PT - JOURNAL ARTICLE AU - Hongming Xu AU - Sunho Park AU - Jean René Clemenceau AU - Nathan Radakovich AU - Sung Hak Lee AU - Tae Hyun Hwang TI - Deep learning approach to predict tumor mutation burden (TMB) and delineate its spatial heterogeneity from whole slide images AID - 10.1101/554527 DP - 2020 Jan 01 TA - bioRxiv PG - 554527 4099 - http://biorxiv.org/content/early/2020/07/20/554527.short 4100 - http://biorxiv.org/content/early/2020/07/20/554527.full AB - Purpose Tumor Mutation Burden (TMB) is a potential genomic biomarker that could help to identify patients benefiting from immunotherapy across many cancers. Various tissue-based sequencing approaches have been widely used to determine the TMB status. However, the clinical utility of these approaches is often limited, due to time, cost, and tissue availability constraints. These methods could also provide inconsistent TMB status driven by spatial intratumor heterogeneity. Hematoxylin and Eosin (H&E) stained whole slide images (WSI) are routinely used for cancer diagnosis, thus mostly available for cancer patients. We present a deep learning based computational pipeline using WSIs for predicting patient-level TMB status and quantifying its spatial heterogeneity within tumor regions.Experimental Design In an experiment to predict patient-level TMB status, we used The Cancer Genome Atlas (TCGA) Urothelial Bladder Carcinoma (BLCA) and Lung Adenocar-cinoma (LUAD) cohorts. To investigate spatial heterogeneity of TMB status within a tumor and its prognostic utility, we used TCGA BLCA cohort.Results In an experiment of patient-level TMB predictions, our proposed method achieved the Area Under ROC (AUROC) scores of 0.752 (95% CI, 0.683-0.802) and 0.742 (95%, 0.682-0.794) for TCGA BLCA and LUAD cohorts, respectively, which are better compared to those of state-of-the-art methods. In another experiment to investigate spatial heterogeneity of TMB per patient, we predicted TMB status for each tile in each WSI for patients from TCGA BLCA cohort. We calculated entropy of TMB prediction probabilities in the WSI to determine whether the patient has high or low spatial TMB heterogeneity within the tumor. Kaplan Meier (KM) analysis showed that incorporating spatial heterogeneity of TMB information with patient-level TMB status based on WSIs could improve identification of patient subgroups with distinct survival outcome (a log rank test P<0.05).Conclusions Our proposed deep learning based approach using WSIs can predict patient-level TMB status with good accuracy, sensitivity and specificity compared to state of the art methods. The spatial analysis of TMB heterogeneity could provide a prognostic utility to better select patient subgroups.Code Availability https://github.com/hwanglab/tcga_tmb_predictionCompeting Interest StatementThe authors have declared no competing interest.