SigMoS

SigMoS performs discovery and model selection of mutational signatures in cancer genomics by using a Negative Binomial non‑negative matrix factorization framework to account for overdispersed mutational counts and patient-specific variability.

Key Features:

• Negative Binomial NMF: Models observed mutational counts with a Negative Binomial distribution and incorporates a patient-specific dispersion parameter within the NMF framework.
• Model selection procedure: Implements a cross-validation‑inspired procedure to determine the optimal number of mutational signatures, contrasted with Poisson-based rank comparison methods.
• Parameter estimation: Uses derived update rules tailored for the Negative Binomial NMF to estimate model parameters.
• Robustness to misspecification: Demonstrates improved robustness in selecting the correct number of signatures under overdispersion relative to existing methods, as shown in simulation studies.
• Residual analysis: Provides residual analysis to investigate model fit and to detect and validate overdispersion in mutational count data.

Scientific Applications:

• Mutational signature discovery: Identification and interpretation of mutational signatures to study mutational processes in cancer genomics.
• Breast and prostate cancer studies: Applied to real datasets from breast and prostate cancer patients for empirical signature inference.
• Simulation benchmarking: Used in simulation studies to compare performance and robustness against alternative signature extraction methods.

Methodology:

Integrates a Negative Binomial distribution into NMF with a patient-specific dispersion parameter; estimates parameters using derived update rules; applies a cross-validation‑inspired model selection procedure to choose the number of signatures; and conducts residual analysis to assess model fit and overdispersion.