Risk estimation of distant metastasis in node-negative, estrogen receptor-positive breast cancer patients using an RT-PCR based prognostic expression signature

We chose time from surgery to distant metastasis, also referred to as distant metastasis-free survival (DMFS), as the primary endpoint. Subjects were considered to have an event at the time of diagnosis of distant metastases or were censored at the earliest occurring date of contra-lateral recurrence, death without recurrence or last follow-up. The definition of DMFS endpoint, its events and censoring rules were aligned with those adopted by the National Surgical Adjuvant Breast and Bowel Project (NSABP) for the prognostic molecular marker studies [4]. We also analyzed the endpoint of overall survival (OS), which was defined as time from surgery to death from any cause.

Five 10 μm sections of each paraffin block were used for RNA extraction. A macrodissection on the samples was performed to isolate RNA from the cancer cell areas which had been marked by a pathologist on a guide slide. Total RNA was extracted from the FFPE tissue sections using a modified commercially available isolation kit (Zymo Research, Orange, CA). Briefly, the FFPE section slides were deparaffinized in xylene, washed consecutively with 100%, 90%, and 70% ethanol, air dried at room temperature and the tissues transferred to a tube. Following digestion with proteinase K for 18 to 24 hours at 55°C, the samples were spun down and the supernatants transferred to new tubes. A mixture of 100% ethanol and extraction buffer was added to the supernatant and loaded onto Zymo-Spin II Columns. The columns were treated with a series of washes that include a DNase treatment step. Total RNA was eluted with TE buffer that was heated to 65°C.

The quality of the RNA extracted from formalin fixed tissues varies and depends on a variety of factors, including the fixation process used, age of the samples, and storage conditions. Additionally, the use of formaldehyde can cause extensive cross linking of tissue components. In many cases, only a small fraction of the recovered RNA can be amplified by RT-PCR. To determine the amount of amplifiable RNA, we quantified the expression level of an endogenous gene, NUP214, in each sample by comparing it to a serially diluted Universal Human Reference RNA standard (Stratagene, La Jolla, CA). Approximately 0.5 ng of amplifiable RNA was used to profile each gene.

We selected 197 candidate genes (see Additional file 2) from the published literature that include the prognosis genes reported by van't Veer et al [1] and Dai et al [14], the gene signature for response in tamoxifen treated women reported by Paik et al (8), and the ER status genes reported by West et al [15]. In addition, three endogenous "housekeeping" genes (NUP214, PPIG, and SLU7) were included and used to normalize expression levels of the other genes (see Additional file 3).

The amount of PCR-amplifiable RNA from the training set was insufficient to profile all 197 genes. Consequently, RNAs from the training set were enriched by pre-amplification with the MessageAmpII aRNA amplification kit (Ambion, Austin, TX) whereas RNAs in the validation sets were used directly. To assess the effect of the enrichment, we profiled and compared the14 genes using 50 paired enriched and unenriched samples from the training set. Good correlation (R² = 0.9931) was observed between the metastasis scores generated with the enriched vs. unenriched samples (see Additional file 4).

A single-step RT-PCR with SYBR^® Green was used for gene expression profiling essentially as previously described [18]. The assays were performed on the Prism 7900 Real-Time PCR system using the following thermocycling parameters: 50°C for 2 minutes; 95°C for 1 minute; 60°C for 30 minutes; 95°C for 15 seconds and 60°C for 30 seconds for 42 cycles. A Universal Human Reference RNA control was amplified with the appropriate candidate gene for each run. All assays were performed in duplicate. PCR primers were designed to amplify all known splice-variants, and the size of the PCR product was designed to be shorter than 150 bp to accommodate degraded RNA in archived FFPE samples.

The relative changes in gene expression were calculated by the ΔΔCt method [19]. The expression of each of the genes was first normalized to three endogenous control (HSK) genes then further normalized to a calibrator, reference RNA pool (Universal human reference RNA, Stratagene, La Jolla, CA). The ΔΔCt values of 197 genes which gave acceptable expression levels were used for statistical analyses. In the validation set, we only profiled 3 normalization genes and the 14 cancer-related genes that were selected in the training set for the prognostic signature, and ESR1 gene for ER status determination.

After developing an ER mRNA cutoff for estrogen receptor status on separate samples (Iverson et al, J Mol Diagn, in press) and demonstrating a high concordance with IHC determination in the training and validation sample sets, we chose to use ER expression as criteria for ER status for consistency between sample sets (see Additional file 1).

The MIB-1 monoclonal antibody to Ki-67 (AMAC, Inc, Westbrook, ME) was used at 1:200 dilution in PBS. Following standard preparation of slides, staining was visualized using biotinylated anti-mouse (Vector Laboratories) and Strepavidin-horseradish peroxidase (Zymed Laboratories); DAB was used as the chromogen, and hematoxylin as the counterstain. The invasive cancer on a slide was reviewed for immunoreactivity. The slide was first scanned using a 10× objective, and regions with high labeling chosen for counting at high power (40×). The Ki-67 LI was calculated as the fraction of positively stained nuclei in at least 1000 invasive cancer cells from multiple high power fields. Tumors above the median labeling index were categorized as high Ki-67; those below the median labeling index, as low Ki-67.

For gene selection in the training set, the expression levels of each gene were standardized to have mean zero and variance of one, and missing values were imputed using a nearest neighbor algorithm [20]. We used the semi-supervised principal component (SPC) method [21, 22] available in the PAM software package [23] with a Cox regression model [24] for time to distant metastasis to each gene. The genes were ranked by their univariate Cox scores. The first principal component of the genes that reached a certain threshold of the univariate Cox score was computed and applied in a Cox model with the principal component as a single variable. Internal cross-validation was used to determine the optimal threshold to select genes to optimize the Cox score with the principal component of the expression of the selected genes. With this procedure, 37 genes were selected from the training set. The number of genes included in the prognosticator was further reduced by a regression of the supervised principal component on the expression values of the 37 genes (see Additional file 5) while imposing a constraint on the size of the regression coefficients. This procedure, known as the Lasso [25], resulted in a linear combination of the expression values of 14 genes (the remaining gene coefficients were effectively shrunk to zero) that provided a good approximation of the supervised principal component. For simplification, since the regression coefficients of the 14 genes were of similar magnitude (see Additional file 6), a summary score was calculated as the sum of the 14 ΔΔCT measurements for each subject, and since lower values of the score were associated with higher probability of metastasis, the final metastasis score (MS) for each subject is defined as the negative of the summary score. The creation of the MS is described in the Additional file 7.

Differences in patient characteristics were assessed with the Wilcoxon rank-sum or Kruskal-Wallis test for continuous or ordinal measures and with the chi-square test for discrete measures. Cox proportional hazards models were used to estimate hazard ratios and Wald tests of the coefficients from these models were used to assess statistical significance of the variables. The MS was modeled both as a continuous variable as well as in discrete groups of high (≥ -23.5) and low (< -23.5) risk, the cut-point of which was determined as the median MS in the training set. Other covariates included in the multivariable Cox models included years of age (at surgery), tumor size (cm), histologic grade of tumor. HER2 status had also been ascertained on the validation set using the FDA approved scoring method. These data were used as a covariate for an additional multivariable model in the validation set. Estimates of distant metastasis free and overall survival for the high and low MS groups were calculated with the method of Kaplan and Meier [26] and confidence intervals for point estimates of survival were calculated using the complementary log-log transformation [27]. The probability of distant metastasis in 5 years and 10 years for individual patients was calculated from the survivor function as estimated by an accelerated failure time model including the continuous MS as the independent variable and assuming the event times have a Weibull distribution [28]. Statistical analysis was performed with PAM [23], SAS software version 9.1 [29] and R software version 2.4.1 [30].

Time dependent receiver operator characteristic (ROC) curves and area under the curve (AUC) to predict distant metastases within 5 years and 10 years and death within 10 years were estimated using the method described by Heagerty [31] with nearest neighbor estimation of the bivariate distribution of time and continuous MS [32]. Sensitivity and specificity were estimated for each time of interest using the cut-point defining high and low MS risk groups. Approximate 95% confidence intervals for the various diagnostic summary measures were calculated based on the standard errors estimated from 500 bootstrap samples [33]. Ten year risk estimates for relapse and mortality based on the Adjuvant! Online calculator [34] were obtained for the patients in the untreated validation set and subsequently used to plot the 10-year time dependent ROC curves for visual comparison with the ROC curves based upon the MS.

In order to obtain representative community-based untreated samples we chose a N- and ER + cohort design from similar hospital settings from US and UK. The characteristics of the subjects included in the training and validation sets are presented in Table 1. The patients in the training set were older and had a higher proportion low grade and small tumors than the patients in the validation set. In addition, subjects in the untreated validation set were followed for a longer period of time (median = 15.6 years) than those in the training set (median = 8.7 years) and the treated validation set (median = 6.3 years).

To assess potential selection bias in the training set we compared the selected 142 patients to 67 patients who met the entry criteria but were not included because of missing clinical information. No significant differences were found between the included and excluded subjects with respect to age (median = 64.0 and 63.7 respectively; p = 0.51) or tumor size (median = 1.28 cm and 1.23 cm respectively; p = 0.39). However, tumor grade tended to be higher among the included subjects than the excluded subjects (47% and 32% with ≥ Grade 2 tumors respectively; p = 0.02) (see Additional file 8). Similarly, in the untreated validation set we compared the selected 279 patients to 118 patients who met the entry criteria but were not included due to insufficient RNA (n = 112) or missing RNA expression values (n = 6) for one or more of the 14 signature genes. No significant differences between included and excluded patients were found with respect to age (median = 56.4 and 55.6 respectively, p = 0.47), or tumor grade (71% and 78% with Grade 2 tumors respectively, p = 0.18) (see Additional file 8). Although median tumor size was 2 cm for both included and excluded patients, a test comparing the distributions of ordered ranks for the two groups indicated a significant difference (p = 0.02). Upon further inspection, we found a somewhat larger proportion of patients with tumor size greater than 2 cm among the included patients than among the excluded patients (40% vs. 31% respectively).

The median MS in the training set was -23.5 and ranged from -72.8 to 16.8. The MS performed well in the training set which is expected given that the outcome of subjects in this sample set was used to select the genes. Each unit increase in the MS was associated with a 4.2% increase (p = 0.0001) in risk of distant metastasis and a 2.4% increase (p = 0.003) in risk of death. The hazard ratio (HR), after adjusting for age, tumor size and tumor grade, comparing subjects above the median MS to subjects below the median MS was 3.2 (95% CI 1.27 to 7.87; p-value = 0.014) for risk of distant metastasis (Table 2) and 2.0 (95% CI 1.05 to 3.83; p-value = 0.036) for risk of death (Table 3). However, since these estimates of risk are likely biased due to the selection of genes from this same training set we next tested the MS in an independent validation set.

The MS was also associated with DMFS (p < 0.0001) and overall survival (p = 0.0004) when tested in the validation sample set from Guy's Hospital. Each unit increase of the MS resulted in an estimated 4.9% (95% CI 1.03 to 1.07) increase in risk of DMFS and a 2.4% (95% CI 1.01 to 1.04) increase in risk of death. Applying the identical cutpoint (< -23.5 vs. ≥ -23.5) as used in the training set resulted in 107 patients designated as low risk and 172 as high risk. In univariate analysis, the HR for the high risk group compared to the low risk group was 4.71 (95% CI 2.33 to 9.51, p < 0.0001) for DMFS (Table 2) and 2.26 (95% CI 1.51 to 3.38, p < 0.0001) for overall survival (Table 3). After adjustment for age, tumor size and tumor grade in Cox multivariate analysis the HR for the high versus low risk group were 4.02 (95% CI 1.91 to 8.44, p = 0.0002) and 1.97 (95% CI 1.28 to 3.04, p = 0.002) for DMFS (Table 2) and overall survival (Table 3), respectively. Among subjects in the high risk group, the estimated rates of DMFS were 83% (95% CI 76 to 88%) at five and 72% (95% CI 64 to 78%) at ten years of follow-up as compared to rates of 98% (95% CI 93 to 100%) and 96% (95% CI 90 to 99%) at five and ten years respectively among subjects in the low risk group (Figure 1). Rates of overall survival were 91% (95% CI 85 to 94%) at five years and 68% (95 CI 61% to 75%) at ten years among high risk subjects compared to 98% (95% CI 93 to 100%) and 91% (95% CI 84 to 95%) at five and ten years respectively among low risk subjects (Figure 1). The HR for the high versus low risk group did not differ significantly (p > 0.05) by age group (≤ 55, > 55) or by tumor grade (1, 2 or 3) for either the DMFS or overall survival endpoints. The HR for the endpoint of overall survival was 3.52 (95% CI 1.83 to 6.75) in subjects with tumor diameters of 2 cm or less compared to 1.23 (95% CI 0.70 to 2.17) in subjects with tumor diameters larger than 2 cm and this difference in hazard ratios was significant (p-value for interaction = 0.012). A similar trend, though not significant (p-value for interaction = 0.14), was seen for the endpoint of DMFS where the HR was 7.59 (95% CI 2.07 to 27.8) in subjects with small tumors (≤ 2 cm) and 2.53 (95% CI 1.05 to 6.12) in subjects with larger tumors (> 2 cm). The risk estimates of the high versus low risk group remained significant (p = 0.0002 for DMFS; p = 0.002 for overall survival) and were essentially unchanged (HR = 4.16 for DMFS; HR = 2.05 for overall survival) in additional multivariate models that included HER2 expression as a covariate although large changes would be unexpected due to the small number of HER2 positive subjects (N = 19).

While the threshold for determining the high and low risk groups based on the median value of the MS in the training set resulted in significant differences in DMFS estimates for these groups in the validation set, the risk for individual patients may be of a more continuous nature. Figure 2 shows the estimated probabilities of DMFS at 5 and 10 years based on a parametric regression model of the event times among the validation set subjects. Using this model, the 5 and 10 year probabilities of DMFS are 94.7% and 90.2% respectively for subjects with MS scores equal to the 25^th percentile in the validation set (MS = -27.4), 91.7% and 84.8% respectively for subjects with MS score equal to the median (MS = -17.3) and 88.0% and 78.5% respectively for subjects with MS score equal to the 75^th percentile (MS = -9.1).

The sensitivity and specificity of the MS high and low risk groups to predict distant metastases were 96% (95% CI 89 to 100%) and 43% (95% CI 37 to 49%) respectively at 5 years and 93% (95% CI 87 to 100%) and 46% (95% CI 39 to 53%) respectively at 10 years. Sensitivity and specificity of the MS risk groups to predict death from any cause at 10 years were 84% (95% CI 75 to 94%) and 45% (95% CI 39 to 52%) respectively.

ROC curves of continuous MS to predict distant metastasis within 5 years and 10 years had AUC (95% CI) of 0.74 (0.66 to 0.81), and 0.71 (0.65 to 0.78). A ROC curve to predict death in 10 years had an AUC of 0.69 (0.62 to 0.77). Hence, MS are predictive of both distant metastases and deaths. Visual comparison of the MS and Adjuvant! Online ROC curves (Figure 3) for distant metastasis and overall survival at 10 years indicate the MS may provide additional diagnostic value, particularly for the distant metastasis endpoint.

In the study of 45 tamoxifen-treated patients (see Additional file 9 for clinical and pathological characteristics of the patients), each unit increase in the MS resulted in an estimated 4.9 percent increase in the hazard for DMFS although this association did not reach statistical significance (p = 0.085). Sixteen patients were classified as high-risk and 29 were classified as low-risk by MS using the cut point of -23.5. In univariate analysis, the HR for DMFS comparing high risk to low risk subjects was 3.61 (95% CI 0.86 to 15.1, p = 0.079) (Table 2). In multivariate analysis the HR was 3.50 (95% CI 0.58 to 21.2, p = 0.172). Five-year distant-metastasis-free survival rates were 73% (95% CI 40 to 89%) in the high risk group and 92% (95% CI 72 to 98%) in the low risk group (Figure 1).

Using the median as the cutpoint for the Ki-67 LI, univariate analysis indicated this proliferation marker is significantly associated with DMFS (HR = 2.4, (95% CI 1.12 to 5.11, p = 0.02) and overall survival (HR = 1.71, (95% CI 1.00 to 2.93, p = 0.05) but is no longer significant after adjustment for MS in a multivariate analysis. The HR for Ki-67 were reduced to 1.55 (95%CI 0.69 to 3.45, p = 0.29) and 1.32 (95% CI 0.75 to 2.33, p = 0.33) for DMFS and overall survival respectively after adjustment for MS while the MS retained significance when adjusted for Ki-67 LI; the HR for the high versus low risk group were 3.64 (95% CI 1.49 to 8.90, p = 0.005) and 2.21 (95% CI 1.23 to 3.97, p = 0.008) for DMFS and overall survival respectively (Table 4). Moreover, the correlation coefficient, R², between Ki-67 LI and MS was 0.30. We did not find a strong correlation (R² = 0.18) between Li-67 LI and the mRNA level of this gene (see Additional file 10).

A list of the 14 genes and overlaps with previously reported signatures is shown in Table 5. Subsets of genes of the previously reported signatures appear in the 14-gene signature: 9 genes overlap with the cell proliferation signature identified by Dai et al [14]; 2 genes overlap with the 21-gene panel by Paik et al [4], and 6 genes overlap with 70-gene panel by van't Veer et al [1]. The gene function analysis revealed that BUB1, CCNB1, MYBL2, PKMYT1, RACGAP1, CENPA, ORC6L, TK1, DIAPH3, RFC4, MELK, and RFC4 are associated with one or more functions of cell cycle control, apoptosis or DNA recombination and repair. UBE2S is a gene involved in ubiquitination. There are two genes, PRR11 and DC13, whose functions are not known. Pathway analyses revealed that the 14 genes in our prognostic signature are involved in a variety of biological functions, but a majority of genes are involved with cell proliferation. Ten of the 14 genes are associated with the TP53 pathways which have been found to be coordinately over-expressed in tumors of poor outcome.