Thirty-Gene Pharmacogenomic Test Correlates with Residual Cancer Burden after Preoperative Chemotherapy for Breast Cancer

Patients and Methods

Patients. As part of an ongoing pharmacogenomic marker validation study, patients treated with preoperative T/FAC chemotherapy were offered fine needle aspiration (FNA) of the primary breast tumor or ipsilateral axillary lymph node metastasis before starting treatment. RNA was extracted from the FNA samples, and gene expression profiling was done on Affymetrix U133A chips as previously described (10, 11). A total of 74 patients were included in the current analysis; 51 from the original validation set and 23 additional new cases. None of these patients were included in the predictor discovery. The goal of the current analysis was to examine how the pharmacogenomic prediction results correlate with pathologic response as a continuous variable. The previous study examined the accuracy of the pharmacogenomic test to predict dichotomous response outcome, pCR versus RD. All patients were treated with preoperative weekly or 3-weekly paclitaxel × 12 weeks, followed by four courses of fluorouracil, doxorubicin, and cyclophosphamide (FAC). When the tumor becomes <2 cm on physical examination or imaging, metallic coils were placed in the tumor bed under ultrasound guidance to identify the primary tumor site at the time of surgery. Definitive breast surgery (mastectomy or breast conserving surgery) followed by axillary dissection or sentinel lymph node biopsy was determined by the surgeon's discretion and patient's preference. This clinical study was approved by the institutional review board of M. D. Anderson Cancer Center, and informed consent was signed by all patients.

Pathologic response assessment. H&E-stained slides from the surgical resection specimen were reviewed microscopically to evaluate RCB. Briefly, RCB in the posttreatment surgical resection specimen is determined from bidimensional diameters of the primary tumor bed in the resection specimen (d₁ and d₂), the proportion of the primary tumor bed that contains invasive carcinoma expressed as % cellularity (f_inv), the number of axillary lymph nodes containing carcinoma (LN), and the diameter of the largest metastasis in the involved axillary lymph node (d_met). Largest bidimensional measurements of the residual tumor mass in the breast were recorded from the macroscopic description in the pathology report and were confirmed after review of the corresponding slides. If multiple tumor masses were present, the dimensions of the largest mass were recorded. The bidimensional measurements of the primary tumor bed were combined as follows: d_prim = . To estimate the proportion of invasive carcinoma (f_inv) within the cross-sectional area of the primary tumor bed, we first estimated the overall percent area of carcinoma (%CA) from microscopic review of the corresponding slides and then corrected for the component of in situ carcinoma (%CIS): f_inv = (1 − (%CIS)/100) × (%CA)/100. To estimate the extent of residual cancer in the regional lymph nodes, we reviewed the slides from resected axillary lymph nodes and recorded the number of lymph nodes that contained a metastasis (LN), and the greatest diameter of the largest metastasis (d_met). All diameters were recorded in millimeters. A free, interactive, online RCB calculator is available in a web site⁶ (,9).

We previously defined two cutoff points for the RCB score that resulted in four pathologic response categories that correlated with survival in 241 T/FAC-treated patients (9). None of the current 74 patients were included in the previous analysis that defined the RCB cutoff values. To determine the first cutoff point, we fitted a multivariate Cox regression model that included all the clinical covariates and a dichotomized (above or below the cutoff) RCB score. The optimal RCB cutoff point was selected that maximized the profile log-likelihood of the Cox model. This occurred at the value of 3.28 that corresponds to the 87.5th quantile of the RCB distribution. A second cutoff point was determined similarly by maximizing the log-likelihood of a Cox model that included all clinical covariates and the second dichotomized RCB factor (above or below the second cutoff). This cutoff point occurred at the value of 1.36 or the 40th quantile of the RCB distribution. These two cutoffs define four groups: (a) RCB-0 (RCB score 0), corresponding to pCR defined as no residual invasive carcinoma in the breast and axillary lymph nodes; (b) RCB-I (RCB scores 0 < RCB ≤ 1.338), corresponding to minimal residual invasive disease; (c) RCB-II (RCB scores 1.338 < RCB ≤ 3.278), corresponding to moderate RD; and (d) RCB-III (RCB scores >3.278), corresponding to extensive RD. These four different categories of residual cancer had very different survival (9). The rates of the 5-year relapse-free–survival were 95% in the RCB-0 and 98% in the RCB-I groups. In contrast, these were 84% in the RCB-II and 46% in the RCB-III groups (P < 0.001 in the log-rank test) The RCB categories I, II, and III were all lumped together as RD in our previous analysis (10).

Statistical analysis. The previously reported and fully defined pharmacogenomic response predictor was applied to the normalized gene expression data from each case. This predictor takes the expression values of 30-probe sets and applies diagonal linear discriminant analysis prediction rules to generate a prediction score. Prediction scores >0 predict for pCR, and scores <0 (i.e., negative scores) predict for RD. Pearson's correlation was done to determine the correlation between the 30-gene predictor scores and the RCB scores and the RCB categories. The one-way ANOVA test was used for the comparison of means of the 30-gene predictor scores between various RCB categories. Paired two-way comparison t tests were used to calculate differences among these groups. All tests were two tailed, with a P value of <0.05 considered significant. The SPSS 12.0 software package (SPSS Inc.) was used for statistical analyses.

Results

Correlation between the pharmacogenomic prediction scores and the RCB scores. Gene expression results and pCR versus RD outcome information were available on all 74 cases. RCB categories (0, I, II, or III) could be assigned to 73 patients and RCB scores to 72 patients due to missing slides for review. One of the two patients with missing RCB scores had progression on preoperative chemotherapy and was categorized as RCB III. Patient and tumor characteristics are shown in Table 1 . The mean 30-gene predictor score was −0.60 (range, 54.35 to −42.14) and the mean RCB score was 1.80 (range, 0 to 4.45). The correlation between the 30-gene predictor scores and the RCB scores showed a Pearson's R = −0.501 (P < 0.0001; Fig. 1A ). Similar correlation was found between the 30-gene predictor scores and the RCB categories (Pearson's R = −0.506, P < 0.0001; Fig. 1B). One-way ANOVA analysis showed that the mean 30-gene predictor scores were significantly different between the four RCB categories (P < 0.0001; Fig. 2A and B ; Table 2 ). In the two-way comparison t tests, the mean 30-gene predictor score in RCB-0 cases was significantly different from the mean values in RCB-II and RCB-III (P = 0.0005 and P = 0.0002, respectively). However, there was no significant difference between the mean 30-gene predictor scores for RCB-0 and RCB-I response categories (P = 0.94), and there was also no significant difference between the predictor scores for RCB-II and RCB-III response (P = 0.14). These results indicate that some of the “false-positive” pCR prediction results from the pharmacogenomic test include patients who have minimal RD and, therefore, may benefit from chemotherapy to the same extent as patients with pCR.

• Download figure
• Open in new tab
• Download powerpoint

Fig. 1.

A, Correlation between the 30-gene predictor scores and the RCB scores. B, correlation between the 30-gene predictor scores and the RCB categories.

• Download figure
• Open in new tab
• Download powerpoint

Fig. 2.

A, 30-gene predictor scores corresponding to RCB categories. B, distribution of the 30-gene predictor scores (means).

Correlation between pharmacogenomically predicted response and RCB categories. RCB-0 response (pCR) was observed in 20 (27.4%) cases, RCB-I was observed in 5 (6.9%), RCB-II was observed in 36 (49.3%), and RCB-III was observed in 13 (16.4%) patients. A total of 33 patients were predicted to have pCR, and 40 were predicted to have RD by the pharmacogenomic test. The overall agreement between the 30-gene predictor class (pCR versus RD) and the RCB classes (RCB-0 versus RCB-I, RCB-II, and RCB-III) was 70.2% (Table 3 ). The 30-gene predictor predicted RCB-0 outcome as a pCR in 80%, RCB-II and RCB-III were predicted as RD in 64% and 92%, respectively. RCB-I (near-pCR) was predicted more frequently (60%) as pCR than as RD (40%). Among the 25 patients with complete (RCB-0) or near-complete (RCB-I) pathologic response who are projected to have excellent survival, 19 (76%) were predicted to achieve pCR by the pharmacogenomic test.

Discussion

In this study, we examined the correlation between a 30-gene pCR predictor score determined before treatment and the extent of residual invasive cancer measured after completion of 6 months of preoperative chemotherapy. Residual invasive cancer was measured as a continuous variable using a recently developed RCB score. Our findings showed a correlation between the pharmacogenomic predictor scores that were determined at diagnosis and the RCB scores determined after definitive surgery (Pearson's R = −0.501, P < 0.0001). The RCB scores can be used to define four categories of pathologic response (RCB-0, RCB-I, RCB-II, and RCB-III) with distinct long-term survival outcome (9). The 30-gene pharmacogenomic predictor correctly predicted 80% (n = 16/20) of the patients with RCB-0 (pCR) response. About 60% of cases (3/5) with RCB-I (minimal RD) were also predicted as pCR. These patients are considered as false-positive cases using the dichotomous pCR versus RD classification; however, their long-term survival is projected to be very similar to those with pCR. There was no significant difference in the mean pharmacogenomic prediction scores of the RCB-0 (mean, 15.12; SD, 20.9) and RCB-I (mean, 15.93; SD, 30.4) response categories. About 48% of cases with moderate amount of RD, RCB-II (mean prediction score, −6.2; SD, 17.5) were misclassified as pCR by the pharmacogenomic test, but only 1 of 12 (8%) patients with extensive residual cancer, RCB-III (mean prediction score −14.6, SD 17.1), was misclassified as pCR. These results suggest that the genomic test correctly identifies ∼76% (n = 19/25) of patients who will achieve either pCR (RCB-0) or near-pCR (RCB-I) to preoperative T/FAC chemotherapy. Equally importantly, 92% of those who will end up with extensive RD (RCB-III) are correctly identified as RD by the test.

These results corroborate our previous finding that the 30-gene pharmacogenomic test predicts chemotherapy sensitivity (12). The pharmacogenomic prediction scores inversely correlate with the extent of residual invasive cancer (i.e., RCB score). The lower the prediction score, the higher is the residual invasive cancer burden (Fig. 2).

Other investigators also reported that it is possible to identify gene signatures that are associated with pathologic response to chemotherapy (13–20). However, no other proposed response predictors were assessed in independent validation, and no detailed analysis of response prediction scores and extent of residual cancer was previously reported. The observation that a 30-gene pharmacogenomic prediction score, which was optimized to predict dichotomous outcome (i.e., pCR versus RD), correlated with the extent of residual invasive cancer when cancer was present increases our confidence in this genomic test. Our data add to the growing body of literature that indicates that there are large-scale gene expression differences between highly chemotherapy-sensitive and less sensitive cancers. Whether these emerging pharmacogenomic predictive tests can improve medical decision making and lead to improved patient outcome needs further investigation.