Integrating Deep Learning and SHAP for Breast Cancer Classification and Biomarker Discovery Using Gene Expression Data

This study explores the application of deep learning models combined with SHAP (SHapley Additive exPlanations) for breast cancer classification using gene expression data. Our model demonstrated exceptional performance, achieving a mean accuracy of 0.9964 across 5-fold cross-validation and perfect scores through ensemble learning. To further validate the model’s generalizability, we tested it on three independent datasets (GSE45827, GSE7904, and GSE42568), where it achieved an accuracy of 0.9867, 1.0 and 0.9914. SHAP analysis provided valuable insights into gene importance, identifying key genes such as DSCAM-AS1, KRT19, ESR1, and KRT5 as significant contributors to cancer prediction. Validation against the MalaCards database confirmed the relevance of most top-ranked genes, while novel candidates were also identified. Notably, some well-known breast cancer-related genes were not top-ranked, likely due to dataset limitations. This research underscores the potential of combining deep learning and SHAP for precise gene ranking and the discovery of novel biomarkers.