GRridge

GRridge applies adaptive group-regularized ridge regression to improve predictive modeling of high-dimensional biological data by integrating multiple sources of co-data (genomic annotations, external p-values) for logistic, linear, and Cox models.

Key Features:

• Adaptive group-regularization: Partitions variables into groups based on co-data and applies group-specific ridge penalties to incorporate structured co-data information.
• Empirical Bayes penalty estimation: Derives analytical empirical Bayes estimates for group-specific penalties that adapt to the informativeness of the co-data.
• Efficient penalty tuning: Tunes a single global penalty parameter via cross-validation to optimize regularization strength.
• Multi-type co-data integration: Integrates multiple types of co-data (e.g., genomic annotations, external p-values) with minimal additional computation.
• Variable selection and diagnostics: Facilitates post-hoc variable selection by distinguishing near-zero from larger regression parameters and supports prediction diagnostics via cross-validation using ROC curves and AUC metrics.

Scientific Applications:

• Cancer genomics prediction: Improves predictive performance over conventional logistic ridge regression and group lasso in cancer genomics studies.
• Methylation-based classification: Applied to discriminate precancerous cervical lesions from normal cervix tissue using methylation microarray data.
• Sparse model selection: Maintains high predictive performance when selecting a limited subset of variables (for example, 42 predictors) from high-dimensional data.

Methodology:

Uses co-data to define variable groups and inform analytical empirical Bayes group-specific penalties within a ridge regression framework for logistic, linear, and Cox models; tunes a single global penalty by cross-validation and enables post-hoc variable selection and prediction diagnostics via ROC/AUC.