Combining heterogeneous subgroups with graph-structured variable selection priors for Cox regression.

Area of Science:

• Bioinformatics
• Computational Biology
• Cancer Research

Background:

• Predicting cancer patient risk using molecular data like gene expression is crucial.
• Small and heterogeneous patient cohorts pose challenges for traditional prediction models.
• Subgroup analysis can lack statistical power, while pooling data may introduce bias.

Purpose of the Study:

• To develop a new Bayesian approach for cancer risk prediction and biomarker identification.
• To improve statistical power and prediction performance in small, heterogeneous cohorts.
• To enable information sharing across cohorts while maintaining individual cohort specificity.

Main Methods:

• A novel Bayesian statistical framework utilizing continuous molecular measurements and survival outcomes.
• Incorporation of a graph structure to link predictors within and across different cohorts.
• Information sharing mechanism to enhance statistical power in variable selection.

Main Results:

• The proposed Bayesian method effectively identifies important predictors (genes) for cancer risk.
• It provides distinct risk prediction models tailored for each individual cohort.
• The graph-based approach facilitates the discovery of functionally related gene pathways and cross-cohort prognostic markers.

Conclusions:

• The developed Bayesian approach outperforms standard methods in prediction accuracy.
• It significantly increases power for variable selection, particularly in small sample size scenarios.
• This method offers a robust solution for cancer research with limited or heterogeneous patient data.