Utilizing machine learning to identify biomarkers of chemoresistance in breast cancer: a complementary analysis with in vitro resistance platforms

Resumo

Objective: This study aimed to use a machine learning algorithm to identify biomarkers of resistance to neoadjuvant
chemotherapy (NACT) in breast cancer and validate these findings in a preliminary patient cohort, comparing results
with an in vitro resistance platform. Methods: Clinicopathological data from breast cancer samples before NACT were
analyzed, using public datasets and a proprietary database. Differential analyses compared patients with residual disease
versus pathological complete response (pCR) to NACT. The XGBoost algorithm, a tree-based machine learning technique,
and the SHAP tool were employed for interpretation. Additionally, tumor samples from patients with primary invasive
breast cancer referred to NACT were collected and cultured in a chemoresistance platform. The samples were tested with
cytotoxic drugs to classify the tumors based on cell viability. Results: These datasets included 1,012 patients exhibiting
heterogeneous data. The XGBoost algorithm achieved 82% accuracy in classifying samples into pCR and residual disease,
with SHAP analysis highlighting age, estrogen receptor status, and grade as key resistance predictors. Among ten patient
samples, five achieved pCR, one had a good response with microinvasion, and four presented poor responses. In the chemoresistance platform, patients with pCR exhibited low resistance to the drugs used in chemotherapy, and those with
poor responses demonstrated high rates of intermediate-to-high resistance to the drugs already used. Distinct resistance
patterns to treatments not used in clinics were observed, suggesting these drugs could be alternative treatment options.
The algorithm predicted NACT response with 81.8% accuracy in this cohort. Conclusion: These findings highlighted the
capacity of the XGBoost algorithm in predicting breast cancer resistance, and in combination with the chemoresistance
platform, allow the development of personalized therapeutic strategies.