Low Recombination Proficiency Score (RPS) Predicts Heightened Sensitivity to DNA-Damaging Chemotherapy in Breast Cancer

Breast cancer remains a leading cause of cancer death for women worldwide (1). Molecular-based cancer tests have been developed to augment the standard clinical and pathologic features used to tailor treatments to individual breast cancer patients (2). While some of these tests provide prognostic information (3–5) and/or predict responsiveness to therapies in general (5–7), relatively few are effective in predicting tumor sensitivity to individual types of treatment. Notable exceptions include the detection of estrogen receptor (ER) expression, human epidermal growth factor receptor 2 (HER2) gene amplification, and BRCA1/2 mutations, which predict sensitivity to endocrine therapy, trastuzumab, and inhibitors of poly (adenosine diphosphate-ribose) polymerase 1 (PARP1), respectively (8–11).

Human tumors exhibit a wide range of responsiveness to DNA-damaging treatments. This variation is hypothesized to result, at least in part, from differential efficiencies in DNA repair pathways. Homologous recombination (HR) mediates resistance to several classes of DNA-damaging therapies by catalyzing the repair of double-stranded breaks and by promoting tolerance to lesions that disrupt replication forks (12). HR also serves a genome-stabilizing role by ameliorating endogenous DNA lesions (12). In the context of patient care, HR deficiency in tumors can induce opposing net outcomes in terms of patient outcome. The underlying genomic instability associated with HR deficiency can fuel malignant progression and generate more aggressive tumor characteristics; however, HR deficiency can simultaneously induce sensitivity to DNA-damaging treatments, which creates a therapeutic opportunity (13).

We hypothesized that measuring HR proficiency in individual tumors can be used to guide clinical cancer care. We previously described the Recombination Proficiency Score (RPS), which provides prognostic and predictive data on individual cancers by indirectly quantifying their HR repair proficiency (13). RPS is calculated based on the expression levels for four genes involved in DNA repair pathway preference [RIF1, PARI (PARPBP), RAD51, and Ku80 (XRCC5)]; high expression of these genes yields a low RPS value. We found that these four RPS-defining genes provided a superior prediction of HR activity when compared with algorithms that utilize smaller or larger combinations of genes. Cells with a low RPS exhibit HR suppression and frequent DNA copy number alterations, which are characteristic of error-prone repair processes that arise in HR-deficient backgrounds. Consistent with this observed genomic instability, tumors with low RPS exhibit more malignant clinical and pathologic features. The RPS system was initially studied in patients with non–small cell lung carcinomas (NSCLC), which revealed that HR suppression plays a central role in promoting malignant progression. The adverse prognosis associated with a low RPS was diminished in NSCLC patients that received adjuvant platinum-based chemotherapy, suggesting that HR suppression and associated sensitivity to DNA-crosslinking drugs counteract the adverse baseline prognosis associated with a low RPS.

Our preliminary work also suggested that low RPS breast cancers similarly present with abundant malignant features, including ER loss, large-segment genomic deletions/amplifications, HER2 gene amplification, and TP53 gene mutations (13). Here, we test the hypothesis that RPS has utility in predicting responsiveness of breast cancers to therapy, by demonstrating that low RPS tumors are especially vulnerable to DNA-damaging chemotherapy. Importantly, we find that basal, HER2-enriched, and luminal B breast cancer subtypes (14) exhibit relatively low RPS values, which may help explain the aggressive clinical behaviors of these subtypes (5). Although RPS values associate with these molecularly defined intrinsic subtypes of invasive breast cancer, RPS independently predicts sensitivity to anthracycline-based chemotherapy when controlled for subtype.

All datasets were identified (as of August 19, 2016) and downloaded from publicly available electronic databases (PubMed, GEO, ArrayExpress) using the following keywords: breast cancer, preoperative chemotherapy, and neoadjuvant chemotherapy. After reviewing all retrieved studies, we identified 28 relevant datasets. Five neoadjuvant chemotherapy breast cancer datasets (Table 1; Supplementary Table S1) were selected for subsequent analysis after elimination of ineligible studies or patients based on criteria outlined in Supplementary Fig. S1. Briefly, patients were excluded from analysis if they received trastuzumab or if their chemotherapy regimen did not include all three of the following: an anthracycline, cyclophosphamide, and a taxane. Available data regarding pathologic complete response (pCR) incidence, age, tumor grade, estrogen receptor (ER), and human epidermal growth factor receptor 2 (HER2) statuses, tumor size, and nodal status were required for inclusion. Incidence of pCR following chemotherapy was defined as no residual invasive cancer in the breast and axillary lymph nodes. Post-chemotherapy residual cancer burden (RCB) information was available for 358 patients. RCB values indicate the extent of residual disease as follows: 0 = none, I = minimal, II = moderate, and III = extensive (15).

The gene expression data for these five publicly available datasets were previously generated on a uniform microarray platform (Affymetrix HG-U133A). Three additional independent breast cancer datasets (Supplementary Table S2), also generated on the Affymetrix HG-U133A microarray platform, were selected and combined to serve as an independent frozen reference cohort, which enabled a per-subject calculation of gene expression z-scores in the five study datasets. We performed background correction, quantile normalization, robust-weighted-average summarization, and log₂-tranformation on each raw CEL file using the Single-Channel Array Normalization (SCAN) algorithm (16). SCAN is a microarray normalization algorithm that only uses data from within each array to estimate and remove probe-specific bias and background noise; hence, every sample can be processed without dependence on other samples or requirement of external reference sets. Samples with identical raw signals across all probesets and identical array scan dates were identified and removed, because these likely represented redundant subjects shared by different studies (Supplementary Table S3). Of these potentially redundant samples, 90% exhibited identical clinical information and 10% showed inconsistencies in clinical information between studies.

Our previous studies defined RPS as a linear combination of four signature genes (13): |$RPS\ = \ - 1\ *\ ( {RIF1\ + \ PARPBP\ +$| |RAD51\ + \ XRCC5} )$|⁠, where values denote median-normalized and log₂-transformed gene expression measurements for an individual sample. This formula was used to determine RPS values for BRCA-mutant and sporadic clinical breast cancers (GSE40115 dataset). Basal and luminal B intrinsic breast subtypes were selected for this analysis to minimize the impact of breast subtype on differences in RPS values between BRCA mutants and sporadic tumors.

As a modification to the previously described method for calculating RPS, the mean (μ) and standard deviation (σ) of the independent frozen reference cohort was used to transform SCAN-normalized gene expression signals (x) to z-scores using the following formula: |z = {\frac{{x - \mu }}{\sigma }$| (Supplementary Fig. S2). Non-uniform weighting of each gene was used to develop a logistic model predicting probability of pCR after neoadjuvant chemotherapy (Supplementary Fig. S3). To determine the weight of each gene, we built a random forest (RF) model to estimate the variable importance of each gene. The five neoadjuvant chemotherapy datasets were combined into a single cohort after SCAN normalization and z-transformation, and samples were randomly split into training and independent test sets. The training set was used to tune the parameters and select the best model using 10-fold cross validation with custom evaluation metrics to handle imbalanced classes (17). After training, the test set was used to independently assess the performance of the final model. In each case, we used various splitting percentages from 50%–50% to 90%–10% for training and test sets. We started with a full model with 10 predictors, including the four gene expression values, as well as clinical and pathologic factors: |$pCR \sim RIF1\ + \ PARPBP\ +$| |$RAD51\! \, +\! \ XRCC5\ + \ Age \lt 50 \ + \ ERplus\ + \ HER2plus + T3/4$| |$+ \ {\hbox{\rm node-positive}} + {\curr Grade}\ 3\gt$|⁠. Variables with importance scores <20 (scaled to have a range of 0 to 100) were eliminated, leaving only the four RPS genes and ER status as predictors (Supplementary Fig. S4): |$pCR \sim RIF1 + PARPBP + RAD51 + XRCC5$| |$+\ ERplus$|⁠. The performance of this reduced model was evaluated in the test set with 81% sensitivity and 67% specificity, respectively (Supplementary Table S4).

To further reduce the complexity of the predictor, samples were stratified into ER-positive (n = 301) and ER-negative (n = 212) subsets to eliminate the impact of ER status in the model. Using the expression levels for the four genes as predictors, we built two separate models in ER-positive and ER-negative subsets of samples (50%–50% for training and test sets): |pCR\sim RIF1\ + \ PARPBP\ + \ RAD51\ + \ XRCC5$| (Supplementary Fig. S3). The performance of the ER-negative model was superior in predicting pCR as compared with the ER-positive model. Therefore, we used the variable importance estimates in the ER-negative model to calculate the weight of each gene in contribution to RPS_b. This generated the final breast cancer-specific RPS model: |$RPSb\ = \ - 1*( {0.2171423\ *\ RIF1 +$| |$0.1946173 \, *\, PARPBP\, + \,0.2783017\, *\, RAD51\, + \, 0.3099387 \ *} }$| |XRCC5} )$|⁠, with the sum of gene weight coefficients equal to 1.

Continuous RPS_b values were compared with expected probability of pCR using logistic regression models (Fig. 1; Supplementary Table S5). We built logistic regression models to fit the relationship between probability of pCR (y variable, binary) and RPS_b value (x variable, continuous) using the Generalized Linear Models method (R function glm, link function logit). We fitted a logistic regression model (R function glm) on each of the ER-positive and ER-negative subsets. Hosmer–Lemeshow tests (R function hoslem.test) were used to assess whether the models fit the respective data. χ² tests were used to assess for differences between RCB class frequencies as a function of RPS_b value. Multivariate logistic regression analysis was used to assess for interaction between RPS_b and class of neoadjuvant chemotherapy (anthracycline- vs. non–anthracycline-based) in the GSE21997 dataset. Multivariate logistic regression analysis was used to determine the odds ratio of achieving a pCR as a function of RPS_b, after controlling for standard clinical and pathologic factors or intrinsic breast subtype. Annotations of intrinsic breast subtypes as defined by the PAM50 algorithm (5) were available for 457 patients (Supplementary Table S5). In these analyses, we omitted patients with so called “normal breast-like” intrinsic subtypes of breast cancer, because controversy exists as to whether this pattern of gene expression is simply an artifact due to normal tissue contamination of tumor specimens (18).

All analyses were performed using JMP 9.0 (SAS Institute Inc.) and R software. The predictive model training and testing was performed using R package caret, Classification and Regression Training, R package, version 6.0-52 (17).

We examined whether HR suppression in low RPS breast cancers predicts sensitivity to DNA-damaging chemotherapies. This hypothesis is supported by the low RPS values that exist in the setting of BRCA1 or BRCA2 mutations, relative to sporadic breast tumors with wild-type BRCA1 and BRCA2 (Supplementary Fig. S5). However, breast cancers can also exhibit HR suppression and sensitivity to DNA-damaging chemotherapies, or so-called BRCAness, even in the absence of a BRCA gene mutation. Therefore, we examined several prospectively collected publicly available clinical breast cancer datasets, which were identified based on specific criteria highlighted in Supplementary Fig. S1. We specifically focused on patients whose treatment consisted of preoperative chemotherapy followed by tumor resection. This is a particularly powerful clinical setting for evaluating chemo-sensitivity of tumors, because rates of tumor responsiveness are quantified histologically using standardized methods. These criteria yielded 513 evaluable patients (5 studies summarized in Table 1 and Supplementary Table S1), all of whom received relatively uniform chemotherapy regimens that included an anthracycline, a taxane, and cyclophosphamide (18–22). Anthracycline-based regimens were deemed appropriate for this analysis given that anthracyclines intercalate between stacked DNA base pairs, resulting in double-stranded DNA breaks that involve HR for their repair (23). Importantly, data on these 513 patients were collected in accordance with prospective trials, suggesting a minimal degree of selection bias and a favorable degree of patient uniformity with respect to tumor and treatment characteristics.

We developed a method to measure RPS from tumors on a per-sample basis. This important detail enables the calculation of stand-alone scores for any given patient, rather than relying on percentiles within larger study cohorts. This method also ensures reproducibility of normalized expression signals, eliminating test set bias introduced during data preprocessing (24). This was accomplished by first constructing a frozen reference cohort, consisting of three independent clinical breast cancer datasets comprising 756 total patients (Supplementary Table S2). The expression of each RPS-defining gene was then determined for 513 study subjects, using the mean (⁠|{\rm{\mu }}$|⁠) and standard deviation (⁠|{\rm{\sigma }}$|⁠) precomputed from the frozen reference cohort to transform the gene expression signals to a z-score value. Variable importance measures were computed from optimized and validated RFs models for each RPS-defining gene (RIF1, XRCC5, RAD51, and PARPBP), together with other clinical–pathologic variables (age, T stage, N stage, grade, ER status, and HER2 status) that are commonly used to evaluate breast cancers (Supplementary Fig. S4). Notably, each of the four RPS genes and ER status ranked among the most important factors in terms of predicting pCR following chemotherapy. Among the four RPS-defining gene expression levels themselves, there were slight differences in their individual importance scores. For this reason, the definition of RPS was modified to account for the relative contribution of each gene in breast cancer (hereafter termed RPS_b to denote breast cancer-specific score): |$RPSb\ = \ - 1*( {0.2171423\$| |$*\ \ RIF1\ + \,0.1946173\ *\ PARPBP\ + \ 0.2783017\ *\ \,RAD51\ \,+$| |0.3099387\ *\ XRCC5} )$|⁠. This makes use of SCAN-normalized, z-transformed gene expression values as described in the Materials and Methods.

We evaluated the probability of achieving a pathologic complete response (pCR) to preoperative anthracycline-based chemotherapy as a function of tumor RPS_b (Fig. 1). Logistic regression analysis demonstrated an inverse relationship between RPS_b and probability of pCR (P = 0.020) in ER-negative tumors. In ER-negative tumors with RPS_b values ≤ −1.4, the probability of pCR exceeded 50%. In contrast, less than 25% of ER-negative tumors with RPS_b values ≥ 1.2 exhibited a pCR. As expected, pCR rates were considerably lower in ER-positive tumors than ER-negative tumors (10.6% vs. 35.9%, P < 0.0001). Among these ER-positive tumors, logistic regression analysis also confirmed an inverse relationship between RPS_b and probability of pCR (P = 0.023). In addition, in ER-positive tumors with RPS_b values ≤ −1.2, the probability of pCR exceeded 20%.

RPS_b was directly compared with clinical and pathologic factors that can influence response to chemotherapy, using univariate (Supplementary Fig. S6) and multivariate logistic regression analyses (Table 2). The multivariate analysis demonstrated that RPS_b (considered as a continuous variable) is an independently predictive factor for pCR rates after anthracycline-based chemotherapy (P = 0.0091). Over the range of RPS_b values (−2.7 to 1.9), there was a 1.5-fold increased multivariate odds ratio of achieving a pCR for each unit decrease in RPS_b. Therefore, RPS provides additional predictive information on tumor sensitivity which exceeds that provided by standard clinical–pathologic factors.

While pCR is a commonly utilized measure of chemo-sensitivity in breast cancer, the binary nature of this endpoint carries practical limitations. For this reason, residual cancer burden (RCB) measurements have been developed to document graded levels of pathologic tumor response. This distinction is clinically relevant, because the degree of pathologic response can influence management decisions, like the choice between breast-conserving surgery versus mastectomy. We evaluated RCB responses as a function of tumor RPS_b. We found that low RPS_b tumors are approximately twice as likely to achieve an RCB-0/I (pCR or minimal residual disease) as compared with high RPS_b tumors (Fig. 2). In ER-negative tumors, 48% of low RPS_b tumors demonstrated an RCB-0/I response as compared with 23% of high RPS_b tumors (P = 0.042). Similarly, in ER-positive tumors, 27% of low RPS_b tumors achieved an RCB-0/I response as compared with 11% of high RPS_b tumors (P = 0.032). This suggests that RPS testing may serve a useful role in both ER-negative and ER-positive tumors.

We tested the hypothesis that the predictive utility of RPS_b is specific to DNA-damaging chemotherapy by evaluating tumor response rates after neoadjuvant treatment with single-agent doxorubicin (a DNA-damaging anthracycline) versus docetaxel (a non-DNA-damaging taxane). Multivariate logistic regression analysis confirmed a significant interaction between low RPS_b value and chemotherapy type (favoring doxorubicin over docetaxel) in the prediction of RCB response (likelihood ratio χ²test, P = 0.0044). As expected, neoadjuvant doxorubicin generated RCB-0/I response rates in 25% of low RPS_b tumors versus 0% of high RPS_b tumors (Supplementary Fig. S7). By contrast, the lower RPS_b tumors did not respond better to neoadjuvant docetaxel, relative to higher RPS_b tumors. In fact, lower RPS_b tumors instead exhibited a greater resistance to docetaxel, which is consistent with published observations suggesting taxane resistance exists in BRCA-mutant tumors (25). These findings support the conclusion that the predictive power of RPS_b is specific to DNA-damaging therapies.

In clinical practice, breast cancers are commonly categorized based on ER status and HER2/neu gene amplification, and the resulting subcategories associate with varying levels of treatment sensitivity. We observed that HER2⁺ and triple-negative tumors exhibit lower RPS_b values as compared with hormone receptor–positive tumors (Supplementary Fig. S8). More recently, breast cancers have been further categorized into molecularly defined subtypes based on gene expression patterns (14). Consistent with previous reports (5), we observed variable degrees of chemo-sensitivity associated with these distinct subtypes (Fig. 3). Therefore, we investigated for a potential relationship between RPS_b and intrinsic subtypes. Among luminal subtypes, we observed significantly higher RPS_b values in luminal A tumors (mean ± standard error of mean: 0.56 ± 0.062), relative to luminal B tumors (luminal B/HER2⁻: −0.16 ± 0.087; luminal B/HER2⁺: −0.80 ± 0.25; both 2-tailed Student t test P < 0.0001). Both of the non-luminal subtypes also exhibited relatively low RPS_b values (basal: −0.10 ± 0.059; HER2-enriched: −0.12 ± 0.12) relative to luminal A tumors (both 2-tailed Student t test P < 0.0001).

Tumor subtypes with highest pCR rates (basal, HER2-enriched, and luminal B/HER2⁺) exhibit relatively low RPS values (Fig. 3). Given this apparent relationship between these two factors in breast cancer, we investigated whether the treatment-predictive power of RPS_b is independent of subtype. RPS_b was analyzed together with intrinsic breast subtype using multivariate logistic regression analyses (Table 3). This demonstrated that RPS (considered as a continuous variable) is independently predictive of pCR (considered as a binary variable) following treatment with anthracycline-based chemotherapy. Over the range of RPS_b values (−2.7 to 1.9), there was a 1.7-fold increased multivariate odds ratio of achieving a pCR for each unit decrease in RPS_b. Therefore, RPS_b provides predictive information on tumor sensitivity that exceeds existing clinical–pathologic factors or intrinsic subtyping alone.

RPS provides an estimate of DNA repair efficiency in human malignancies, wherein low RPS tumors harbor HR suppression. We previously demonstrated that low RPS breast cancers exhibit genomic instability and adverse clinical features, suggesting that HR suppression in these tumors promotes malignant progression and the emergence of aggressive phenotypes (13). We now show that low RPS_b breast tumors simultaneously exhibit a heightened sensitivity to DNA-damaging therapy, which creates a novel therapeutic opportunity. Therefore, RPS_b provides both prognostic and predictive characterization of individual breast cancers: it identifies the tumors likely to behave most aggressively, while simultaneously identifying the tumors most likely to benefit from DNA-damaging therapies.

Numerous gene expression classifier systems have been proposed to the predict sensitivity of breast cancers to anthracycline-based chemotherapy, however relatively few have undergone extensive clinical validation (26). The OncotypeDx 21-gene recurrence score (RS), which remains the most extensively validated test of its class, is indicated primarily for ER-positive cases (3, 6, 7). By comparison, RPS_b appears to be effective in both ER-positive and ER-negative cases. Additionally, our results demonstrate that RPS_b and intrinsic subtyping are independently predictive of sensitivity to anthracycline-based chemotherapy on multivariate analysis. This suggests that RPS_b testing and intrinsic subtyping may be superior to either test alone.

The low RPS_b values we observed in basal, HER2-enriched and luminal B subtypes may provide mechanistic explanations for their clinical behaviors (5) and their propensity for genomic instability (27, 28). These subtypes are known to exhibit a relatively high risk for recurrence in patients that receive no systemic therapy, suggesting that they harbor more malignant baseline characteristics (5). However, these subtypes are more likely to respond to pre-operative chemotherapy (5). Our results suggest that both of these clinical behaviors may be explained by HR suppression, and that this can be readily quantified by RPS_b testing.

Some sporadic breast cancers exhibit behaviors that are reminiscent of tumors occurring in BRCA germline mutations carriers, and this feature has been termed BRCAness (29). While BRCAness can result from somatic mutations in HR-promoting genes (30), it is more commonly identified based on characteristic patterns of resulting genomic instability. Various detection methods have been proposed, including the identification of large-segment DNA copy number alterations (31–35) or smaller DNA sequence signatures (36) that associate with HR loss. Low RPS_b measurements may share some commonality with these competing tests of BRCAness. It is worth considering, however, that these DNA-based detection methods recognize the consequences of HR deficiency; therefore, test results are not expected to recognize the eventual emergence of treatment resistance after therapy. Further work is ongoing to test whether RPS_b (which instead depends on gene expression levels) may provide a more dynamic real-time approximation of HR activity over the course of treatment.

We chose to study RPS_b in the context of commonly used anthracycline-based chemotherapy regimens because numerous large clinical datasets were available for such analyses. However, HR deficiency is more commonly associated with tumor sensitivity to other drug classes, including PARP1 inhibitors and DNA cross-linkers (10–12). Several recent studies using platinum-based chemotherapy and PARP1 inhibitors have observed correlations between measured BRCAness and tumor sensitivity (37–39). For example, the I-SPY-2 investigators recently demonstrated impressive sensitivity of triple-negative breast cancers to pre-operative therapy with veliparib plus carboplatin (39). Similarly, Telli and colleagues observed relatively high response rates to platinum-based chemotherapy in breast tumors that harbor low homologous recombination deficiency (HRD) scores (37). Further work is needed to test whether RPS_b can similarly predict efficacy for these drug classes, thereby enabling ideal selection of therapies to individual patients.

S. Pitroda, R. Weichselbaum, and P. Connell are listed as co-inventors on a patent on “Methods and Compositions Relating to Cancer Therapy with DNA Damaging Agents,” which is owned by the University of Chicago. P.P. Connell is a consultant/advisory board member for Shuwen Biotech Co. Ltd. No potential conflicts of interest were disclosed by the other authors.

Conception and design: S.P. Pitroda, P.P. Connell

Development of methodology: S.P. Pitroda, R. Bao, J. Andrade, P.P. Connell

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): S.P. Pitroda

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): S.P. Pitroda, R. Bao, J. Andrade, R.R. Weichselbaum, P.P. Connell

Writing, review, and/or revision of the manuscript: S.P. Pitroda, R. Bao, R.R. Weichselbaum, P.P. Connell

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): S.P. Pitroda

Study supervision: S.P. Pitroda, R.R. Weichselbaum, P.P. Connell

We are indebted to the curators of publicly available datasets.

This work was supported by funding from the National Institutes of Health (NIH grant 2R01CA142642, P.P. Connell), the Ludwig Foundation for Cancer Research (R.R. Weichselbaum), the Chicago Tumor Institute (R.R. Weichselbaum), a gift from the Foglia Foundation (R.R. Weichselbaum), and Stand Up To Cancer—Ovarian Cancer Research Fund-Ovarian Cancer National Alliance-National Ovarian Cancer Coalition Dream Team Translational Research Grant (grant number SU2C-AACR-DT16-15, P.P. Connell).

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.