Bayesian data integration and variable selection for pan-cancer survival prediction using protein expression data.

Area of Science:

• Biostatistics
• Translational Oncology
• Bioinformatics

Background:

• Precision medicine requires accurate prognostic prediction using molecular data.
• Identifying actionable proteins linked to clinical outcomes like patient survival is a significant challenge.
• Integrating multi-tumor data and handling censored survival outcomes adds complexity.

Purpose of the Study:

• To develop translational models for identifying major actionable proteins associated with patient survival.
• To address statistical and computational challenges in high-dimensional proteomic data analysis.
• To integrate data across different tumor types and accommodate censored survival data.

Main Methods:

• Developed Bayesian hierarchical survival models, specifically the accelerated failure time model.
• Employed sparse horseshoe prior distributions for regression coefficients to identify key proteomic drivers.
• Incorporated a correlation structure among prior distributions to borrow strength across tumor groups.

Main Results:

• The proposed integrative model effectively links different tumors using correlated prior structures.
• Analysis of The Cancer Proteome Atlas (TCPA) data demonstrated the model's efficacy.
• Simulations confirmed the utility of the Bayesian hierarchical approach for prognostic prediction.

Conclusions:

• The developed Bayesian hierarchical survival models successfully address challenges in multi-tumor proteomic data analysis.
• The integrative model facilitates the identification of major proteomic drivers for improved prognostic prediction.
• This approach advances precision medicine by linking molecular information to clinical outcomes.