An unsupervised cross-validation method to select the optimal number of mutational signatures. A data set of mutational counts is split into training and validation data.Signatures are estimated in the training data and then used to predict the mutations in the validation data.