Stratification of Intermediate-Risk Endometrial Cancer Patients into Groups at High Risk or Low Risk for Recurrence Based on Tumor Gene Expression Profiles

Endometrial carcinoma is the fourth most common cancer in women and the most common gynecologic malignancy (1). The American Cancer Society estimates that there will be >40,000 new cases and 6,800 deaths due to endometrial cancer in the United States in 2004 (1). The overall 5-year survival rate for all stages of endometrial cancer is >75% (2). This reflects the fact that >75% of endometrial cancers are confined to the uterine corpus (stage I) at the time of diagnosis, and these tumors are associated with an overall 5-year survival of >80% (2).

Initially, endometrial cancers were staged using the International Federation of Gynecology and Obstetrics (FIGO) 1971 clinical staging system. However, prospective surgical staging trials revealed that clinical staging underestimates the true extent of extrauterine disease in clinical stage I and II patients (3–5). These studies determined that depth of myometrial invasion, grade of the tumor, and cervical extension were poor prognostic features that were associated with an increased risk of extrauterine spread, nodal metastases, and recurrence (3–5). As a result of these findings, FIGO established a new surgical staging system for endometrial cancer in 1988 that included these prognostic factors.

Early-stage endometrial cancers with poor histopathologic prognostic features have been designated “intermediate risk” because they are associated with a 15% to 20% risk of recurrence and a reduced rate of surgical cure (6). Traditionally, these patients have been offered adjuvant radiotherapy, which decreases vaginal and pelvic recurrences from 6% to 15% (without therapy) to 2% to 4% (7, 8). However, the use of adjuvant radiotherapy has had no impact on the rate of distant metastases or overall survival (7–9). Because most women with endometrial cancer present with stage I and II disease and the majority do not recur, it would be of substantial benefit to better define those women who are likely to recur. This would prevent unnecessary use of adjuvant radiotherapy and eliminate its associated morbidity for the majority of women with early-stage endometrial cancer (10). In addition, those women with early-stage endometrial cancer identified as being at higher risk of recurrence could participate and possibly benefit from clinical trials using novel adjuvant therapies.

One approach to this problem involves the generation of genome-wide gene expression profiles of tumors, which have been shown to correlate with various clinicopathologic features as well as important prognostic subgroups. With respect to endometrial cancer, gene expression profiling using cDNA microarrays indicates that different histologic subtypes possess unique gene expression profiles (11). In addition, gene expression profiling of early-stage endometrial cancers supports the existence of two highly distinct molecular subgroups that segregate with relatively good correlation to tumor grade (12). In other human cancer types, such as B-cell lymphoma, melanoma, and lung and breast cancers, expression profiles identify high-risk patient populations that cannot be identified using traditional clinicopathologic criteria (13–16). The purpose of this study was to determine whether patients with early-stage, intermediate-risk endometrial cancer could be stratified into high-risk and low-risk subgroups for risk of recurrence based on tumor gene expression profiles.

Patients and Clinicopathologic Characteristics. This study was approved by the Institutional Review Board of the Memorial Sloan-Kettering Cancer Center. Primary tumor specimens from 75 patients with pathologically confirmed endometrial cancer from April 1996 to February 2002 were obtained from institutional tissue banks; all tumor specimens had been flash frozen and stored at −80°C. All cases had pathologic features consistent with intermediate-risk classification as defined by deep myometrial invasion, high-grade histology, cervical extension, or positive peritoneal washings. This classification included any grade 3 tumor limited to the uterine corpus (FIGO stages Ia-Ic), grade 2 tumors with any myometrial invasion (FIGO stages Ib and Ic), grade 1 tumors with >50% myometrial invasion (FIGO stage Ic), occult stage II, and tumors limited to the uterus with malignant peritoneal cytology (FIGO stage IIIa; ref. 17). Patients with stage IIIA disease defined as such by the presence of extrauterine spread to surrounding tissues are generally defined as “high risk” for recurrence (17) and were excluded from the study. The number of intermediate-risk tumors within each FIGO staging group is shown in Table 1. The majority of endometrial cancers analyzed were of endometrioid histology (85%). The mean age at time of diagnosis was 67 years. Of these women, 59% had comprehensive surgical staging that included para-aortic and pelvic lymph node dissection. The majority of women in this study received adjuvant radiotherapy (80%). Of the 75 intermediate-risk patients, 13 had recurred and 62 had not recurred with a median follow-up of 24 months. Of note, the patient population analyzed in this study shared no overlap with that analyzed in a previous study from our laboratory that identified two distinct subsets of endometrial cancers through an unsupervised analysis (12).

RNA Isolation, Probe Preparation, and Microarray Hybridization. Total RNA was isolated from tumor specimens using RNeasy columns (Qiagen, Valencia, CA), and all samples were treated on the column with RNase-free DNase. Quality of the RNA was ensured before labeling by analyzing 20 to 50 ng of each sample using the RNA 6000 NanoAssay and a Bioanalyzer 2100 (Agilent, Palo Alto, CA). Samples with a 28S/18S ribosomal peak ratio of 1.8 to 2.0 were considered suitable for labeling. For samples meeting this standard, total RNA (2 μg) was used for cDNA synthesis using oligo(dT) primer and the SuperScript Double-Stranded cDNA Synthesis kit (Invitrogen, Carlsbad, CA). Synthesis, linear amplification, and labeling of cRNA were accomplished by transcription in vitro using the MessageAmp RNA kit (Ambion, Austin, TX) and biotinylated nucleotides (Enzo Diagnostics, Farmingdale, NY). Labeled and fragmented cRNA (10 μg) were then hybridized to the Human Genome U133A GeneChip (Affymetrix, Santa Clara, CA), which contains 22,215 oligonucleotide-based probe sets, at 45°C for 16 hours. Automated washing and staining were done using the Affymetrix Fluidics Station 400 according to the manufacturer's protocols, and probe intensities were measured using the argon laser confocal GeneArray Scanner (Hewlett-Packard, Palo Alto, CA).

Statistical Analysis of Gene Expression Data. Raw expression data were analyzed using the Microarray Analysis 5.0 software (Affymetrix). Data were normalized to a target intensity of 500 to account for differences in global chip intensity, and expression values were then transformed using the logarithm base 2. Probe sets with very low average expression were eliminated because their expression measurements were not reliable. A threshold of 6 on the log scale was used for this purpose.

Gene expression was associated with survival using the univariate Cox proportional hazards regression model. Adjustment for multiple comparisons was done using the false discovery rate procedure of Benjamini and Hochberg (18), and a 5% false discovery rate cutoff was used. Other clinical variables were associated with time until recurrence using either the univariate Cox model or the log-rank test. A statistical technique was used using time until recurrence of disease as the end point. Specifically, the genes most significant for risk of recurrence were used to create a risk score (15). The risk score for each subject was a linear combination of the gene expression values for the top genes identified by the univariate Cox model weighed by their estimated regression coefficients from the modeling. If the risk score was high, assuming the gene expression values were predictive, the subject would be more likely to recur. The risk score was used to stratify subjects into “high-risk” or “low-risk” groups based on a cut point. The cut point chosen was the 80th percentile of the risk scores, which corresponded to the upper limit of the proportion of patients who were at risk of recurring in this population.

Leave-one-out cross-validation was used in determining the risk scores. Specifically, the gene selection by Cox modeling and the risk score weights were determined on all but one left-out sample and the risk score from this model was calculated for the left-out sample. This left-out risk score was compared with the risk scores for all the other samples and subjected to the 80th percentile threshold. This process was repeated once for every sample. Because we did not know a priori how many genes to choose for the analysis, we took a majority vote of the risk group assignments using the top 50, 75, and 100 genes. The time to recurrence curve was estimated for the high-risk and low-risk groups using the Kaplan-Meier method, and the difference in time to recurrence between the two groups was tested using the log-rank test. Clinicopathologic characteristics of the low-risk and high-risk groups were compared using the Fisher exact test or the Wilcoxon rank sum test, as appropriate.

Kaplan-Meier estimates of the proportion of women in different clinicopathologic groups who had not recurred within 2 years are summarized in Table 2. There was no association between recurrence and the clinicopathologic variables studied, such as patient's age at diagnosis, body mass index (BMI), histologic grade of tumor, histologic type, stage, depth of myometrial invasion, or whether the patients had received adjuvant radiotherapy. In addition, there was no significant difference between risk of recurrence and whether the patients had been comprehensively surgically staged. The estimates in Table 2 are for recurrence within 2 years because most endometrial cancers recur within this period, and there is very little information to be gained beyond 2 years in this patient group.

The genes most associated with time to recurrence are listed in Table 3. There was no single gene that was significantly correlated with recurrence after adjusting for multiple comparisons. However, a leave-one-out cross-validation approach was used to develop a risk score that stratified subjects into high-risk and low-risk for recurrence. The log-rank test showed that the high-risk and low-risk groups differed significantly in their time to recurrence (P = 0.04; Fig. 1). The estimated hazard ratio (95% confidence interval) was 3.07 (1.0-9.4). Ignoring the censoring, there were 13 patients identified as high risk, 5 who had recurred, and 62 patients identified as low-risk, 8 who had recurred.

Clinicopathologic characteristics of the endometrial cancer cases in low-risk and high-risk subgroups are summarized Table 4. There were no significant differences between low-risk and high-risk groups, determined by gene expression profiles, and clinicopathologic variables, such as patient's age at diagnosis, BMI, histologic grade of tumor, histologic type, FIGO stage, myometrial invasion, and whether patients received adjuvant radiotherapy. However, there was a significant difference in the proportion of women who had been comprehensively surgically staged in the low-risk and high-risk subgroups (P = 0.03). Of the 13 women who were identified as high-risk for recurrence, 9 (69%) were not surgically staged. This was in contrast to 22 of 62 (35%) women identified as low risk for recurrence who were not surgically staged. Because of this difference, the high-risk and low-risk groups were again compared using a log-rank test, but one that stratified for surgical staging. The P for this test was of borderline significance (P = 0.06). Notably, however, of the 13 patients who recurred, only 3 had recurrence involving pelvic or para-aortic lymph nodes, and 2 of these patients with nodal recurrences had been comprehensively surgically staged. The clinicopathologic characteristics of those patients in this study that recurred are listed in Table 5. There is no feature that distinguishes this group of patients from those patients who did not recur (or have not recurred).

These data on early-stage, intermediate-risk endometrial cancers indicate that gene expression profiling can stratify patients into low-risk and high-risk subgroups for risk of recurrence. There was not a “signature” gene list identified in this study that was significantly correlated with risk of recurrence. However, when gene expression data using the genes most associated with recurrence were used to create a risk score, women with early-stage endometrial cancer with poor prognostic histopathologic features identifying them as intermediate risk could be separated into subgroups for their risk of recurrence. The newly identified high-risk subgroup was at significantly increased risk of recurrence compared with the low-risk subgroup. The increased risk of recurrence for the high-risk group was independent of well-established poor clinicopathologic prognostic features. Known poor clinicopathologic prognostic features, such as high histologic grade, deep myometrial invasion, cervical extension, and nonendometrioid histology, were not correlated significantly with this molecular risk stratification. In our previous study that identified two highly distinct subclasses of endometrial carcinoma that clustered according to tumor grade (12), there was not significant overlap between the gene list associated with high-grade tumors in that study and the gene list associated with high risk for recurrence in the present study.

However, whether these women had comprehensive surgical staging, including pelvic and para-aortic lymph node dissection, did correlate significantly with risk stratification into low risk or high risk for recurrence. The majority of women in the low-risk group (65%) were comprehensively surgically staged compared with those in the high-risk group (31%). Based on surgical staging studies, patients with disease apparently confined to the uterus but with deep myometrial invasion, high grade or both may have a 25% to 35% risk of having pelvic lymph node metastases (4). Therefore, it is possible that a small group of patients who may have been upstaged (stage IIIc) had they undergone complete surgical staging has been identified by the gene expression profiles of their primary tumors and is captured in this high-risk group. However, surgical staging alone was not significantly correlated with increased risk of recurrence; therefore, those women who were not comprehensively surgically staged were not more likely to recur. More importantly, however, as summarized in RESULTS, only one patient with recurrence that was not surgically staged had a nodal recurrence; therefore, the difference in surgical staging between the two groups cannot account for the difference in recurrence rates between the two risk groups identified.

Alterations in single cancer–related genes, such as the ERBB2 oncogene and the TP53 tumor suppressor gene, have been associated with decreased progression-free and overall survival in endometrial cancer (19). For stage I patients who overexpress ERBB2 (HER-2/neu), there is a significant decrease in 5-year progression-free survival from 97% to 62% (20). In multivariate analysis, p53 protein overexpression is a significant independent prognostic factor for decreased overall and progression-free survival (21–24). However, the majority of endometrial cancers do not have specific genetic alterations when genes are assessed independently, and when present, most do not confer risk independent of well-recognized poor prognostic clinical and pathologic features. Genome-wide characterization may better define subtypes of endometrial cancer with different clinical outcomes, such as risk of recurrence, as was shown in this study.

Stratification of subjects into high-risk and low-risk did not correlate perfectly with recurrence and therefore was not diagnostic. This was not unexpected because the number of patients with an outcome event used to determine genes most associated with recurrence was small. This study was done to provide proof of principle that a genome-wide characterization method can potentially stratify patients for their risk of recurrence and possibly provide useful prognostic information. This approach warrants further study using a larger sample of patients who have recurred to establish a list of genes most associated with recurrence and possibly to help guide clinical decision-making for this clinically problematic group of patients with endometrial cancer.