Background: Precision treatments for melanomas require mutational profiling for BRAF and NRAS oncogenes. Therefore, rapid and accurate screening assays are a critical adjunct to diagnostic molecular tests. We examined whether deep machine learning computer algorithms can predict BRAF and/or NRAS mutations of digitized hematoxylin-and-eosin (H&E) slides of tumors with known mutation status. Methods: We first trained the network on H&E stained images from The Cancer Genome Atlas (TCGA) of 95 melanomas (BRAF, n = 50; NRAS, n = 13; WT/WT, n = 32). We also tested an additional cohort of 164 invasive melanomas (BRAF, n = 60; NRAS, n = 52; WT/WT, n = 52) from New York University (NYU) Melanoma database. Tumors were annotated using Aperio ImageScope software. 299x299 pixel “tiles” partitioned at 20X magnification were used to train a Convolutional Neural Network (CNN) model (Inception v3). CNNs were trained for: (1) binary classification (BRAF mutant vs. BRAF WT) on the TCGA dataset, and (2) three-way classification (BRAF mutant vs. NRAS mutant vs. WT/WT) on the NYU cohort. In each classification, 75% of the images were used for training, 15% for validation, and 15% for testing. Results: Model performance was measured by Area Under the Curve (AUC) on the held-out independent test set (Table 1). Prediction of mutated BRAF from the TCGA dataset resulted in AUC 0.852 with a sensitivity (SENS) of 0.75 and a specificity (SPEC) of 0.78. Predictive accuracy of the NYU cohort achieved AUC 0.764 for mutated BRAF (SENS 0.89 and SPEC 0.75) and AUC 0.76 for mutated NRAS (SENS 0.79 and SPEC 0.73). Conclusions: Distinct morphological changes by BRAF and NRAS mutations in melanoma can be detected on histopathological images by deep learning algorithms with robust accuracy rates among different cohorts. Additional training with larger sample size can potentially improve predictive performance.Table 1.