Mutational signatures are characteristic patterns of mutations caused by endogenous mutational processes or by exogenous mutational exposures. Much research has focused on the problem of inferring mutational signatures as latent variables in somatic mutation data from multiple tumors. However, the problem of determining which signatures are present in a given sample and how many mutations each signature is responsible for has received negligible attention. In particular, there has been little systematic benchmarking of various approaches to this problem. This problem is referred to as "signature attribution" in a single sample. We show that this is a challenging problem, because there are often many combinations of signatures that can reconstruct the mutational spectrum of a given sample reasonably well. We benchmarked the accuracy of five approaches to signature attribution, including a new approach we call Presence Attribute Signature Activity (PASA), on large synthetic data sets. These data sets recapitulated the single-base, insertion-deletion, and doublet-base mutational signature repertoires of 9 cancer types. For single-base substitution mutations, PASA outperformed other approaches on all the cancer types combined. Interestingly, however, the ranking of approaches varied by cancer type. For doublet-base substitutions and small insertions and deletions, the ranking of approaches was more stable, with PASA outperforming other approaches in most, but not all of the nine cancer types. For all mutation types, the ranking of approaches varied by cancer type, and no approach achieved both high precision and recall. We believe these observations reflect the inherent challenges in signature attribution

A new approach to the challenging problem of mutational signature attribution