Statistical learning of mutational signatures in cancer genomics.
Somatic mutations in cancer can be described as a mixture of different mutational signatures. The signatures can often be attributed to exposures such as UV-light or tobacco smoking, and they can be decoded using a mixed membership model called non-negative matrix factorization. I will describe how non-negative matrix factorization works and why it is useful for cancer treatment and precision medicine. I will then proceed by explaining some fundamental methodological problems with applications of non-negative matrix factorization, and discuss novel solutions developed in my group.
The talk is based on very recent joint work with Ragnhild Laursen (Department of Mathematics, Aarhus University), Marta Pelizzola (University of Veterinary Medicine Vienna) and Lasse Maretty (Department of Molecular Medicine, Aarhus University).