Mutational signature learning with supervised negative binomial non-negative matrix factorization.

Area of Science:

• Genomics
• Computational Biology
• Cancer Research

Background:

• Understanding cancer mutational processes is crucial for improving diagnosis and treatment.
• Mutational signatures summarize these processes, but current methods have limitations.
• Existing methods rely on mutation counts and are sensitive to cohort size, often requiring extensive post-processing.

Purpose of the Study:

• To develop a supervised method for extracting more distinctive and robust mutational signatures.
• To leverage cancer type metadata to enhance signature derivation.
• To explore the use of molecular features for mutational signature analysis.

Main Methods:

• A supervised method combining negative binomial non-negative matrix factorization with a support vector machine loss.
• Utilizing cancer type as metadata for signature extraction.
• Adapting the model to incorporate molecular features such as gene expression and mutation status.

Main Results:

• The supervised method yields mutational signatures with lower reconstruction error.
• Signatures derived are more predictive of cancer type compared to unsupervised methods.
• The approach enhances signature robustness, particularly with small or cancer-type limited patient cohorts.
• Demonstrated utility by deriving signatures from APOBEC expression and MUTYH mutation status.

Conclusions:

• Exploiting metadata significantly improves the quality and interpretability of mutational signatures.
• The supervised method offers a more robust and efficient approach to mutational signature analysis.
• This advancement holds promise for novel discoveries in cancer genomics and personalized medicine.