Bayesian multi-study non-negative matrix factorization for mutational signatures

Area of Science:

• Genomics
• Bioinformatics
• Cancer Research

Background:

• Mutational signatures are crucial for understanding cancer development and are typically identified using non-negative matrix factorization (NMF).
• Current NMF methods analyze single datasets, hindering cross-condition signature comparisons.
• This limitation impedes comprehensive analysis of cancer evolution and treatment response.

Purpose of the Study:

• To introduce a novel Bayesian NMF method for joint decomposition of multiple tumor genome datasets.
• To enable rigorous comparison of mutational signatures across different conditions.
• To identify shared mutational signatures and their patterns.

Main Methods:

• Developed a Bayesian NMF framework for simultaneous decomposition of multiple datasets.
• Introduced an unsupervised 'discovery-only' model for novel signature identification.
• Proposed a semi-supervised 'recovery-discovery' model to estimate known and novel signatures.
• Extended models to incorporate covariate effects for enhanced analysis.

Main Results:

• The Bayesian NMF method successfully identifies mutational signatures and their cross-dataset sharing patterns.
• Simulations demonstrate the method's robustness and accuracy.
• Application to colorectal and early-onset breast cancer data provides insights into disease-specific signatures.

Conclusions:

• The proposed Bayesian NMF approach overcomes limitations of single-dataset NMF for signature analysis.
• This method facilitates robust comparisons of mutational signatures across diverse cancer types and conditions.
• The framework offers a powerful tool for advancing cancer genomics research and personalized medicine.