Abstract No. 
2009 San Antonio Breast Cancer Symposium
December 10-13

Breast Cancer Molecular Subtypes Predict Response to Anthracycline/Taxane-Based Chemotherapy

Parker J, Prat A, Cheang M, Lenburg M, Paik S, Perou C

Background: Classifying breast cancer into molecular subtypes (Luminal A/B, Her2-enriched, and Basal-like) has proven to be biologically and clinically informative. Recently, we developed a gene expression based protocol (PAM50) that provides reliable subtype classification as well as independent prognostic utility (Parker et al, J Clin Oncol 2009). In this study, we evaluated the ability of the PAM50 subtypes to predict pathological complete response (pCR) to anthracycline/taxane-based chemotherapy in the neoadjuvant setting.

Methods: Gene expression data were combined from three independent neoadjuvant studies totaling 361 subjects. Complete clinical data (hormone receptor levels, tumor size, node status, and grade) were available for 186 subjects. The I-SPY cohort (n=129; fresh-frozen, Agilent) was treated with doxorubicin/cyclophosphamide (ACx4) followed by paclitaxel (Tx4). The NSABP B-27 cohort (n=103; FFPE, Affymetrix) was treated with ACx4 followed by docetaxel (4 cycles). The third cohort from MD Anderson Cancer Center (MDACC) (n=129; fresh-frozen tissue, Affymetrix) was treated with paclitaxel/T followed by sequential fluorouracil-AC (FAC). Subtype assignments and prognosis scores were generated as previously described (Parker et al., 2009). Univariate and multivariate testing for pCR was performed with Fisher's exact test and logistic regression. Statistical learning methods were also implemented in order to identify an improved predictor of pCR.

Results: There was no difference between the cohorts/studies with respect to estrogen receptor (ER) status (p=0.23), subtype distribution (p=0.77), or pCR rate (p=0.95). ER, PR, HER2, grade, node status, and subtype were all associated with pCR in univariate analyses. Multivariable logistic regression indicated that molecular subtype is an independent predictor of pCR (p<0.01). We found that molecular subtype contains additional pCR-predictive information beyond that of ER status, and that adding ER status to models containing molecular subtype did not significantly improve the model's performance. This indicates that the information predictive of pCR represented by ER status is largely also present in molecular subtype. Among the subtypes, Basal-like and Her2-enriched tumors demonstrated the highest rate of pCR relative to Luminal A tumors (OR 7.4; 2.2-24.8; OR 6.4; 2.0-20.7). The Risk of Relapse based on the subtype (ROR-S) was also a strong predictor of pCR (auROC=0.74), but was improved when using a novel ridge regression model (auROC=0.77; p<0.01). ROR-S and the different pCR predictors evaluated were generally correlated, however, Luminal B samples were consistently classified as having a poor prognosis and a low likelihood of pCR.

Conclusions: In this combined analysis, using ER, PR, HER2, node status, tumor size, and subtype, the intrinsic molecular subtype classification is the most significant predictor of pCR across three anthracycline/taxane-based neoadjuvant cohorts.Interestingly, hormone receptor status is no longer significant when subtype is included in the model. These results indicate robust predictive information provided by the subtype classification, and highlights the chemotherapy insensitivity of the poor prognosis luminal B breast cancer subtype.

View original