Deep Learning model accurately classifies metastatic tumors from primary tumors based on mutational signatures

Abstract

Metastatic propagation is the leading cause of death for most cancers. Prediction and elucidation of metastatic process is crucial for the therapeutic treatment of cancers. Even though somatic mutations have been directly linked to tumorigenesis and metastasis, it is less explored whether the metastatic events can be identified through genomic mutation signatures, a concise representation of the mutational processes. Here, applying mutation signatures as input features calculated from Whole-Exome Sequencing (WES) data of TCGA and other metastatic cohorts, we developed MetaWise, a Deep Neural Network (DNN) model. This model accurately classified metastatic tumors from primary tumors. Signatures of non-coding mutations also have a major impact on the model performance. SHapley Additive exPlanations (SHAP) and Local Surrogate (LIME) analysis into the MetaWise model identified several mutational signatures directly correlated to metastatic spread in cancers, including APOBEC-mutagenesis, UV-induced signatures and DNA damage response deficiency signatures.