Human-interpretable features derived from breast cancer pathology slides detect BRCA1/2 gene mutations

BRCA1/2 gene mutations significantly increase the risk of developing breast cancer.Accurate BRCA1/2 genotyping is crucial for guiding treatment options, such as the use of poly adenosine diphosphate-ribose polymerase inhibitors (PARPi). Computational pathology research that utilizes deep learning models constructed on whole slide images (WSIs) to detect BRCA1/2 mutation status suffer from limited robustness and interpretability. We developed a Bi-directed Self-Attention MultiInstance Learning (BiAMIL) model to detect BRCA1/2 mutations using WSIs. This method employs a bi-directional self-attention mechanism to transform pixel-level features into semantic representations that offer relevant attention scores for BRCA mutation risks. By visualizing the weight maps of tumor tiles, we provided insights into the model's decision-making process. The model achieved an area under the curve (AUC) of 0.847 (95% CI, 0.707-0.987) and 0.839 (95% CI, 0.766-0.912) for the internal test set and the external test set, respectively. The visualization findings demonstrated that nuclear pleomorphism and lymphocytic infiltration are major tissue features associated with BRCA1/2 mutations, while the BRCA wild type is associated with lowgrade tumors. Our deep learning framework is an effective and interpretable method for detecting the status of BRCA1/2 from histopathological images. It has the potential to serve as an important tool with practical clinical applications.