Integrative Bayesian models using Post-selective inference: A case study in radiogenomics.

Area of Science:

• Genomics
• Biostatistics
• Medical Imaging

Background:

• Integrative analyses of genomics and intermediary phenotypes (e.g., imaging) offer insights into disease mechanisms.
• Accurate inference of uncertainty in these models is crucial but challenged by variable selection bias.

Purpose of the Study:

• To develop novel selection-aware Bayesian methods to address bias in integrative models.
• To improve the accuracy of uncertainty estimation in genomic and imaging association studies.

Main Methods:

• Developed selection-aware Bayesian methods using a "selection-aware posterior".
• Employed ℓ1-regularized algorithms for variable selection.
• Utilized a conditional likelihood function with reparameterization for tractable Markov chain Monte Carlo (MCMC) sampling.

Main Results:

• Successfully counteracted model selection bias in integrative Bayesian models.
• Achieved accurate uncertainty estimation by balancing selection quality and inferential power.
• Applied methods to radiogenomic data, identifying key gene pathways associated with patient survival.

Conclusions:

• The developed selection-aware Bayesian methods provide more reliable insights from integrative genomic and imaging analyses.
• These methods enhance the clinical relevance of disease mechanism studies by improving uncertainty estimation.
• The approach is effective for radiogenomic analyses, linking gene pathways to survival outcomes.