Building a prediction model of pathological complete response in breast cancer patients undergoing neoadjuvant chemotherapy

Aim of the study Building a model to predict the pathological complete response (pCR) of breast cancer patients after neoadjuvant chemotherapy (NAC). Introduction Breast cancer (BC) remains the leading cause of cancer-related deaths among women worldwide. The treatment of BC is complex, with multiple strategies tailored to the tumor’s subtype, stage, and molecular characteristics. These strategies include adjuvant and neoadjuvant approaches. In the adjuvant setting, systemic therapies like chemotherapy, hormone therapy, or targeted treatments are administered after surgical removal of the tumor to reduce the risk of recurrence. In contrast, in the neoadjuvant setting, treatments such as neoadjuvant chemotherapy (NAC) are administered before surgery to shrink the tumor, potentially allowing for breast-conserving surgery and providing an early assessment of treatment response. NAC is used in cases of locally advanced and potentially inoperable breast cancer to make the disease amenable to resection, but it can also be used in cases of operable tumors to downstage disease in the breast and axilla. Among these strategies, the optimal treatment for each individual patient remains an ongoing area of research due to the heterogeneity of breast cancer. Pathological complete response (pCR) following NAC has emerged as a surrogate endpoint for long-term survival outcomes, such as invasive disease-free survival (iDFS) and overall survival (OS). Consequently, developing models that can accurately predict pCR could guide personalized treatment strategies, improving patient outcomes. Methods Data of breast cancer patients treated at IRCCS Humanitas were prospectively collected. These included: clinical data, radiomics applied to the breast lesion on the PET/CT scan, gene expression levels from the tumor tissue, analysis of tumoral microbiota, and microbiota data from stool and buccal swabs. Features were selected using LASSO regression, followed by backward elimination via logistic regression. The selected features were used to train different models, that were evaluated via crossvalidation with ROC-AUC. Results Of 108 selected patients, 51 had pCR (47%). The best model in terms of efficacy and interpretability was a logistic regression classifier that reached an average AUC of 0.93. The selected features included 9 genes and 3 bacteria: CDHR5, CIAO3, EIF4E3, IGKV1-17, LTBP2, RNF170, LINC01126, ENSG00000244245, ENSG00000279687, Parascardovia denticolens (buccal), Streptococcus vestibularis (buccal) and Paratractidigestivibacter faecalis (stool). Conclusion Gene expression data from the tumor tissue and microbiota from buccal swabs and stool samples can be helpful in predicting the pCR of breast cancer patients. By tuning the hyperparameters and setting the threshold so to have enough patients with low probability of pCR, the model can provide good Negative Predictive Value, thus allowing clinicians to detect more easily patients that would not benefit from receiving NAC.