BreastPRS is a gene expression assay that stratifies intermediaterisk Oncotype DX patients into high- or low-risk for disease recurrence

Breast Cancer Res Treat (2013) 139:705 715 DOI 10.1007/s10549-013-2604-0 PRECLINICAL STUDY BreastPRS is a gene expression assay that stratifies intermediaterisk Oncotype DX patients into high- or low-risk for disease recurrence Timothy M. D Alfonso Ryan K. van Laar Linda T. Vahdat Wasay Hussain Rachel Flinchum Nathan Brown Linda Saint John Sandra J. Shin Received: 2 May 2013 / Accepted: 8 June 2013 / Published online: 18 June 2013 Ó Springer Science+Business Media New York 2013 Abstract Molecular prognostic assays, such as Oncotype DX, are increasingly incorporated into the management of patients with invasive breast carcinoma. BreastPRS is a new molecular assay developed and validated from a metaanalysis of publically available genomic datasets. We applied the assay to matched fresh-frozen (FF) and formalinfixed paraffin-embedded (FFPE) tumor samples to translate the assay to FFPE. A linear relationship of the BreastPRS prognostic score was observed between tissue preservation formats. BreastPRS recurrence scores were compared with Oncotype DX recurrence scores from 246 patients with invasive breast carcinoma and known Oncotype DX results. Using this series, a 120-gene Oncotype DX approximation algorithm was trained to predict Oncotype DX risk groups and then applied to series of untreated, node-negative, estrogen receptor (ER)-positive patients from previously published studies with known clinical outcomes. Correlation of recurrence score and risk group between Oncotype DX and BreastPRS was statistically significant (P \ 0.0001). 59 Electronic supplementary material The online version of this article (doi:10.1007/s10549-013-2604-0) contains supplementary material, which is available to authorized users. T. M. D Alfonso (&) S. J. Shin Department of Pathology and Laboratory Medicine, New York-Presbyterian Hospital, Weill Cornell Medical College, 525 East 68th Street, Starr 1031E, New York, NY 10065, USA e-mail: tid9007@med.cornell.edu R. K. van Laar W. Hussain R. Flinchum N. Brown L. S. John Signal Genetics, LLC, New York, NY, USA L. T. Vahdat Department of Hematology/Oncology, Weill Cornell Medical College, New York, NY, USA of 260 (23 %) patients from four previously published studies were classified as intermediate-risk when the 120-gene Oncotype DX approximation algorithm was applied. BreastPRS reclassified the 59 patients into binary risk groups (high- vs. low-risk). 23 (39 %) patients were classified as low-risk and 36 (61 %) as high-risk (P = 0.029, HR: 3.64, 95 % CI: 1.40 9.50). At 10 years from diagnosis, the low-risk group had a 90 % recurrence-free survival (RFS) rate compared to 60 % for the high-risk group. BreastPRS recurrence score is comparable with Oncotype DX and can reclassify Oncotype DX intermediate-risk patients into two groups with significant differences in RFS. Further studies are needed to validate these findings. Keywords BreastPRS Oncotype DX Breast cancer recurrence Microarray Introduction Clinicians are increasingly incorporating and utilizing genomic information of patients breast tumors via multigene prognostic signatures to guide treatment recommendations. These molecular assays, combined with traditional clinical and pathologic variables, are used to determine the risk of cancer recurrence and the benefits of adding chemotherapy to a patient s treatment regimen. A number of prognostic and predictive multigene assays are commercially available for this purpose, of which the Oncotype DX is currently the most popular for luminal subtypes of nodenegative (N0) breast cancer patients. BreastPRS (Signal Genetics) is a new molecular characterization assay developed and validated from a metaanalysis of publically available genomic datasets. Breast- PRS is unique in that the 200 genes utilized in its algorithm

706 Breast Cancer Res Treat (2013) 139:705 715 (validated in a large series of breast cancer patients) were significantly associated with RFS, independent of traditional prognostic variables including age, tumor size, ER status, tumor grade, and nodal involvement [1]. In contrast to Oncotype DX, BreastPRS is a binary assay which stratifies patients into low- and high-risk groups. In this study, we sought to (i) translate the previously published 200-gene prognostic signature from fresh frozen (FF) to formalin-fixed paraffin-embedded (FFPE) tissue, (ii) compare the BreastPRS prognostic index to the Oncotype DX assay using FFPE patient specimens analyzed by both methods and correlate recurrence scores with clinicopathogic features, and (iii) use publically available whole genome profiles from series of untreated ER? N0 patients to investigate the ability of BreastPRS to reclassify Oncotype DX intermediate-risk patients into binary risk categories (high- vs. low-risk) with clinically significant differences in outcome. The ultimate goal was to assist clinicians with decision making for patients whose tumors are classified as intermediate-risk by Oncotype DX. Materials and methods Translation of BreastPRS from FF to FFPE tissue The 200-gene prognosis signature within BreastPRS was originally developed from gene expression data generated from FF breast cancer tissue. In order to translate this signature for use with FFPE tissue, RNA from FF and FFPE portions of the same 35 individual breast tumors was obtained from a commercial tissue repository (BioServe, Beltsville, MD, USA) (Table 1). Pre-isolated RNA from the FF portion of each tumor was supplied by BioServe and hybridized to Affymetrix U133 GeneChips according to manufacturer recommendations. For the FFPE counterparts, RNA was isolated and amplified from five 10-lM sections of each specimen using the Ovation FFPE WTA System (NuGen Inc., San Carlos, CA, USA). A minimum tumor cell content of[50 % was verified by the supplier. Amplified cdna was fragmented, labeled, and hybridized to a Human Genome U133 Plus 2.0 GeneChip according to manufacturer recommendations. As an additional FF to FFPE validation series, a 20-patient series of matched FF and FFPE breast cancer data was downloaded from National Center for Biotechnology Information (NCBI) Gene Expression Omnibus (GEO) [2]. In both series, the prognostic signature was applied to each genomic profile (averaged across technical replicates where present) and the resulting risk scores were compared between tissue preservation methods. Passing and Bablok regression, a procedure with no special assumptions regarding the distribution of the samples and the measurement errors, was used to assess linearity and consistency [3].The regression equation from this analysis was then used to adjust the low-/high-risk threshold for FFPE tissue, to maintain the previously published characteristics of the signature [1]. Gene expression profiling of FFPE breast tumors previously analyzed with Oncotype DX Two hundred eighty-four patients with consecutively diagnosed invasive breast carcinoma and known Oncotype DX recurrence scores performed as part of their routine clinical care were identified and retrieved from pathology files at Weill Cornell Medical College. Unstained FFPE slides from representative tumor blocks were used for gene array analysis. RNA was isolated after manual microdissection from tissue slides using Prelude FFPE RNA Isolation Module, part no. 1410-50 (NuGen Inc., San Carlos, CA, USA). The isolated RNA was converted to cdna, amplified, and then hybridized to Affymetrix 133 Plus 2.0 Whole Genome microarrays. Normalized gene expression profiles were generated using the MAS5 probe summarization algorithm and annotated with Bioconductor AnnotationData package, release 2.11 [4]. For qualitycontrol purposes, gene chips with fewer than 25 % probeset detection were excluded from further analysis. Creation of a novel gene expression signature to predict a patient s Oncotype DX risk group from microarray data After excluding the 21 genes used by the proprietary Genomic Health recurrence score algorithm, a new signature was developed to predict the Oncotype DX risk group of a breast cancer specimen using Affymetrix U133 Plus 2.0 data. To create the signature, the significance of association between each Affymetrix probe and Oncotype DX risk group (low, intermediate, or high) was determined, using an F test. A strict P value threshold of 1 9 10-6 to reduce the number of false positive probes was used to select the most discriminatory genes whose association with Oncotype DX risk group in the training series was visualized with hierarchical clustering. In order to apply the gene signature to archival breast cancer specimens with outcome data, the gene signature was used to train a diagonal Linear Discriminant Analysis classifier (LDA) using partial cross validation [5]. The trained algorithm, which will be referred to as the Oncotype DX approximation algorithm, was applied to gene expression profiles assembled from public repositories (NCBI GEO) from patients from four previously published studies [6 9] and classified each patient as low, intermediate, or high Oncotype DX risk group. Only those patients with ER?, N0 disease who did not receive adjuvant therapy were included in this analysis. Those predicted to be intermediate-risk

Breast Cancer Res Treat (2013) 139:705 715 707 Table 1 Patient demographics from breast cancer series used in this study FFPE formalin-fixed paraffin embedded, ER estrogen receptor, NA not available, Ax axillary, CT classical type, PT pleomorphic type Characteristic Fresh frozen to FFPE series Weill Cornell Oncotype DX series Archival ER?, node-negative series Number: 55 246 260 Age (mean years) 60 (NA: 20) 57 49 (NA: 34) Histologic grade 1 43 54 2 116 150 3 50 42 Lobular (CT/PT) 32/5 NA NA 0 14 T Size T1 214 146 T2 31 110 T3 1 3 NA Ax. lymph node status Positive 37 0 Negative 209 260 ER status Positive 32 242 260 Negative 20 3 0 NA 3 1 0 Stage I 1 215 146 II 28 31 110 III 0 0 3 NA 26 0 1 Recurrence Yes 56 No 204 Follow-up (median years) 9.9 Gene expression repository IDs ArrayExpress: E-TABM-108 GSE47109 NCBI GEO: GSE11121, GSE4922, GSE6352, GSE7390 Gene expression platform Affymetrix GeneChip U133 Plus 2.0 (100) Illumina HumanRef-8 Expression BeadChip (90) Affymetrix GeneChip U133 Plus 2.0 Affymetrix GeneChip U133 Plus 2.0 were reanalyzed using the BreastPRS prognostic signature and the reclassified high- and low-risk groups were compared to known outcomes. A flow diagram of the study design is shown in Fig. 1. Data management and statistical analysis Statistical analyses were performed using Microsoft Excel, R and Medcalc (version 12.3.0). All t tests were two sided and P values below 0.05 were considered statistically significant. To compare differences in RFS between risk groups on a univariate level, Kaplan Meier analysis and log-rank testing were performed. To evaluate the differences between risk groups on a multivariate level, Cox proportional hazards regression analysis was used, including tumor grade, size, and BreastPRS risk group. Results were analyzed using the Signal Genetics ResultsPX platform and R (www.r-project.org). The

708 Breast Cancer Res Treat (2013) 139:705 715 Fig. 1 Flow diagram summarizing analytical aspects of validation study BreastPRS prognostic score was compared with known clinicopathologic variables (age, tumor size, nodal status, ER/PR/HER2 status, Ki-67, lymphovascular invasion) and Oncotype DX recurrence score for each patient. Data from the 284 patient training series have been deposited at NCBI GEO under accession GSE 47109. Results Patient demographics from the series used in this study are summarized in Table 1. Translation of BreastPRS from FF to FFPE tissue The previously published 200-gene prognosis score was calculated for 35 matched pairs of microarray profiles for FF and FFPE breast cancer specimens which passed RNA and GeneChip quality metrics, as described. The r 2 and intraclass correlation coefficients for paired measurements were 0.67 and 0.90, respectively. Passing and Bablock regression analysis revealed the relationship between tissue types exhibited no significant deviation from linearity (P = 0.91), as shown in Fig. 2a. When the regression equation [FFPE score =-0.265? 1.051 9 FF score] is applied to the previously determined FF-based classification threshold for high-/low-risk (i.e., [-0.38 = highrisk), the FFPE threshold was determined to be [-0.63. This adjustment ensures that the performance of the high-/ low-risk groups is consistent with those previously published for FF tissue, for example in terms of recurrence-free and overall survival. In order to further verify the threshold adjustment, microarray data from a published series of 20 paired FF and FFPE breast cancer specimens were analyzed [2] (Fig. 2b). These specimens were profiled (with technical replication) using the Illumina Whole-Genome DASL Assay which contains 24,526 transcripts. Probe mapping identified 169 of the 200 (85 %) genes in the BreastPRS prognostic signature, therefore the complete algorithm could not be applied. The prognostic signature was calculated on this subset and Passing and Bablok regression showed the relationship between tissue preservation methods to be [FFPE-score = 1.038 9 FF score - 0.154]. In general, both datasets revealed a linear relationship between the BreastPRS prognostic score when measured in paired FF and FFPE specimens. Regression analysis showed that an adjustment to the classification threshold of -0.246 was necessary to perform BreastPRS using FFPE tissue and maintain the previously published performance characteristics of the low- and high-risk groups. Performing a similar comparison on a different patient series profiled on a different platform (Illumina) showed an adjustment of a similar magnitude would be necessary if BreastPRS was to be performed on this system. For comparison purposes with our own dataset (Affymetrix), we computed a 169-gene version of the prognostic score and applied it to the 35 pairs of FF and FFPE specimens in this validation series (Fig. 2c). Regression analysis of this subset algorithm revealed the regression analysis to be similar to that observed for the Mittempergher/Illumina series: FFPE score =-0.214? 1.005 9 FF score. Comparison of BreastPRS prognostic indices to official Oncotype DX recurrence scores Of the 284 cases of invasive carcinoma with known Oncotype DX recurrence scores, 246 cases passed all RNA and

Breast Cancer Res Treat (2013) 139:705 715 709 Fig. 2 Passing and Bablok Regression of the BreastPRS prognostic score calculated on paired FF and FFPE specimens. a Affymetrix 200-gene signature in 35 pairs of FF/FFPE RNA. FFPE score = -0.265? 1.051 9 FF score. b Illumina 169-gene signature in 20 chip-related quality controls. The two prognostic algorithms were compared using whole genome profiles from these cases (Fig. 3a). Both metrics were designed to be continuously associated with risk of disease recurrence, reflected in the interclass correlation coefficient of 0.73 (95 % CI: 0.65 0.79), indicating a moderately strong positive relationship between the two metrics. Figure 3b shows a polynomial regression line fitted to the dataset, suggesting the spread of risk scores may be greater in BreastPRS as compared to Oncotype DX, where more scores are compressed in the low-to-middle range distribution. For Oncotype DX, low-risk = \18, intermediate-risk = 18 30, and highrisk = [30. BreastPRS classified patients as only high- or low-risk, based on the threshold of high-risk = 34 or greater. Of the 30 high-risk Oncotype DX cases, 27 (90 %) were classified as high-risk by BreastPRS (Table 2). Ninety-five low-risk Oncotype DX cases (76 %) were classified as lowrisk by BreastPRS. Interestingly, a majority (60 %) of cases classified as intermediate-risk by Oncotype DX were classified as low-risk by BreastPRS, a group which has previously been shown to not benefit from adjuvant chemotherapy treatment [10]. pairs of FF/FFPE RNA. FFPE score = 1.038 9 FF score - 0.154. c Affymetrix 169-gene signature in same series of patients as a. FFPE score =-0.214? 1.005 9 FF score. FF fresh frozen, FFPE formalin-fixed paraffin-embedded, RNA ribonucleic acid Identification of Oncotype DX intermediate-risk patients in public gene expression profile data repository After identifying a novel set of genes able to hierarchical cluster Affymetrix profiles into three groups which correspond to Oncotype DX low-, intermediate-, and high-risk groups, an Oncotype DX approximation algorithm was created to predict Oncotype DX risk groups of genomic profiles from other datasets. Partial cross validation of the 120-gene Oncotype DX approximation algorithm on the Oncotype DX training series reproduced accuracy of the hierarchical clustering in terms of risk group assignment and official Oncotype DX classification (Fig. 4). Overall, 72 % of patients were predicted correctly by the 120-gene algorithm with a mean cross-validation sensitivity of 68 %, specificity of 84 %, positive predictive value of 69 %, and negative predictive value of 85 %. Variation between the official and predicted Oncotype DX risk groups may be attributable to intratumoral heterogeneity, differences in measuring gene expression using qpcr versus microarray, or the general reproducibility of the 21-gene signature itself.

710 Breast Cancer Res Treat (2013) 139:705 715 [11]. With this method, only an additional 3 % of patients analyzed were classified into the same risk group as the official Oncotype DX risk group (data not shown), a statistically insignificant difference, as determined by a t test of proportions (P = 0.63). Therefore, when a tumor is profiled using the Affymetrix platform, it appears that using the 21 gene (patented) Oncotype DX signature does not reproduce the commercial qpcr assay with significantly greater accuracy than using a novel 120-gene Oncotype DX approximation algorithm. Within the 120-gene Oncotype DX approximation algorithm predicted risk groups, it was the intermediate-risk group which differed most between the official 21-gene PCR method risk groups and those predicted by the 120-gene approximation algorithm. Of the true intermediate-risk patients, 73 % were classified as intermediaterisk by the 120-gene approximation algorithm. For gene set comparison purposes, the Affymetrix implementation of the official 21-gene signature was also evaluated. Using the 21-gene set, 69 % of true intermediate-risk patients were classified as such. This suggests that the 120-gene Oncotype DX approximation algorithm is more accurate at identifying true Oncotype DX intermediate-risk patients in previously published breast cancer gene expression datasets. Performing additional algorithm incorporating gene-reselection in each loop of the cross-validation procedure did not improve classification accuracy (data not shown). Fig. 3 a Scatter plot of 246 FFPE breast cancer specimens analyzed by both Oncotype DX and BreastPRS. Cases are colored according to the risk group of each assay. Both prognosis scores were created to be continuous markers of worsening prognosis with higher scores correlating with higher rates of outcome and poor overall survival. b Polynomial regression fitted line exploring the relationship between the Oncotype DX recurrence score and BreastPRS prognosis index. This plot suggests that the spread of risk scores may be greater in BreastPRS as compared to Oncotype DX, where more scores are compressed in the low-to-middle range distribution. OT Oncotype DX, FFPE formalin-fixed paraffin-embedded Table 2 Direct comparison of risk group stratification by Oncotype DX (21-gene qpcr) and BreastPRS (200-gene Affymetrix signature) Oncotype DX Risk Total BreastPRS BreastPRS Low-risk High-risk Low 125 95 (76 %) 30 (24 %) Intermediate 91 55 (60 %) 36 (40 %) High 30 3 (10 %) 27 (90 %) Total 246 153 93 v 2 P value for risk group association \0.0001 For comparison purposes, a separate cross-validation exercise was performed using the 16 genes used by Genomic Health to perform the Oncotype DX assay (after renormalizing to the 5 Oncotype normalization genes) as published Application of the 120-gene Oncotype DX approximation algorithm to expression profiles from historical datasets to classify patients into predicted Oncotype DX risk groups The 120-gene Oncotype DX approximation algorithm was applied to the whole genome expression profiles of 260 ER?, N0, untreated (i.e., no Tamoxifen or chemotherapy) breast cancer patients from four previously published studies [6 9], with recurrence-related outcome data available. Application of the 120-gene Oncotype DX approximation algorithm classified 169 (65 %) as low-risk, 59 (23 %) as intermediate-risk, and 32 (12 %) as high-risk. Kaplan Meier analysis of these risk groups approached statistical significance in 15-year RFS (log-rank test P value 0.088) (Fig. 5a). In a multivariate analysis, the difference in RFS between the high-risk and low-risk groups was significant (P = 0.043), but not between the high- and intermediate-risk groups (P = 0.55). Because this series of patients were compiled from four historical datasets, each with their own patient selection criteria, the proportion of patients in each risk group was not expected to resemble that observed in the general population or target demographic of the Oncotype DX assay. Despite this, the proportions observed are similar to that reported in

Breast Cancer Res Treat (2013) 139:705 715 711 Fig. 4 Hierarchical clustering of the 120-gene Oncotype DX approximation algorithm identified by comparing whole genome profiles of patients samples previously analyzed by Oncotype DX the NSABP B-14 trial, i.e., low-risk (51 %), intermediaterisk (22 %), and high-risk (27 %) [11]. When the 200-gene BreastPRS algorithm is applied to expression profiles from this set of patients (Fig. 5b), 129 (49.6 %) were classified as low-risk and 131 (50.4 %) were classified as high-risk with a significant difference in 15-year RFS (log-rank test P value = 0.0001). In a multivariate model with grade and tumor size, the BreastPRS signature remains statistically significant (P \ 0.0001) with a hazard ratio of 3.94 (95 % CI: 1.99 7.78) as shown in Table 3. BreastPRS was then applied to the 169 cases classified as low-risk by the 120-gene Oncotype DX approximation

712 Breast Cancer Res Treat (2013) 139:705 715 Fig. 5 Kaplan Meier analyses of a all ER?, node-negative, untreated archival patients stratified by 120-gene Oncotype DX approximation algorithm risk groups (n = 260) P = 0.088 and b by 200-gene BreastPRS algorithm risk groups (n = 260) P = 0.0001, HR: 3.00 (95 % CI: 1.77 5.08). c Archival ER?, node-negative, untreated breast cancer patients predicted to be Oncotype DX intermediate-risk by the 120-gene approximation signature (from a), reclassified as high- or low-risk by BreastPRS (n = 59) P = 0.029, HR: 3.64 (95 % CI: 1.40 9.50). ER estrogen receptor, HR hazard ratio, CI confidence interval algorithm to investigate the inconsistent risk group assignments of cases from the Weill Cornell series that were classified as low-risk by Oncotype DX and high-risk by BreastPRS (24 % of low-risk cases, Table 2). Sixtythree of the 169 (37 %) low-risk cases were classified as high-risk by BreastPRS. Kaplan Meier analysis was

Breast Cancer Res Treat (2013) 139:705 715 713 performed on these 63 and compared with the remaining 106 that were classified as low-risk by both methods. The 63 cases classified as high-risk by BreastPRS experienced significantly shorted RFS compared to the cases classified as low-risk by both methods [hazard ratio: 2.96 (95 % CI: 1.39 6.28), P = 0.0010]. At 10 years from diagnosis, the chance of recurrence is 25 % lower in the low-risk group compared to the high-risk group (90 vs. 65 %) and 15 % lower at 15 years from diagnosis. Subgroup analysis of Oncotype DX intermediate-risk patients by BreastPRS To test the hypothesis that BreastPRS is able to reclassify Oncotype DX intermediate-risk patients as high- or lowrisk, the 59 (ER?, N0, untreated) intermediate-risk patients identified in the retrospective series were analyzed by BreastPRS (Fig. 5c). Twenty-three (39 %) patients were classified as low-risk and 36 (61 %) as high-risk. The hazard ratio of a high-risk classification was 3.64 (95 % CI: 1.40 to 9.50) and log-rank testing indicated that the results were statistically significant (P = 0.029). At 10 years from diagnosis, the low-risk group had a 90 % RFS rate, compared to 60 % for the high-risk group. When adjusted for tumor grade and size in a multivariate cox proportional hazards model (Table 3), BreastPRS was the closest variable to achieving statistical significance (P = 0.055, HR: 3.58, 95 % CI: 0.97 13.14). These data indicate that BreastPRS is able to reclassify Oncotype DX intermediaterisk patients as high- or low-risk for disease recurrence, independent of clinical variables of size and tumor grade. Discussion While current guidelines recommend consideration of chemotherapy for the majority of patients with invasive breast carcinoma [12], most patients with small, ER? tumors will not gain additional benefit from adding adjuvant chemotherapy to Tamoxifen, and can likely be spared the toxicities of the former. Clinicians have traditionally relied on clinical and pathologic factors including tumor size, axillary lymph node status, histologic grade, and hormone receptor status when assessing the need for adjuvant chemotherapy in patients with early breast cancer. Recent advances in gene expression profiling and microarray technology have led to a greater understanding of the biology of breast cancer at the molecular level. This coincides with advances in imaging techniques that have led to an increase in the detection of smaller invasive cancers. Clinicians are increasingly incorporating genomic data from patients tumors via multigene prognostic signatures to gather additional information, particularly the risk of distant recurrence, to aid in deciding as to whether Table 3 Multivariate analysis of BreastPRS and the 120-gene Oncotype DX approximation algorithm NCBI GEO National Center for Biotechnology Information Gene Expression Omnibus, LDA linear discriminant analysis, HR hazard ratio, CI confidence interval Description NCBI GEO archival series with BreastPRS (n = 260) NCBI GEO archival series with 120-gene Oncotype DX approximation algorithm (n = 260) 120-gene predicted Oncotype intermediate-risk patients with size and grade information (n = 59) Cox proportional hazards ratio and P value Covariate P value HR 95 % CI of HR Grade 2 vs. 1 0.274 1.64 0.68 3.93 Grade 3 vs. 1 0.825 1.12 0.40 3.13 Size T2 vs. T1 0.011 2.15 1.19 3.85 Size T3 vs. T1 0.005 5.97 1.74 20.51 BreastPRS high- vs. low-risk \0.001 3.94 1.99 7.78 Grade 2 vs. 1 0.224 1.73 0.7154 4.22 Grade 3 vs. 1 0.593 1.34 0.45 4.03 Size T2 vs. T1 0.037 1.85 1.04 3.29 Size T3 vs. T1 \0.001 13.08 3.71 46.08 120-gene Oncotype approximation high- vs. intermediate-risk 0.545 1.32 0.54 3.23 120 gene Oncotype approximation highvs. low-risk 0.043 2.49 0.16 0.97 Grade 2 vs. 1 0.926 0.90 0.11 7.37 Grade 3 vs. 1 0.834 0.77 0.07 8.40 Size T2 vs. T1 0.087 2.73 0.86 8.63 BreastPRS highvs. low-risk 0.055 3.58 0.97 13.14

714 Breast Cancer Res Treat (2013) 139:705 715 or not to add chemotherapy to a patient s treatment regimen. The 21-gene Oncotype DX RT-PCR test (Genomic Health, Redwood City, CA, USA) is currently the most widely used breast cancer prognostic assay, largely due to its performance on routinely prepared FFPE tissue blocks. The assay was developed and validated from large series of patients from National Surgical Adjuvant Breast and Bowel Project (NSABP) trials, as well as other independent studies and has been shown to be both prognostic of breast cancer recurrence and predictive of chemotherapy benefit [10, 11, 13, 14]. The Oncotype DX assay stratifies patients into low-, intermediate-, and high-risk groups, corresponding to risk of recurrence at 10 years in ER?, N0 patients treated with Tamoxifen. High-risk patients have been shown to benefit from chemotherapy. The benefit of adjuvant chemotherapy in the intermediate-risk group is uncertain and is currently under investigation [15, 16]. BreastPRS is a 200-gene microarray-based prognostic signature that was generated and validated on multiple independent series of breast cancer patients using publicly available gene expression profiles [1]. In a prior study, the BreastPRS algorithm was applied to expression profiles of 1,016 patients, and separated them into risk groups with significant differences in recurrence-free and overall survival [1]. In untreated, N0 patients, the sensitivity and specificity of the assay for predicting RFS were 88 and 44 %, respectively, with positive and negative predictive values of 30.5 and 92 %, respectively. In this study, we compared the BreastPRS and Oncotype DX assays using FFPE tissue from a series of patients treated at our institution, as well as publically available gene expression profiles from the GEO, a public repository maintained by the NCBI (http://www.ncbi.nlm.nih.gov/geo/). GEO serves as a central database for high-throughput microarray and next-generation sequencing data that is submitted by researchers and is freely available to the public. A major strength of Oncotype DX is the ability to perform the assay on FFPE tissue, as preserving tumor samples as FF tissue is not practical for routine clinical care at most institutions. Moreover, the ability of an assay to be performed on FFPE tissue allows retrospective study of large cohorts of patients for validation. The BreastPRS algorithm was originally developed from expression profiles of FF tumor tissue. The aim of this study was to determine whether the BreastPRS algorithm could be translated to FFPE and this was accomplished using matched pairs of FF and FFPE tissue from the same tumors. Here we show that a linear relationship exists between BreastPRS scores generated from FF and FFPE tissue with an intraclass correlation coefficient of 0.90, indicating strong positive agreement. Overall the FFPE specimens resulted in lower BreastPRS scores when compared to the paired FF specimen, therefore linear regression analysis was used to adjust the threshold for high-/low-risk group classification accordingly. Next, we compared the Oncotype DX and BreastPRS algorithms on a series of patients with known Oncotype DX results performed as a part of their clinical care. We found significant correlation between the prognostic metrics generated by Oncotype DX and BreastPRS when applied to genomic profiles from the studied patients, with 90 % of Oncotype DX high-risk cases predicted to be also high-risk by BreastPRS. Among low-risk Oncotype DX patients, 24 % were reclassified as high-risk by BreastPRS, a change in classification that would have important prognostic and treatment implications. Because outcome data were not available for the Weill Cornell series, we investigated this inconsistency in patients from the archival ER?, untreated, N0 series used in this study. BreastPRS was applied to tumors classified as low-risk by the 120-gene Oncotype DX approximation algorithm. Patients in this group that were reclassified as high-risk by BreastPRS showed significantly shorter RFS compared with those classified as low-risk by both assays. This finding raises the possibility that there is a subset of patients classified as low-risk by Oncotype DX who may benefit from chemotherapy. Certainly, this observation warrants further investigation and confirmation by other independent studies. Finally, a retrospective head-to-head comparison of BreastPRS and Oncotype DX was performed by applying a novel 120-gene Oncotype DX approximation signature, trained on commercial Oncotype DX results, to previously published series of untreated, N0, ER? patients with outcome data. BreastPRS resulted in a more statistically significant log-rank test P value and fewer recurrences in the low-risk group, as compared to the microarray-based 120-gene approximation of Oncotype DX. Subgroup analysis of the retrospective cases classified as Oncotype DX intermediate-risk was then performed. BreastPRS was able to reclassify these patients as either high- or low-risk for recurrence, with the reclassified groups having a highly significant difference in outcome (hazard ratio 3.64, P = 0.029). We found that 61 % of those classified as intermediate-risk using the 120-gene Oncotype DX approximation algorithm were classified as high-risk by BreastPRS and would likely benefit from adjuvant chemotherapy. It will be interesting to see whether the findings of the TAILORx trial mirror these findings. The gene lists used by BreastPRS and Oncotype DX overlap by two genes only [Cyclin B1 (CCNB1) and Ki67 (MKI67)]. Despite this, DAVID gene ontology analysis of both gene lists reveals a number of common gene families between the two sets (Supplementary Table). Genes involved in cell cycle regulation and apoptosis are significantly represented in both sets, however, BreastPRS also contains genes involved in metabolism, intracellular

Breast Cancer Res Treat (2013) 139:705 715 715 organization, hormone receptor binding, migration, and immune function. Conversely, Oncotype contains genes involved in metal binding and tissue development; categories not significantly represented in BreastPRS. The submission and use of GEO data sets by researchers continues to increase and these data sets represent a powerful tool to perform meta-analyses using large volumes of samples. The use of GEO data sets is likely to increase as the availability of FFPE tissue from large trials with longterm follow up becomes depleted. Using a combination of FFPE material and expression profiles from GEO, we have shown that the 200-gene BreastPRS prognosis algorithm is comparable to the 21-gene Oncotype DX assay and effectively separates Oncotype DX intermediate risk cases into binary categories with significant differences in RFS. Additional validation studies are forthcoming with the goal of commercializing BreastPRS as a stand-alone breast cancer prognostic assay that can be performed on FFPE tissue. Disclosures TMD declares no conflict of interest. RKVL is the Head of Bioinformatics and New Product Development for Signal Genetics and owns stock in the company. LTV declares no conflict of interest. WH, RF, NB, and LSJ are employees of Signal Genetics. SJS is a paid consultant of Signal Genetics. References 1. Van Laar RK (2011) Design and multiseries validation of a webbased gene expression assay for predicting breast cancer recurrence and patient survival. J Mol Diagn 13(3):297 304 2. Mittempergher L, de Ronde JJ, Nieuwland M et al (2011) Gene expression profiles from formalin fixed paraffin embedded breast cancer tissue are largely comparable to fresh frozen matched tissue. PLoS One 6:e17163 3. Passing H, Bablok (1983) A new biometrical procedure for testing the equality of measurements from two different analytical methods. Application of linear regression procedures for method comparison studies in clinical chemistry, part I. J Clin Chem Clin Biochem 21:709 720 4. Gentleman RC, Carey VJ, Bates DM et al (2004) Bioconductor: open software development for computational biology and bioinformatics. Genome Biol 5:R80 5. Dudoit S, Fridlyand J, Speed T (2002) Comparison of discrimination methods for the classification of tumors using gene expression data. J Am Stat Assoc 97:77 87 6. Schmidt M, Bohm D, von Torne C et al (2008) The humoral immune system has a key prognostic impact in node-negative breast cancer. Cancer Res 68:5405 5413 7. Ivshina AV, George J, Senko O et al (2006) Genetic reclassification of histologic grade delineates new clinical subtypes of breast cancer. Cancer Res 66:10292 10301 8. Loi S, Haibe-Kains B, Desmedt C et al (2007) Definition of clinically distinct molecular subtypes in estrogen receptor-positive breast carcinomas through genomic grade. J Clin Oncol 25:9 1246 9. Desmedt C, Piette F, Loi S et al (2007) Strong time dependence of the 76-gene prognostic signature for node-negative breast cancer patients in the TRANSBIG multicenter independent validation series. Clin Cancer Res 13:3207 3214 10. Paik S, Tang G, Shak S et al (2006) Gene expression and benefit of chemotherapy in women with node-negative, estrogen receptor-positive breast cancer. J Clin Oncol 24:3726 3734 11. Paik S, Shak S, Tang G et al (2004) A multigene assay to predict recurrence of tamoxifen-treated, node-negative breast cancer. N Engl J Med 351:2817 2826 12. Ma XJ, Patel R, Wang X et al (2006) Molecular classification of human cancers using a 92-gene real-time quantitative polymerase chain reaction assay. Arch Pathol Lab Med 130:465 473 13. Habel LA, Shak S, Jacobs MK et al (2006) A population-based study of tumor gene expression and risk of breast cancer death among lymph node-negative patients. Breast Cancer Res 8:R25 14. Albain KS, Barlow WE, Shak S et al (2010) Prognostic and predictive value of the 21-gene recurrence score assay in postmenopausal women with node-positive, oestrogen-receptorpositive breast cancer on chemotherapy: a retrospective analysis of a randomised trial. Lancet Oncol 11:55 65 15. Zujewski JA, Kamin L (2008) Trial assessing individualized options for treatment for breast cancer: the TAILORx trial. Future Oncol 4:603 610 16. Sparano JA (2006) TAILORx: trial assigning individualized options for treatment (Rx). Clin Breast Cancer 7:347 350