Machine learning-based prediction of clinical outcomes in luminal b breast cancer using somatic variant signatures

Abstract

The application of machine learning techniques in oncology has enabled the integration of complex genomic data for predicting clinical outcomes. In Luminal B breast cancer, high tumor heterogeneity poses a major challenge for prognostic stratification and therapeutic decision-making. This study aimed to develop predictive models based on somatic variants to assess tumor aggressiveness. A machine learning pipeline was implemented using supervised algorithms, including XGBoost, Support Vector Machine, and Artificial Neural Networks. Feature selection was based on predictive importance, prioritizing biologically relevant genes. Model performance was evaluated using metrics such as area under the ROC curve, sensitivity, specificity, and F1-score. The results demonstrated high predictive performance, with XGBoost achieving the best results (AUC = 0.88), followed by Neural Networks (AUC = 0.87) and SVM (AUC = 0.85). Interpretability analysis revealed that genes such as PIK3CA, TP53, and ERBB2 were the main contributors to model predictions. These findings highlight the potential of machine learning approaches in identifying genomic patterns associated with tumor aggressiveness, supporting precision medicine strategies.