Deep learning infers clinically relevant protein levels and drug response in breast cancer from unannotated pathology images

Abstract

AbstractThe computational pathology has been demonstrated to effectively uncover tumor-related genomic alterations and transcriptomic patterns. Although proteomics has indeed shown great potential in the field of precision medicine, few studies have focused on the computational prediction of protein levels from pathology images. In this paper, we assume that deep learning-based pathological features imply the protein levels of tumor biomarkers that are indicative of prognosis and drug response. For this purpose, we propose wsi2rppa, a weakly supervised contrastive learning framework to infer the protein levels of tumor biomarkers from whole slide images (WSIs) in breast cancer. We first conducted contrastive learning-based pre-training on tessellated tiles to extract pathological features, which are then aggregated by attention pooling and adapted to downstream tasks. We conducted extensive evaluation experiments on the TCGA-BRCA cohort (1978 WSIs of 1093 patients with protein levels of 223 biomarkers) and the CPTAC-BRCA cohort (642 WSIs of 134 patients). The results showed that our method achieved state-of-the-art performance in tumor diagnostic tasks, and also performed well in predicting clinically relevant protein levels and drug response. To show the model interpretability, we spatially visualized the WSIs colored the tiles by their attention scores, and found that the regions with high scores were highly consistent with the tumor and necrotic regions annotated by a 10-year experienced pathologist. Moreover, spatial transcriptomic data further verified that the heatmap generated by attention scores agrees greatly with the spatial expression landscape of two typical tumor biomarker genes. In predicting the response to drug trastuzumab treatment, our method achieved a 0.79 AUC value which is much higher than the previous study reported 0.68. These findings showed the remarkable potential of computational pathology in the prediction of clinically relevant protein levels, drug response, and clinical outcomes.