Bayesian Hierarchical Varying-sparsity Regression Models with Application to Cancer Proteogenomics

Area of Science:

• Biomedical data integration
• Cancer genomics and proteomics
• Personalized medicine

Background:

• Cancer is a heterogeneous disease requiring personalized treatment strategies.
• Identifying patient-specific prognostic biomarkers is crucial for effective cancer therapy.
• Integrating multi-omics data (proteomics and genomics) with clinical data can improve biomarker discovery.

Purpose of the Study:

• To propose a novel regression framework, Bayesian hierarchical varying-sparsity regression (BEHAVIOR), for selecting clinically relevant disease markers.
• To integrate proteogenomic and clinical data for robust biomarker identification.
• To identify patient-level, genomically driven prognostic protein markers.

Main Methods:

• Developed Bayesian hierarchical varying-sparsity regression (BEHAVIOR) models.
• Modeled flexible protein-gene relationships.
• Induced sparsity in protein-gene and protein-survival associations.
• Applied the framework to The Cancer Genome Atlas (TCGA) proteogenomic pan-cancer data.

Main Results:

• BEHAVIOR models demonstrated superior performance in protein marker selection and survival prediction compared to competing methods in simulations.
• Identified prognostic proteins and pathways shared across multiple cancers.
• Discovered cancer-specific prognostic proteins and pathways.

Conclusions:

• BEHAVIOR provides a powerful framework for identifying patient-specific prognostic biomarkers by integrating proteogenomic and clinical data.
• The identified biomarkers can inform personalized treatment strategies for various cancers.
• The study highlights the utility of multi-omics data integration in advancing precision oncology.