Prediction of clinicopathological features, multi-omics events and prognosis based on digital pathology and deep learning in HR(+)/HER2(−) breast cancer

BACKGROUND: Breast cancer has the highest incidence and mortality rates among women worldwide. Hormone receptor (HR)(+)/human epidermal growth factor receptor 2 (HER2)(−) breast cancer is the most common molecular subtype, accounting for 50–79% of breast cancers. Deep learning has been widely used in cancer image analysis, especially for predicting targets related to precise treatment and patient prognosis. However, studies focusing on therapeutic target and prognosis predicting in HR(+)/HER2(−) breast cancer are lacking. METHODS: This study retrospectively collected hematoxylin and eosin (H&E)-stained slides of HR(+)/HER2(−) breast cancer patients between January 2013 and December 2014 at Fudan University Shanghai Cancer Center (FUSCC) and scanned to generate whole-slide images (WSIs). Then, we built a deep-learning-based workflow to train and validate model to predict clinicopathological features, multi-omics molecular features and prognosis; the area under the curve (AUC) of the receiver operating characteristic (ROC) and the concordance index (C-index) of the test set were used to assess model effectiveness. RESULTS: A total of 421 HR(+)/HER2(−) breast cancer patients were included in our study. Regarding clinicopathological features, grade III could be predicted with an AUC of 0.90 [95% confidence interval (CI): 0.84–0.97]. Regarding somatic mutations, TP53 and GATA3 mutation could be predicted with AUCs of 0.68 (95% CI: 0.56–0.81) and 0.68 (95% CI: 0.47–0.89), respectively. Regarding gene set enrichment analysis (GSEA) pathways, the G2-M checkpoint pathway was predicted with an AUC of 0.79 (95% CI: 0.69–0.90). Regarding markers of immunotherapy response, intratumoral tumor-infiltrating lymphocytes (iTILs), stromal tumor-infiltrating lymphocytes (sTILs), CD8A, and PDCD1 were predicted with AUCs of 0.78 (95% CI: 0.55–1.00), 0.76 (95% CI: 0.65–0.87), 0.71 (95% CI: 0.60–0.82), and 0.74 (95% CI: 0.63–0.85), respectively. In addition, we found that the integration of clinical prognostic variables and deep features of images can improve the stratification of patient prognosis. CONCLUSIONS: Using a deep-learning-based workflow, we developed models to predict the clinicopathological features, multi-omics features and prognosis of patients with HR(+)/HER2(−) breast cancer using pathological WSIs. This work may contribute to efficient patient stratification to promote the personalized management of HR(+)/HER2(−) breast cancer.