<p>Cancer classification traditionally focuses on tissue of origin and histology appearance. Traditional “hard” classification for a specific diagnostic class yields binary yes/no class prediction. However, newer therapeutic approaches have demonstrated that tumors overlap biologically and may respond to a specific therapeutic approach regardless of traditional histological classifications. Instead of histological classification, we propose the use of a “Functional classification” based on machine learning models, leveraging high-dimensional RNA expression data. This approach identifies not only the predicted diagnostic class with the highest probability, but also the second, third, and other possible diagnostic classes. RNA from various types of cancers was sequenced and quantified using next generation sequencing. We used these RNA expression profiles in several functional classification models trained on 3484 tumors across 25 cancer types and tested these models using 1716 tumors. The Extreme Gradient Boosting Classifier showed the greatest accuracy, followed by the Multi-Layer Perceptron and random forest models. We demonstrate that using RNA profiling and machine learning achieve high accuracy in both hard and functional classifications of tumors. However, significantly better accuracy in classification was achieved using functional classification reflecting biological overlapping between tumors. Functional classification shows proximity between tumors and enables exploring treatments used in other close diagnostic class.</p>