Advances in uncertainty quantification in the analysis of mutational signatures in cancer genomics
In clinical oncology, the number and patterns of DNA mutations are increasingly used to investigate tumor etiology and identify biomarkers relevant for treatment selection. These patterns are extracted by dimensionality reduction using non-negative matrix factorization (NMF), with a linear decomposition of an m x n mutation-count matrix (m mutation types across n tumors) into signature and exposure matrices. Signatures represent recurrent mutational patterns linked to the latent generating processes, while exposures quantify their contribution to each tumor.
Uncertainty has been quantified using conditional probabilities of mutation types and Poisson-based sensitivity analyses. Here, we propose a new bootstrap-based alternative for sampling residuals from the fitted NMF decomposition. Bootstrap matrices are then refitted to obtain 95% confidence intervals. This framework provides a complementary assessment of the robustness of NMF-derived mutational signatures in cancer genomics studies.
Keywords: genomics patterns signatures