The BRCA1ness signature is associated significantly with response to PARP inhibitor treatment versus control in the I-SPY 2 randomized neoadjuvant setting

The collaborative European Union-funded effort FP7 RATHER project (Rational Therapy for Breast Cancer) aims to integrate gene expression profiling, copy number variation, kinome variation and kinase activation status in an effort to identify new targets for therapy of difficult-to-treat breast cancer subtypes, including TN breast cancer (www.​ratherproject.​com). The RATHER project retrospectively identified 128 TN breast cancer patients with long-term follow-up in total: 70 from Netherlands Cancer Institute (NKI), Amsterdam, the Netherlands and 58 from Addenbrooke’s Hospital, Cambridge, UK.

The primary inclusion criterion for the RATHER cohort was availability of sufficient isolated frozen RNA tissue in the tissue bank and diagnostic information indicative of TN breast cancer. We enriched for frozen tumors with 30% or greater tumor content (2 x 8-μm serial sections, hematoxylin and eosin stained). The local medical ethics authorities of both centers approved the collection protocols. Sectioning of tumor tissue and RNA isolation were performed as described previously [37].

Samples with RIN value > 5 according to 2100 Bioanalyzer (Agilent Technologies) assessment were selected for further analysis. Gene expression data were generated as described previously [37]. Briefly, feature signal intensities were processed and extracted using the ‘limma’ R package with background subtraction using an offset of 10 and log₂ transformed data. Probe intensities were quantile normalized with in-house R scripts and missing values (including probes with signal intensities < 1 after preprocessing) were imputed by the 10 nearest-neighbor method. A biobank batch effect was adjusted using ComBat [38]. Genes with multiple probes were summarized by the first principal component of a correlating subset.

The multiplex ligation-dependent probe amplification (MLPA) method was used to generate copy number profiles for the determination of the BRCA1-like status of the tumors. The assay was performed, fragments analyzed and data normalized according to the manufacturer’s protocol (MRC-Holland). Class prediction (BRCA1-like/non-BRCA1-like) was carried out on the normalized data according to published instructions [16].

Signature development was performed using Partek Genomics Suite (partek.com) (categorical signature) and Matlab (https://​mathworks.​com/​) (translation to continuous score) on 128 samples. Top variable genes (variance >1 across all samples) were used for the model input (N = 2049). The classification model of diagonal linear discriminant analysis (DLDA) with equal prior probabilities was run to select the signature genes. Groups of genes ranked by their significance in a univariate ANOVA examining the BRCA1-like/non-BRCA1-like MLPA status were tested (groups from 1 to 100, in increments of 1). Single-level leave-one-out cross validation (LOOCV) with the maximum number of partitions was used to internally validate and calculate the performance of the model.

The significant number of genes in the model (n = 77, Additional file 1: Table S1) was selected based on the ROC area under the curve (AUC as specified by Partek). After the model is run, each sample is allocated a posterior probability for each class (BRCA1-like and non-BRCA1-like) and the sample is assigned to the class with the highest posterior probability (BRCA1ness). This categorical signature was then transferred to the diagnostic setting to better comply with quality and regulatory requirements using a nearest centroid model; a robust method with both reasonable and favorable characteristics for many measurements on a modest amount of patients [39]. Briefly, raw full genome data were normalized and class centroids were calculated (median per gene) for each of the 77 genes for each class (BRCA1ness/non-BRCA1ness) using the discovery set. These calculated centroids were used as the template for BRCA1ness/non-BRCA1ness. Pearson correlations of each new sample with the BRCA1ness/non-BRCA1ness templates were calculated (Additional file 1: Table S1) and combined into a single continuous score by subtracting the correlation to the non-BRC1Aness template from the correlation to the BRC1Aness template. In order for a sample to be classified as BRCA1ness, a threshold was established with a high sensitivity while preserving specificity close to 0.75. The classification threshold was set at –0.3; that is, a sample with a BRCA1ness score > –0.3 was classified as BRCA1ness and a sample with a score < –0.3 was classified as non-BRCA1ness.

The I-SPY 2 TRIAL is a standing multicenter, phase 2 platform trial to screen experimental regimens in combination with standard chemotherapy in the neoadjuvant treatment of breast cancer. Patients are adaptively randomized into one of four experimental arms or a control arm (Fig. 1) [35]. In this portion of the I-SPY 2 TRIAL, eligible patients received weekly paclitaxel at 80 mg/m² (T) i.v. for 12 doses alone (control), or in combination with an experimental regimen. Patients randomized to V-C received 50 mg of veliparib by mouth twice daily for 12 weeks and carboplatin at AUC 6 on weeks 1, 4, 7 and 10 concurrent with weekly paclitaxel. Following paclitaxel ± V-C, all patients received doxorubicin 60 mg/m² and cyclophosphamide 600 mg/m² (AC) i.v. every 2–3 weeks for four doses with myeloid growth factor support as appropriate per protocol followed by surgery that included axillary node sampling. The V-C arm was open to HER2-negative patients; and was graduated in the TN group. The BRCA1ness signature was one of the qualifying dichotomous biomarkers assessed as a predictor of response to V-C relative to standard chemotherapy.

To assess the BRCA1ness signature in this validation set as a specific biomarker of V-C response, gene expression data from 116 HER2-negative patients (V-C, n = 72 and concurrent controls, n = 44) were analyzed. A Customized Agilent 44 K array (Agendia) was used to evaluate the 77-gene signature BRCA1ness classification. The association between BRCA1ness classification and response in the V-C and control arms alone was assessed using Fisher’s exact test, and the relative performance between arms (biomarker × treatment interaction, likelihood ratio test) using a logistic model. We included adjustment for hormone receptor status (HR/TN) and tumor size in our model. Our sample size is small, and thus statistical calculations (p values) are descriptive rather than inferential. This analysis does not adjust for multiplicities of other biomarkers evaluated in the trial but outside this study.

We developed a BRCA1ness signature using whole genome gene expression data. The signature has been developed on fresh frozen (FF) breast tumors that were categorized as either BRCA1-like or non-BRCA1-like using a DNA copy number MLPA-based classifier [16], and endeavors to predict BRCA1-like tumors with a high sensitivity/specificity rate.

Forty-eight percent of the tumors (61/128) in the discovery cohort were classified as BRCA1-like and the remainder was assigned to the non-BRCA1-like class. We employed the gene expression profiles of the tumors, DLDA and the labels assigned by the MLPA-based classifier to train a classifier that distinguishes between the two classes. Using the ROC area under the curve model (AUC) as the performance criterion we identified a 77-gene signature that resulted in the highest performance (Table 1).

Unsupervised hierarchical clustering of the genes in the 77-gene signature indicates separation between the classes (Fig. 2). We transferred the signature to a diagnostic setting using a nearest centroid-based algorithm. The sensitivity and specificity for detecting BRCA1-like status as defined by MLPA were 96.7% and 73.1%, respectively. Using Ingenuity Pathway Analysis (Qiagen) to identify key biological processes associated with the 77 genes in the BRCA1ness 77-gene signature, we found cellular assembly and control and DNA replication, recombination and repair to be among the top associated pathways and functions (Fig. 3a). In addition, we observed serine and glycine biosynthesis to be associated with the 77-gene signature genes, indicating that these genes may be responsible for reprogramming of metabolic processes, which can lead to tumor progression (Fig. 3a). Supporting the pathway and molecular function results, network analysis revealed a network centered upon cell the cycle control regulator cyclin A (Fig. 3b).

The BRCA1ness signature was applied to 116 HER2-negative patients (V-C, n = 72 and concurrent controls, n = 44). Fifty-five patients were classified as BRCA1ness. Fourteen percent of these patients were hormone receptor-positive (ERα/PR) and HER2-negative. The distribution of pathological complete response (pCR) rates among BRCA1ness and non-BRCA1ness groups is shown in Fig. 4a [36] and Table 2. Association between the BRCA1ness classification and patient response was seen in the V-C arm (OR = 3.2, p = 0.03) but not in the control arm (OR = 0.39, p = 0.45) (Fig. 4b). A significant biomarker × treatment interaction (p = 0.025) was also observed. Although there is enrichment for TN samples in the BRCA1ness group in univariate analysis (Table 2), this interaction remains significant upon adjusting for HR (p = 0.023) (Fig. 4c).

In addition we found that the interaction also remains significant when adjusting for tumor size and HR (p = 0.038). We use the likelihood ratio test to formally demonstrate that the logistic regression with the addition of the HR (and tumor size) terms does not provide a better fit to the data. When the hormone receptor-positive BRCA1ness classified patients are added to the graduating TN subset, the OR associated with V-C is 4.03. This is comparable to that of the TN alone (OR = 4.04), while increasing the prevalence of ‘biomarker-positive’ predicted V-C sensitive patients by 8%.