Single‑center weakly supervised deep learning prediction of KRAS, NRAS, BRAF, and HER2 status in colorectal cancer from histopathology images using internal cross‑validation

Research has shown that mutations in the KRAS, NRAS, and BRAF genes are linked to resistance to anti-EGFR therapies in colorectal cancer (CRC) patients. HER2-targeted therapies are increasingly being recommended for individuals with HER2 overexpression. The evaluation of KRAS, NRAS, BRAF, and HER2 statuses has become an important part of precise diagnosis for CRC. However, conventional molecular or protein testing can be time-consuming and expensive. This study aims to predict the status of KRAS, NRAS, BRAF, and HER2 through the analysis of whole-slide pathology features from CRC samples stained with Hematoxylin-Eosin (H&E) for KRAS, NRAS, and BRAF, and by utilizing Immunohistochemistry (IHC) for HER2. In this study, 435 CRC patients were enrolled from Jiangsu Province Hospital of Chinese Medicine. Using the clustering-constrained attention-based multiple-instance learning (CLAM) model, we constructed four models for predicting the statuses of KRAS, NRAS, BRAF, and HER2 based on whole-slide images (WSIs). This single‑center study used patient‑level internal cross‑validation to train and evaluate weakly supervised CLAM models for predicting KRAS, NRAS, BRAF, and HER2 status from whole‑slide images. The mean area under the receiver operating characteristic (ROC) curve (AUC) values (95% CI) were KRAS 0.8958 (0.8575, 0.9340), NRAS 0.9367 (0.8893, 0.9829), BRAF 0.9876 (0.9744, 1.0000), and HER2 3 + versus non‑3 + 0.99 (0.98–1.00). Given the extremely small NRAS+ (n = 14) and BRAF+ (n = 21) cohorts, these estimates are statistically fragile and should be interpreted as hypothesis‑generating pending external validation. Our model-generated heatmaps showing KRAS, NRAS, BRAF mutation patterns and HER2 expression levels generally matched the regions identified by the pathologists. This method provides new insights to predict gene mutations and protein expression using deep learning. This single-center study used patient-level internal cross-validation. Robustness and clinical applicability cannot be assumed without external, multi-center validation, and the present results should be interpreted as hypothesis-generating.