Chemo-prAIdict Breast: A deep learning solution for predicting residual disease on biopsies of breast cancer patients treated with neoadjuvant chemotherapy.

[BACKGROUND] Predicting chemosensitivity before treatment could help tailor neoadjuvant chemotherapy (NAC) in early breast cancer (eBC). Pathological complete response (pCR) is associated with better long term survival, but yet no robust baseline predictor is available.

[PATIENTS AND METHODS] We developed Chemo-prAIdict Breast, a deep learning model using whole slide images (WSIs) of diagnostic biopsies to predict residual disease (RD) after NAC. Two large French cohorts were analyzed (n = 1140 initially included, 928 analyzed after selection): the prospective multicenter PRIMUNEO cohort (n = 500, 438 after selection) for training and internal validation, and the CGFL retrospective cohort (n = 640, 490 after selection) for independent external validation. Patients were stratified by molecular subtype: HER2-amplified (HER2 +), ER-positive/HER2-negative (ER+/HER2 -), and triple-negative (TN).

[RESULTS] In external validation, Chemo-prAIdict Breast outperformed standard clinicopathological features, achieving AUCs of 0.652 (p = 0.001) in HER2 + , 0.814 (p = 0.003) in ER+ /HER2 -, and 0.677 (p = 0.001) in TN tumors. Robustness was confirmed using paired consecutive biopsy sections from 421 patients: predictions were strongly correlated within patients (Pearson r = 0.933 for HER2 +, 0.932 for ER+/HER2 -, 0.939 for TN; all p < 0.001).

[CONCLUSIONS] While prospective studies with modern treatment regimens are needed to establish clinical utility, Chemo-prAIdict Breast is a new tool for identifying eBC that are differentially sensitive to standard NAC, and could help to select the most appropriate treatment strategy in HER2 + , ER+ /HER2- and TN eBC.