Prospective comparison of clinical and genomic multivariate predictors of response to neoadjuvant chemotherapy in breast cancer.

PURPOSE

Several different multivariate prediction models using routine clinical variables or multigene signatures have been proposed to predict pathologic complete response to combination chemotherapy in breast cancer. Our goal was to compare the performance of four conceptually different predictors in an independent cohort of patients.

EXPERIMENTAL DESIGN

Gene expression profiling was done on fine-needle aspirations of 100 stage I to III breast cancers before preoperative paclitaxel, 5-fluorouracil, doxorubicin, and cyclophosphamide combination chemotherapy. Pathologic response was correlated with prediction results from a clinical nomogram, a human cancer-derived genomic predictor (DLDA30), a cell line-based genomic predictor [in vitro coexpression extrapolation (COXEN)], and an optimized cell line-derived (in vivo COXEN) predictor. None of the 100 test cases were used in the development of these predictors.

RESULTS

The in vitro COXEN using a combination of four individual drug sensitivity predictions derived from cell lines was not predictive [area under the receiver operator characteristic curve (AUC), 0.5; 95% confidence interval, (95% CI), 0.41-0.59]. The clinical nomogram (AUC, 0.73; 95% CI, 0.65-0.80) and the DLDA30 (AUC, 0.73; 95% CI, 0.66-0.80) genomic predictor had similar performances. The in vivo COXEN that used informative genes from cell lines but was trained on a separate human data set also showed significant predictive value (AUC, 0.67; 95% CI, 0.60-0.74). These three different prediction scores correlated with each other and were significant in univariate but not in multivariate analysis.

CONCLUSIONS

Three conceptually different predictors performed similarly in this validation study and tended to identify the same patients as responders. A genomic predictor that relied solely on a composite of individual drug sensitivity predictions from cell lines did not show any predictive value.