The genome of cancer cells carries somatic mutations caused by DNA damage and repair process operations like ultraviolet and tobacco during the cellular lineage between the fertilized egg and the cancer cell. These mutational processes with specific mutation patterns did not characterize well. Sequencing technologies give us an opportunity to obtain a somatic mutation catalog from whole-genome and thus to discover the mutational process signature in human cancer. With these mutation signatures, we can understand the pathogenetic mechanism of all cancer and as a result, we can detect cancers in the primary stages. In this project, we intend to design a computational framework from the somatic mutation catalog caused by cancer so we can decipher mutation signatures. to do this, we use a probabilistic graphical model called compound latent Dirichlet allocation with variational Bayes inference which is able to consider all mutations of all cancers together and analyze them based on cancer type. This model can analysis all cancer genome together. In previous methods, the genomes of each type of cancer were examined separately and were not able to detect common mutation signatures between different cancer types.With utilizing a variational lower bound, we can find the most probable number of mutation signatures. Finally, to evaluate the model we use simulated data produced by the generative process of the LDA model. Also with applying our model to real mutations data, we find a new signature.