Multi-attribute embedding network for breast cancer subtyping prediction reveals intrinsic molecular properties hidden in H&E images

Molecular subtyping plays a critical role in breast cancer management, yet immunohistochemistry (IHC) methods for detecting protein expression levels are hindered by time constraints and interpretational challenges. The association between molecular expression and pathological morphology suggests the potential to predict subtypes from haematoxylin and eosin (H&E) images, offering deeper molecular insights beyond protein expression. However, acquiring molecular information from whole slide images (WSIs) of breast H&E slides remains challenging due to gigapixel resolution and weak labelling. Here, a weak-supervised learning framework based on multi-attribute feature embedding for IHC subtype prediction directly from breast H&E images is presented. Leveraging a multi-centre breast cancer dataset comprising 3,721 H&E images from 1,688 patients, the model achieves the highest mean area under the ROC curve (AUC=0.868) compared to state-of-the-art methods. Validation through IHC testing on 20 cases demonstrates the potential clinical utility of pathological deep features. Patch probability visualization reveals potential morphological associations with molecular subtypes, which generally aligns with protein expression truth in IHC images. Furthermore, multi-omics analysis involving pathological deep features facilitates the exploration of important molecular properties. Global feature visualization elucidates pathological heterogeneity, while differential gene expression identifies potential gene markers and functional heterogeneity among breast cancer subtypes. By combining pathology and transcriptome analysis, the study reveals correlations between pathological features and gene expression. Finally, survival analysis suggests that subtyping probabilities and pathological deep features are potential survival indicators for clinical prognosis.