e23529 Background: Soft tissue sarcomas (STS) are histologically and molecularly heterogenous with variable biologic behavior and differential therapeutic responses. Molecular analysis of STS is commonly utilized to identify targetable mutations or identify STS subtypes that have predictive and prognostic implications. To address the challenge of rapid, accurate classification of STS subtypes, we sought to develop an approach using a form of artificial intelligence called deep learning to identify meaningful hidden information directly from digital H+E slide images. We trained a deep learning model using genotype and phenotype data from The Cancer Genome Atlas (TCGA) and companion diagnostic digital H+E slide images, in order to identify therapeutic biomarkers and prognostic molecular subtypes. Methods: Digital whole slide images (WSIs) from TCGA were downloaded and regions of tumor were annotated by pathologists. WSIs were carved into 604 um by 604 um square tiles within the tumor region and saved at an image size of 598 px by 598 px for a resolution of ~1um/px. These tiles were used as input to a deep convolution neural network (Xception model pretrained on imagenet). Performance results are reported at the as the average of three-fold cross-validation, averaged at the patient level. Results: 206 sarcomas available through the TCGA were used as training set. Sarcoma histologies include 80 leiomyosarcoma (LMS) (27 uterine LMS and 53 non-uterine LMS), 50 dedifferentiated liposarcoma (DDLPS), 44 undifferentiated pleomorphic sarcoma (UPS), 72 myxofibrosarcoma (MFS), and 10 synovial sarcoma (SS). FNCLCC grade 1 (n = 14), grade 2 (n = 112), and grade 3 (n = 80). Histology could be predicted automatically with very high accuracy for LMS (Area under Receiver-operator curve (AUROC) 0.93 and 0.95 for uterine and non-uterine LMS respectively), DDLPS (AUROC 0.70), UPS (AUROC 0.76), and MFS (AUROC 0.76). FNCLCC grade could be predicted for grade 3, 2, 1 with AUROC of 0.79, 0.73, and 0.81 respectively. High vs. low mitotic rate (threshold 15) could be predicted with AUROC of 0.70. Within non-uterine LMS, prognostic molecular cluster could be predicted with AUC of 0.62 for cluster with worse recurrence-free and disease specific survival. Within DDLPS, DNA methylation cluster 2 associated with worse disease specific survival was predicted with AUC of 0.80. Conclusions: Deep learning algorithm analysis of histologic slides was able to predict clinically meaningful characteristics and molecular clusters with reasonable accuracy. Retrospective and prospective validation is needed.