Transcriptomic classifiers do not reliably predict lymph node involvement in breast Cancer: Insights from three large cohorts

Lymph node (LN) status remains one of the most critical prognostic indicators in breast cancer. A reliable molecular surrogate for LN involvement could offer substantial clinical benefits, including improved risk stratification and more selective use of surgical staging. This study investigates the potential of mRNA expression profiles to serve as predictors of LN status.

mRNA expression data from three large cohorts, TCGA (n = 749), METABRIC (n = 1895), and SCAN (n = 3273), were used to generate six train-test permutations. Two machine learning algorithms, AdaBoost and XGBoost (gradient boosting), were applied to classify LN status and Nottingham histological grade (used as a benchmark). A total of 15,707 genes were assessed as predictors. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC). In a separate analysis, time-to-event data (SurvObj) were modelled using the FastCox library in R to generate 1000 random survival models. The survival probabilities generated were then repurposed to predict LN status and grade.

For LN prediction, the highest AUC achieved was 0.613 using XGBoost and 0.604 using AdaBoost. In contrast, histological grade molecular prediction yielded significantly higher accuracy, with maximum AUCs of 0.901 (XGBoost) and 0.880 (AdaBoost). Cramér's V analysis across all genes showed limited concordance with LN status. Additionally, models derived from survival probabilities consistently yielded AUCs below 0.6 when applied to LN prediction, indicating that strong survival predictors did not translate into accurate LN classifiers.

Despite utilizing a comprehensive transcriptomic dataset and advanced machine learning algorithms, mRNA expression alone did not reliably predict LN status. Currently, no mRNA-based model serves as a clinically viable alternative to surgical LN assessment. Future efforts should explore integrative approaches that combine molecular, spatial, and clinical data to enhance our understanding and prediction of lymphatic dissemination in breast cancer.