PD-CR

PD-CR implements a primal-dual optimization framework for feature selection and classification with rejection in high-dimensional metabolomics datasets, projecting data into a lower-dimensional space while simultaneously optimizing selected features and prediction accuracy.

Key Features:

• Primal-Dual Framework: Utilizes a primal-dual optimization technique that handles various loss functions and facilitates convergence analysis.
• Classification with Rejection: Performs classification with rejection by providing confidence scores and enabling rejection of uncertain predictions to reduce the false discovery rate (FDR).
• Feature Selection and Biological Relevance: Selects biologically relevant metabolites from complex datasets to identify key biomarkers.
• Confidence Scoring: Provides a confidence score for each prediction analogous to probability assessments used in clinical decision-making.
• Dimensionality Reduction via Projection: Projects high-dimensional data into a lower-dimensional space under a constrained optimization scheme while maintaining essential information.

Scientific Applications:

• Method Comparison: Demonstrated superior performance compared to Partial Least Squares Discriminant Analysis (PLS-DA) and Random Forests in metabolomics studies.
• Urinary Metabolomics (Lung Cancer): Applied to a urinary metabolomics dataset differentiating lung cancer patients from healthy controls.
• Glial Tumor IDH Status: Applied to glial tumor samples to distinguish isocitrate dehydrogenase (IDH)-mutated from IDH wild-type samples.
• Biomarker Identification: Identified biologically significant metabolites in both datasets, supporting clinical relevance.

Methodology:

PD-CR minimizes a quadratic cost function via a primal-dual optimization approach that simultaneously optimizes feature selection and prediction accuracy while projecting high-dimensional data into a lower-dimensional space.