This study presents a computational framework for deciphering mutational signatures in human cancer genomes. The framework models mutational processes as a blind source separation problem, enabling the identification of mutational signatures from somatic mutation catalogs. The approach is based on nonnegative matrix factorization (NMF), which allows the extraction of mutational signatures from cancer genomes by decomposing the mutational catalog into a combination of mutational signatures and their respective exposures. The framework is robust to a wide range of parameters and can be applied to both genome and exome sequences. The study demonstrates that the framework can accurately identify mutational signatures from simulated and real data, including a wide variety of mutation types such as substitutions, indels, strand bias, and kataegis. The framework was tested on 100 simulated cancer genomes and successfully identified reproducible mutational signatures, with high similarity to the original signatures used to generate the data. The results show that the framework can distinguish between mutational processes even when the number of mutations is low, and that the accuracy of signature extraction improves with the number of cancer genomes analyzed. The study also shows that the framework can be applied to exome sequencing data, revealing mutational signatures from breast cancer exomes. The results indicate that exome data can be used to extract mutational signatures, although with less precision and comprehensiveness compared to whole genome data. The framework was further applied to analyze mutational signatures in breast cancer genomes, revealing five distinct mutational signatures, with the fifth signature being characterized by kataegis. The study highlights the importance of considering sequence context in the analysis of mutational signatures, as incorporating additional sequence context can reveal new dependencies and provide insights into the underlying biological processes. The framework was also applied to analyze strand bias in mutational signatures, revealing that certain mutational signatures exhibit a context-specific strand bias, which may be restricted to exons. Overall, the study demonstrates the effectiveness of the computational framework in deciphering mutational signatures from cancer genomes, providing insights into the mutational processes underlying human cancer. The framework has the potential to be applied to a wide range of cancer types and can be used to identify mutational signatures with high accuracy and reproducibility.This study presents a computational framework for deciphering mutational signatures in human cancer genomes. The framework models mutational processes as a blind source separation problem, enabling the identification of mutational signatures from somatic mutation catalogs. The approach is based on nonnegative matrix factorization (NMF), which allows the extraction of mutational signatures from cancer genomes by decomposing the mutational catalog into a combination of mutational signatures and their respective exposures. The framework is robust to a wide range of parameters and can be applied to both genome and exome sequences. The study demonstrates that the framework can accurately identify mutational signatures from simulated and real data, including a wide variety of mutation types such as substitutions, indels, strand bias, and kataegis. The framework was tested on 100 simulated cancer genomes and successfully identified reproducible mutational signatures, with high similarity to the original signatures used to generate the data. The results show that the framework can distinguish between mutational processes even when the number of mutations is low, and that the accuracy of signature extraction improves with the number of cancer genomes analyzed. The study also shows that the framework can be applied to exome sequencing data, revealing mutational signatures from breast cancer exomes. The results indicate that exome data can be used to extract mutational signatures, although with less precision and comprehensiveness compared to whole genome data. The framework was further applied to analyze mutational signatures in breast cancer genomes, revealing five distinct mutational signatures, with the fifth signature being characterized by kataegis. The study highlights the importance of considering sequence context in the analysis of mutational signatures, as incorporating additional sequence context can reveal new dependencies and provide insights into the underlying biological processes. The framework was also applied to analyze strand bias in mutational signatures, revealing that certain mutational signatures exhibit a context-specific strand bias, which may be restricted to exons. Overall, the study demonstrates the effectiveness of the computational framework in deciphering mutational signatures from cancer genomes, providing insights into the mutational processes underlying human cancer. The framework has the potential to be applied to a wide range of cancer types and can be used to identify mutational signatures with high accuracy and reproducibility.