A Robust Predictor of Response to Preoperative Chemotherapy for Breast Cancer

Abstract

Breast cancer patients with pathological complete response (pCR) to taxane and anthracycline containing preoperative chemotherapy tend to have excellent distant-free and overall survival. However, most of current pCR predictors for guiding chemotherapy tend to perform poor in interlaboratory validation, most likely due to microarray experimental batch effects and insufficient training samples for feature selection. To tackle this difficulty, by merging three separate datasets of patients treated with paclitaxel, fluorouracil, doxorubicin, and cyclophosphamide, we extracted gene pairs with consistent relative expression ordering in patients not achieving pCR and with relative expression reversal in patients achieving pCR. Then, based on combination of these pairs, a pCR predictor was built. The performance of this predictor achieved a sensitivity of 93% and specificity of 62% in an independent dataset, much better than the performances of three previously proposed predictors. In conclusion, the rank-based pCR predictor derived from a large cohort of samples can accurately and robustly predict patients with high probability of achieving pCR after chemotherapy.