Model selection and robust inference of mutational signatures using Negative Binomial non-negative matrix...

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Cancer genomes exhibit mutations that can be classified into distinct mutational signatures.
• Non-negative matrix factorization (NMF) is commonly used to identify these signatures.
• Traditional NMF assumes Poisson distribution for mutation counts, which may not be suitable for overdispersed data.

Purpose of the Study:

• To develop a Negative Binomial NMF model that accounts for overdispersed mutation count data.
• To introduce a robust model selection procedure for determining the correct number of mutational signatures.
• To compare the performance of the proposed method against existing approaches.

Main Methods:

• Implemented Negative Binomial NMF with patient-specific dispersion parameters.
• Developed a novel cross-validation-inspired model selection procedure.
• Conducted simulations to evaluate distributional assumptions and compare methods.
• Applied the model to simulated and real cancer genomic data (breast and prostate).

Main Results:

• The proposed Negative Binomial NMF method accurately estimates mutational signatures, even with overdispersed data.
• The novel model selection procedure is more robust and accurate in identifying the true number of signatures compared to classical methods.
• State-of-the-art methods were found to overestimate the number of signatures when overdispersion is present.
• Residual analysis on real data confirmed the presence of overdispersion.

Conclusions:

• The developed Negative Binomial NMF and model selection procedure offer a more reliable approach for mutational signature analysis.
• The findings highlight the importance of accounting for overdispersion in cancer genomic data.
• The R package SigMoS provides accessible tools for implementing these methods.