Flexible model-based non-negative matrix factorization with application to mutational signatures.

Area of Science:

• Genomics
• Cancer Research
• Bioinformatics

Background:

• Somatic mutations in cancer are often modeled as a mixture of mutational signatures.
• Current signature models include mono-nucleotide and tri-nucleotide interactions.
• Non-negative matrix factorization (NMF) is a common method for signature inference.

Purpose of the Study:

• To develop a novel framework for identifying biologically plausible parametrizations of mutational signatures.
• To specifically estimate di-nucleotide interaction models for cancer mutations.
• To assess the stability and accuracy of di-nucleotide interaction signatures.

Main Methods:

• Developed a novel framework for mutational signature parametrization.
• Employed the expectation-maximization (EM) algorithm for signature estimation.
• Utilized regression in the log-linear quasi-Poisson model.
• Applied the framework to simulated data and real cancer mutation datasets (breast, liver, urinary tract).

Main Results:

• Di-nucleotide interaction signatures are statistically stable and sufficiently complex for fitting mutational patterns.
• These signatures balance data fitting and avoid over-fitting.
• Di-nucleotide signatures offer improved biological plausibility and data fit over mono-nucleotide signatures.
• Parametrization is more stable than with tri-nucleotide signatures.

Conclusions:

• The proposed framework provides a robust method for analyzing cancer mutational signatures.
• Di-nucleotide interaction models represent a significant advancement in understanding cancer mutation processes.
• This approach enhances biological interpretability and model stability in cancer genomics.