Abstract
Replication errors in the genome accumulate from a variety of mutational processes, which leave a history of mutations on the affected genome. The relative contribution of each mutational process has been characterized by non-negative matrix factorization and has lead to deeper insight into both mutational and repair processes contributing to cancer. However current implementations of NMF have left unresolved some specific patterns that should be present in the mutation data and have not generated signatures designed for classification. Here, we use a variant of NMF, termed non-smooth NMF, to generate sparse matrix factorizations of somatic mutation profiles present in 7129 tumors. nsNMF factorization revealed 21 mutational signatures. We found three APOBEC mutational processes clearly segregating with the published APOBEC enzymology and trans-lesion repair processes. We discovered several signatures differed between geographic locations even between closely related tissues.
