization of equal amounts (2 μg) of test and reference DNA to metaphase spreads from normal lymphocytes, seven to 10 separate metaphases were captured and processed using the Quantitative Image Processing System (Quips Pathvysion system; Applied Imaging, Santa Clara, CA). Red, green, and blue fluorescence intensities were analyzed for all metaphase spreads, normalized to a standard length, and statistically combined to show the red to green signal ratio and 95% CIs for the entire chromosome. Copy number changes were detected based on the variance of the red to green ratio profile from the standard of 1. Ratio values of 1.2 and 2.0 were defined as thresholds for gains and amplifications, respectively, and losses were defined as ratio of 0.8 or less.
Prestatistical Data Analysis
CGH analysis detects an array of DNA copy number abnormalities that may vary in size and chromosomal location between samples.10,17 For example, gains of the chromosomal arm 3q may involve different parts of the chromosomal arm, some of which may or may not overlap between patient samples and thus are difficult to compare between samples. To objectify the comparison between samples, we divided the genome into 303 regions according to the cytogenetic regions as detailed in the 400-band International System of Cytogenetic Nomenclature,18 excluding the p arms of acrocentric chromosomes, telomeric regions, and centromeric regions because of proven irreproducible CGH ratio deviations.19 Thus we analyzed the CGH data per cytogenetic band. Because this approach leads to redundancy of abnormalities (one cytogenetic event—for example, isochromosome formation of chromosome 3—is represented as 14 separate genetic events [ie, deletion of eight cytogenetic bands at 3p and gain of eight cytogenetic bands at 3q]), we selected a single representative of each set of neighboring abnormal cytogenetic bands that co-occurred in a sample to represent the entire abnormality. In all cases, the cytogenetic locus identified by this approach was congruent with minimal common region of abnormality for the respective DNA copy number aberration. Because genomic instability is a key feature of many cancers, it is expected that multiple abnormalities may be present in the cancer genome purely as a result of instability and not as a result of clonal accumulation, resulting in a Darwinistic selective advantage.20 Because these abnormalities are unlikely to influence tumor behavior and thus outcome, they represent noise that can potentially confound statistical analysis. In our analysis, we chose to target the issue pertaining to noise rigorously by excluding all abnormalities that occurred in less than 5% of cases (four cases).
Statistical Analysis
Identification of significant prognostic factors among a large number of potential markers represents a major challenge in multiplicity of statistical testing on correlated genetic abnormalities in the current study. Univariate significance tests for CGH genomic abnormalities are expected to yield some nominal P values that are very small by chance alone. Also, there were multiple sets of two or more chromosomal abnormalities that were essentially indistinguishable in their patterns of occurrence in the sample, especially at closely linked loci. Furthermore, the accuracy of the asymptotic χ2 approximation for the log-rank statistic for testing the association of disease-specific survival (DSS) with genetic abnormalities is limited by the proportion of patients with noncensored DSS data (20 of 82 patients) and by the proportions of patients having a particular abnormality (range, four to 11 of 82 patients for nominally significant abnormalities at the final model).
In our analysis, we performed a sequential series of steps to control for issues present in the analysis of CGH data. First, we eliminated rare chromosomal abnormalities found in less than 5% of the sample because any abnormality with low prevalence is likely not to be useful as a prognostic factor. Second, because redundant data do not provide additional prognostic information, we selected a single representative of each set of closely linked bands whose abnormalities co-occurred across patients in the data set (corresponding to the minimal common region of involvement). Then, we submitted the reduced list of chromosomal abnormalities to univariate analysis to assess the strength of relationship with DSS time using the log-rank test. We adjusted the P values by doing a large number of Monte Carlo simulations because the χ2 approximation may not be accurate enough for very small P values when abnormalities occur in only a small number of patients (eg, four to 11 patients). We performed more than 100,000 random permutations of DSS data for each possible number of occurrences of an abnormality among the 82 CGH profiles. We estimated the P value for each abnormality as the proportion of simulations for which an abnormality that occurred in a number of patients yielded a log-rank χ2 statistic at least as large as what was observed in the actual data for that abnormality.
Further, we used Hochberg-Benjamini's13 false discovery rate (FDR) controlling procedure to adjust for multiple testing of the large number of abnormalities when selecting a subset of the most significant abnormalities to submit to stepwise Cox proportional hazards regression, adjusting for significant demographic/clinical variables. The FDR multiplies the i-th smallest P value by a factor m/i, where m is the total number of significant tests. The FDR procedure starts with the largest nominal P value and stops at the first adjusted P value. Hochherg and Benjamini's approach is more powerful than the very conservative Bonferroni method. All abnormalities whose univariate tests resulted in Hochberg-Benjamini's procedure adjusted P value of less than .15 were considered in the stepwise Cox proportional hazards regression model. We again adjusted the P values from the χ2 approximate significant tests the significance of coefficients in the final regression model by doing 10,000 Monte Carlo permutation simulations of the Cox regression.
RESULTS
CGH Analysis
Application of CGH to the analysis of the genomic composition of 82 cases of HNSCC demonstrated a consistent pattern of DNA copy number changes (Fig 1). Gain of genomic loci most commonly involved 3q26-27, 5p14-15.3, 7p12-22, 8q24, 9q34, 11q13, 17q25, 19, and 22q13. Deletions of genomic material most commonly involved 3p, 4q31, 5q11-15, 9p13-21, 11q14-25, 13q31, and 18q21-23. In addition, recurrent high-level amplifications were identified at 2q32, 3q26, 4p15.3 to 16, 5p15, 7q11.2-p12, 7q21, 8p11, 8q24, 9p22-24, 11q13, 12p13, 18p, and 19p.
Ideogram showing DNA copy number changes identified by comparative genomic hybridization analysis of 82 cases of head and neck squamous cell carcinoma. Thin vertical lines on either side of the ideogram indicate losses (left) and gains (right) of the chromosomal region. The chromosomal regions of the high-level amplification are shown by thick lines (right).
Prestatistical Analysis
We extended the assessment of DNA copy number changes across our samples of HNSCC by analyzing each cytogenetic band individually. This analysis showed a total of 595 DNA copy number abnormalities (gains, losses, amplifications) of cytogenetic bands (273 losses, 253 gains/amplifications, and 69 amplifications) that were found in at least one patient's sample (Table 2). However, because this approach is associated with the introduction of redundancy that may influence the statistical power (see Patients and Methods), we chose to select one abnormality among a set of closely linked (neighboring) DNA copy number changes of chromosomal bands (minimal common region). In addition, to reduce noise and improve statistical power, we excluded all abnormalities that occurred in less than 5% of cases. After these modifications, we found 207 copy number abnormalities in at least four patients, including 104 losses, 95 gains/amplifications, and eight amplifications (Table 2).
Reduction of Comparative Genomic Hybridization Profiles to 303 Chromosomal Loci
Outcome Analysis
Thirty-eight abnormalities were associated with DSS at nominal P < .05 (Fig 2). These included five at P < .0001 (gain at 12q24 and 14q11; losses at 5q11, 6q14, and 21q11). Gain at 8p11 and amplification of 11q13 were among seven nominally significant at .0001 < P < .001. All abnormalities significant at P < .05 by the log-rank statistic χ2 approximation remained at P < .05 in the Monte Carlo permutation simulations. Twenty abnormalities were significant at P < .015 in the Monte Carlo simulations and remained significant at P < .15 after the Hochberg-Benjamini adjustment for multiple testing. The identified prognostic abnormalities and tumor-node-metastasis system stage (which was the only univariate prognostic demographic/clinical characteristics associated with DSS) were then submitted to stepwise Cox proportional hazards regression (Table 3). Five (12q24, 5q11, 6q14, 21q11, and 11q13) were jointly significant at P < .05 when adjusted for tumor-node-metastasis system stage in the final regression equation (Table 4). CIs for regression coefficients were not adjusted for multiple comparisons. We did further Monte Carlo simulations that adjusted the joint significance levels based on the χ2 statistics for the regression coefficients (Table 4).
Correlation of individual chromosomal loci of (A) amplification, (B) gains/amplifications, and (C) losses with disease-specific survival. Each plot shows chromosomes 1 to 22, with loci on p arms represented in green and q arms in red.
Univariate Survival Statistics for Top 25 Abnormalities Identified by Comparative Genomic Hybridization (P < .01)
Results From Stepwise Cox Proportional Hazards Regression Model
No significant differences in clinicopathologic features were observed based on the presence or absence of one or more of the identified prognostic abnormalities (Table 5), except sex, which was not a prognostic factor of DSS. Of the 23 patients who had one or more prognostic genetic abnormalities, 87% (20 of 23 patients) were male (P = .009). To graphically show the effect of the five selected genetic abnormalities, we divided patients post hoc into three cohorts: stage I to III patients with no abnormality, stage IV patients with no abnormality, and patients with one or more of the five prognostic chromosomal abnormalities. The survival plots for these three are shown in Fig 3. Our data confirm association between 11q13 overrepresentation and disease-specific survival of HNSCC. Moreover, prognostic significance of 12q24 gain and deletions of 5q11-15, 6q14-21, and 21q11-21 has not been previously described and may aid the research of treatment planning of HNSCC.
Disease-specific survival based on tumor-node-metastasis stage and genetic predictors of outcome in 82 patients with head and neck squamous cell carcinoma (HNSCC).
Distributions of Patient Demographic and Clinical Characteristics With or Without Any of Comparative Genomic Hybridization
DISCUSSION
HNSCC development and progression is driven by an accumulation of genetic events.21 The advent of global genetic screening methodologies, such as CGH, has allowed an improved understanding of the genomic profile of HNSCC, identifying recurrent gains, losses, and amplifications on virtually all chromosomal arms.10 Although the overall genomic profile of HNSCC is highly consistent, individual cases differ significantly in the types of accumulated genetic abnormalities, likely reflecting the heterogeneous clinical phenotype of HNSCC. Accordingly, previous studies have associated individual genomic aberrations with clinicopathologic factors, including lymphatic metastasis, tumor grade, and presence or absence of etiologic factors, including tobacco, alcohol, and human papillomavirus exposure.22-25 Despite these efforts, few studies have assessed the global DNA copy number profile of HNSCC in relationship to its clinical outcome. To date, only one group has associated CGH-detected DNA copy number abnormalities with clinical outcome in a large number of patients with HNSCC.11 This study correlated gains of 3q26 and 11q13 and deletion of 8p21-22 with survival outcomes. However, because of limitations in the statistical approach, especially the lack of correction for multiple testing, the results are difficult to interpret. In this study, we have applied a rigorous statistical approach to simultaneously assess the impact of multiple CGH-detected events on clinical outcome of a large panel of randomly ascertained HNSCCs.
Application of CGH to the genetic analysis of 82 HNSCCs confirmed previous studies showing the presence of a recurrent pattern of DNA copy number changes in these tumors. To objectify the comparison of DNA copy number abnormalities that may vary in size and location between individual tumor cases, we chose to analyze the genomic constitution of our cases for each cytogenetic band. Subsequently, we reduced redundancy introduced by this approach by selecting a single locus representative (minimal common region) of closely linked (neighboring) abnormalities for outcome analyses. In addition, we attempted to reduce noise by excluding abnormalities present in less than 5% of cases, because these were unlikely to have an impact on outcome.
We evaluated 207 individual copy number abnormalities for univariate association with outcome, revealing 38 abnormalities that correlated with outcome at P < .05. Given the large number of genomic abnormalities tested, several of the loci linked with outcome likely represent false-positive results.13 We performed Monte Carlo permutation simulations to adjust for inaccuracies of the χ2 approximations in small samples. To overcome the issues relating to multiplicity of testing, we used Hochberg-Benjamini's FDR procedure to adjust for 207 simultaneous univariate survival analyses. This analysis revealed that five of 38 abnormalities remained significant at a nominal P < .0001 (Table 3). Subsequently, we introduced 20 identified abnormalities (Monte Carlo permutation simulations adjusted P value < .015 and Hochberg-Benjamini's FDR adjusted P value < 0.15) into a Cox proportional hazards model to control for potential dependent relationships with clinicopathologic predictors included in the tumor-node-metastasis staging system (Table 4). It is clear that in the accurate appraisal of large-scale CGH data, several possible statistical approaches are possible.15,17 For example, Jain et al17 recently divided the CGH data into a profile comprising 1,225 channels and used Kendall's Tau permutation testing to assess correlations between genomic abnormalities and clinical outcome of 52 breast carcinomas. Application of our statistical model to the survival analysis in 82 cases of HNSSC revealed DNA copy number instability involving chromosomal regions 11q13, 12q24, 5q11-15, 6q14-21, and 21q11-21 to be associated with poor outcome independent of the tumor-node-metastasis system stage. Although tumor-node-metastasis system stage remained significant in multivariate analysis, the genetic predictors identified in this study were associated with strikingly higher risk ratios (Fig 3). This was well represented in the descriptive Kaplan-Meier DSS curves demonstrating that patients of any stage that featured one of these abnormalities did significantly worse than patients with none of these abnormalities regardless of tumor-node-metastasis system stage. From this analysis, it is clear the behavior of HNSCC is not universally predicted by tumor-node-metastasis system staging and that genomic predictors may augment predictive accuracy. All 11 patients of the 23 having prognostic aberrations identified by CGH and adequate follow-up died of disease within 2.5 years of primary treatment, independent of the tumor-node-metastasis system stage. These findings suggest that genetic predictors may significantly enhance prognostic stratification of HNSCC.
The association between 11q13 overrepresentation and outcome supports the validity of the statistical approach presented here. 11q13 amplification is strongly linked with poor clinical outcome of squamous cell carcinomas of the upper aerodigestive tract.11,26,27 Cyclin D1, a cell cycle regulator whose overexpression affects cellular survival, is a candidate oncogene within 11q13 locus.28 Antisense cyclin D1 treatment inhibits HNSCC cell proliferation, induces apoptosis, and leads to tumor shrinkage, supporting its role in HNSCC pathogenesis.29 Additionally, recent studies suggest that EMS1 and TAOS1 may drive selection for 11q13 amplification.30,31
Similar to 11q13 amplification, 12q24 amplification and deletions of 5q11-15, 6q14-21 and 21q11-21 have been described in squamous cell carcinomas arising in various organs including the head and neck, anogenital tract, and lung.32,33 Recently, Fitzsimmons et al34 reported evidence for the presence of an unknown mortality gene driving selection for 6q14 deletion in advanced tongue carcinomas. However, prognostic significance of the abnormalities has not been assessed. Moreover, gene targets driving selection for these abnormalities remain ill-defined. It is of significance that our previous work using gene array analysis identified several genes within the identified regions whose dysregulation is correlated with outcome of HNSCC.35 These genes include DHFR (5q11-13), MAP3K7 (6q14), EST H75853 (6q14), SCAD (12q24), IFNAR1 (21q22), and NCAM2 (21q21). The role of these genes in HNSCC requires further delineation.
Bockmuhl et al11,36 suggested 3q26.3 amplification and 8p23 deletion as independent prognostic markers of HNSCC. In the present study, these abnormalities were commonly detected, but no outcome associations were identified. To assess the role of 3q26.3 copy number in more detail, we defined precise 3q26.3 copy numbers in an independent panel of 50 HNSCCs by fluorescence in situ hybridization using a novel probe.37 Although outcome was linearly associated with 3q26.3 copy number, only high copy numbers of 3q26.3 were independently associated with survival. Because 3q amplification seems to occur exclusively in patients with a history of tobacco/alcohol use,25 the findings of our study may have been diluted by the inclusion of a significant number of nonsmokers. Indeed, high-level 3q amplification was a significant predictor of outcome in the subset of smokers with HNSCC.
With regard to the generalizability of our findings, several limitations should be taken into account. First, although the present study includes one of the largest cohorts of HNSCC that have been assessed by CGH to date, it is clear that our findings are based on a relatively small number of cases. Second, it needs to be taken into account that our study population was skewed toward cases involving the laryngopharyngeal complex as a result of specimen availability. Nonetheless, this report represents the first study on the impact of CGH-detected DNA copy number changes in HNSCC analyzed using a rigorous statistical approach. The statistical approach outlined here may prove valuable in the analysis of multiple biomedical factors in cancer, such as microarray-based analyses. In addition to 11q13 amplification, we propose amplification of 12q24 and deletions of 5q11-15, 6q14-21, and 21q11-21 as predictors of HNSCC outcome. Akin to 11q13 amplification, these alterations may also have prognostic value in squamous cell carcinomas arising in other anatomic loci, and merit further study.
Authors' Disclosures of Potential Conflicts of Interest
The authors indicated no potential conflicts of interest.
Footnotes
-
Supported in part by the Young Investigator Award from the American Society of Clinical Oncology and the George H.A. Clowes, Jr, MD, FACS, Memorial Research Career Development Award from the American College of Surgeons.
V.B.W. and W.S. contributed equally to this work.
Authors' disclosures of potential conflicts of interest are found at the end of this article.
- Received January 14, 2004.
- Accepted June 30, 2004.
