TCSM

TCSM models mutational processes in cancer by integrating clinical and molecular covariates directly into mutation signature analysis.

Key Features:

• Bayesian topic modeling: Uses Bayesian topic modeling to represent mutation signatures and exposures.
• Covariate-adjusted priors: Adjusts the prior distribution on signature exposure based on observed tumor-level covariates.
• Covariate imputation: Imputes covariates in held-out datasets to enable inference when covariate data are missing.
• Statistical association testing: Evaluates the statistical significance of associations between mutation signatures and covariates.
• Signature resolution: Improves distinction and recovery of true exposure levels for similar mutation signatures.
• Benchmarking: Demonstrates performance comparisons with non-negative matrix factorization and other topic modeling approaches.
• Implementation: Implemented in Python 3.

Scientific Applications:

• Breast cancer signature discovery: Identified five distinct mutation signatures in breast cancer datasets.
• Homologous recombination deficiency prediction: Predicted homologous recombination repair deficiency in additional tumor samples.
• Cross-cancer signature analysis: Uncovered four mutation signatures across a combined melanoma and lung cancer cohort using cancer type as a covariate.
• Tumor classification insight: Provided statistical support that some TCGA-classified lung cancers may represent misdiagnosed metastatic melanomas.
• Validation in simulations and real data: Recovered true exposure levels in simulated data and demonstrated improved recovery in real-world datasets.

Methodology:

Applies Bayesian topic modeling with covariate-adjusted priors on signature exposures, performs covariate imputation in held-out datasets, evaluates statistical significance of signature–covariate associations, and was implemented in Python 3 for benchmarking against non-negative matrix factorization and other topic modeling methods.