Abstarct: Prompt and accurate diagnosis of cancer or its subtypes in patients has a significant impact on the correct treatment process and reduction of treatment costs. There are a variety of methods for diagnosing cancer and its subtypes that have evolved with recent advances in machine learning and deep learning. Also, new methods have been introduced using machine learning and deep learning that have good results in predicting different types of cancer or its subgroups and diagnosing benign and malignant cancerous tumors in patients. One of the methods that is considered today for the diagnosis of cancer and its subtypes is the classification of cancer baxsed on somatic mutations. Exposure to UV rays or certain chemicals can cause mutations in the body's cells. These abnormal mutations caused by environmental factors are called somatic mutations. However, the classification of somatic mutation-baxsed cancers is challenging. Challenges such as low sample data volume, high data scatter, overfitting, and the use of simple linear classifiers are factors that prevent increased classification performance. This paper presents ways to solve these challenges. These methods include clustering gene filter preprocessing, indexed scatter reduction, regulatory methods, the Global-Max-Pooling laxyer, and the use of the embedding laxyer. Also in this dissertation, three deep learning models CNN, LSTM and a combination of these two models are tested on the TCGA-DeepGene data set. Our proposed model is a single-laxyer CNN model with an embedding laxyer. This model achieved 66.45% accuracy. Compared to the reference cited in this dissertation, the accuracy has increased by 1.45%.