Molecular risk assessment of BIG 1-98 participants by expression profiling using RNA from archival tissue

The RNA was isolated from 4 sections (25 μm) of FF material and from 10 paraffin sections (10 μm thick) as described previously [27]. After demodification, the RNA was bound to silica-based columns, DNase I digested and eluted with water. The protocols and reagents for RNA isolation from FF and FFPE tissues were recently incorporated in commercial protocols (RNAready and FFPE RNAready, AmpTec, Hamburg, Germany). RNA qualities were assessed on an Agilent 2100 Bioanalyzer (Agilent Technologies, Inc., Santa Clara, CA, USA). RNA prepared from FF material had a RIN>6 (RNA integrity number), the RIN of RNA from FFPE was 2-3. The percentage of tumor cells in each FFPE block was evaluated on stained tissue sections. From 437 available FFPE samples 43 samples (9.8%) with less than ~30% tumor cells, 10 ER-negative tumor samples and 7 samples (1.6%) with less than 1.5 μg total RNA recovery were excluded from further analysis. Approximately 30% of the sections contained 30-50% tumor cells, and about 60% contained 50-100% tumor cells. Each of the remaining RNAs was tested by quantitative reverse transcription PCR (qRT-PCR) with 3 control genes (GUSB, RPLP0 and UBB). The mean of the three raw Cts (cycle thresholds) was determined. In 35 samples (8%) the mean Ct was >31, indicating poor quality of the RNA. These RNAs were excluded from further analyses. For the remaining 342 RNAs (78.3%), the expression of 34 genes (see Table 1) was measured by qRT-PCR on TaqMan Low Density Arrays (TLDAs) (Applied Biosystems, Foster City, CA, USA) using a one step protocol (Invitrogen, Basel, Switzerland) on an Applied Biosystems 7900HT instrument. Technical replicates were performed for several intact and several partially degraded RNAs from FF and FFPE material, respectively. They revealed Pearson correlation coefficients higher than 0.95 for all 34 assays.

Genes with high correlation to the expression of ER, PgR, Her2 and proliferation related genes were prospectively selected from publicly available microarray data [28]. A complete list of microarray data sets used in the meta-analysis is available at ".http://​breast-cancer-research.​com/​content/​10/​4/​R65/​table/​T1[28] (Additional File 1, Table S1). The scores were defined by giving equal weight to each gene in the four groups (proliferation, estrogen response, progesterone response, Her2 response). Thus, a training set was not used as the scores were based on in silico gene selection.

Raw Ct values were normalized against the mean expression of GUSB, RPLP0 and UBB. Scores for ER (ER_8), PgR (PGR_5), Her2 (HER2_2) and proliferation (PRO_10) were defined as mean expression of all genes in each category (Table 1). A RISK score comprising 25 genes was calculated as follows: RISK_25 = PRO_10+HER2_2-(8 × ER_8+5 × PGR_5)/13. For comparison, ER_4 and PRO_5 scores were calculated based on 4 and 5 genes described previously [27]. The genes corresponding to ER_4 and PRO_5 scores corresponded to the genes used for calculating the recurrence score (RS) [29].

Histopathological data of BIG 1-98 samples were derived from a central review, with the exception of the grade which was locally assessed. The ER and PgR status were dichotomized into positive (≥ 10% immunoreactive cells) or negative (<10%) [30]. Her2 was measured by fluorescence in-situ hybridization or immunohistochemistry (IHC) and tumors were classified according to Rasmussen et al. [31]. The Ki-67 labeling index (LI) was centrally assessed by IHC as described and classified into low or high using the median LI (11%) as cut-off [32]. The same assays and cut-offs were used for the 82 matched samples with the exception of Her2 which was measured using the CB11 monoclonal antibody and using a cut-off of ≥ 50% [33]. Continuous molecular scores were compared to binary IHC parameters using the area under the curve (AUC). The 95% confidence intervals (CI) were estimated by a bootstrap method (100 bootstraps). Two-sided Mann-Whitney tests were used to assess the association between clinicopathological factors and scores.

Primary endpoint of survival analyses was DFS as defined previously [25]. Forty-five events were observed in 342 patients with a median follow-up time (estimated by reverse Kaplan-Meier [34]) of 60 months. DFS was estimated by Kaplan Meier analysis. Patients were classified into low and high PRO or RISK scores using the corresponding median score as cut-off. The differences in survival experience between the two resulting groups were assessed with log rank tests. Univariate and multivariate Cox proportional hazard models were used [35] and hazard ratios (HR), CIs and p-values were obtained. The multivariate models were assessed using the log-likelihood and the deviance of residuals. Likelihood ratio tests (LRT) were used to compare different nested multivariate models. No adjustments were made for multiple testing. Univariate Cox proportional hazard models were applied to estimate the rate of events and to produce corresponding plots.

Gene expression was measured from 34 genes using TLDAs with RNA isolated from FF and FFPE material of 82 breast cancers. These data were used solely for the assessment of the expression profiling from FFPE material. Pearson correlation coefficients between FF and FFPE expression values for each tumor and all assays ranged from 0.91 to 0.98. The mean increase of raw Ct values derived of FFPE compared to matched FF tissues was 1.30 units. This Ct shift was mostly compensated by normalization (Additional File 2, Figure S1. and Additional File 3, Figure S2).

Unsupervised hierarchical clustering demonstrated the stability of gene clusters and revealed an excellent agreement between FF- and FFPE-based expression profiles (Additional File 4, Figure S3). Molecular scores were determined for ER, PGR, HER2 and PRO. A linear relationship of scores was found for RNA from FF and RNA from FFPE material (Figure 1). Pearson correlation coefficients for the four scores were 0.968, 0.974, 0.942 and 0.944, respectively. The distributions of ER_8, PGR_5 and HER2_2 scores are shown as histograms together with the fitted mixture of two Gaussian distributions (Additional File 1, Figure S4) used for discriminating the subtypes.

The agreement between molecular scores and corresponding binary IHC variables was assessed by receiver operating characteristic (ROC) curves and AUC. AUCs and 95% CI were calculated for ER_8 (FF = 0.940 (0.835-1.00), FFPE = 0.931 (0.804-1.00)), PGR_5 (FF = 0.919 (0.828-0.986), FFPE = 0.916 (0.806-0.987) and HER2_2 (FF = 0.961 (0.895-1.00), FFPE = 0.963 (0.915-0.993)). PRO_10 was compared with IHC data for Ki-67 using a cut-off of 11% and the resulting AUCs were 0.798 (0.609-0.900) for FF and 0.810 (0.660-0.907) for FFPE, respectively. In conclusion, the agreement of the IHC with FFPE samples was as good as with FF samples.

Molecular scoring was applied to an independent set of tissue samples from Swiss patients participating in the BIG 1-98 randomized clinical trial and scores were compared to centrally assessed histopathological data by ROC curves. From a total of 437 provided tumor samples 342 ER-positive tumors (78.3%) were suitable for analysis. The AUC was 0.974 (95% CI = 0.946-0.995) for HER2_2 and 0.847 (95% CI = 0.794-0.902) for PGR_5. PRO_10 scores positively correlated with Ki-67 LI (Pearson correlation coefficient 0.51); the AUC was 0.815 (95% CI = 0.768-0.864) for Ki-67 binarized at 11% [32].

The histological grade was assessed according to Elston and Ellis [36]. The PRO_10 score positively correlated with Elston and Ellis scores and with grade (Pearson correlation coefficient 0.453 and 0.409, respectively) (Figure 2). Furthermore, PRO_10 scores were significantly higher in Her2 positive tumors, in tumors larger than 2 cm and in tumors with axillary lymph node metastasis as compared to Her2 negative tumors, T1 tumors and N0 tumors (p ≤ 0.0015, Mann-Whitney tests), respectively (data not shown).

The prognostic values of PRO_10 and RISK_25 scores were assessed by their ability to assign patients to low and high risk groups. Patients were stratified according to histological grade and low or high PRO_10 and RISK_25 scores using the corresponding medians as cut-offs (Figure 3). As expected, patients with grade III tumors had poorer DFS than patients with grade I or grade II tumors (p = 0.0019, panel A). High PRO_10 scores correlated with poorer DFS compared to low scores in all (p = 0.0043, panel B) and in histological grade II tumors (p = 0.0024, panel C). Similarly, RISK_25 discriminated between favorable and poor DFS in all (p = 0.0005, panel D) and in node positive tumors (p = 0.0009, panel E). Univariate Cox proportional hazards regression analysis confirmed these observations.

The PGR_5, PRO_10 and RISK_25 scores were all significant predictors of DFS (p < 0.05) as were histological grade, tumor size, number of positive lymph nodes and Ki-67 LI (Table 2). The PRO_5 score was also a significant predictor of DFS but PRO_10 score was numerically better than PRO_5 in terms of log-likelihood (L) and deviance of residuals (D) (PRO_10: L = -223.35, D = 225.83; PRO_5: L = -224.16, D = 227.57).

Figure 4 shows the estimated rate of recurrence as a function of PRO_10, PGR_5 and RISK_25 scores. The PRO_5, PRO_10 and the RISK_25 scores remained significant predictors of DFS when applied to patients with grade II breast cancer.

The impact of the molecular scores PRO_10 and PGR_5 was further documented in multivariate models comprising clinicopathologic predictors and molecular scores that were significant in univariate analyses.

Multivariate analyses revealed that PRO_10 is a predictor of DFS independent of tumor size (T), number of positive lymph nodes (N), grade (G) and Ki-67 LI. PRO_10 represents proliferation-related genes and it was of interest to compare it to Ki-67. Table 2 shows the results of multivariate analyses including T, N, G and either Ki-67 (model 1) or PRO_10 (model 3) in comparison with a model containing both markers (model 2). The full model (model 2) was significantly better than model 1 (LRT p = 0.0071). No significant difference was found for PRO_10 between models 2 and 3 (LRT p = 0.8075). Thus, adding PRO_10 to T, N, G and Ki-67 significantly improved the model. In contrast, adding Ki-67 to T, N, G and PRO_10 did not bring additional information.

The same procedure was used to evaluate whether PGR_5 further improved model 6 containing T, N, G and PRO_10 (Table 2). The full model including all 5 variables (model 5) performed better than model 4 (T, N, G, PGR_5; LRT p = 0.0089) and model 6 (T, N, G, PRO_10; LRT p = 0.0339). Both, PGR_5 and PRO_10 remained significant in model 5 suggesting that the two scores contain independent information with respect to prognosis and outcome.