ReRa

ReRa performs relevance-redundancy feature selection to identify predictive, non-redundant features from high-dimensional genomic and transcriptomic data for unbalanced classification tasks in translational bioinformatics.

Key Features:

• Relevance-Based Filtering: ReRa applies a customized relevance-based filtering step to select predictive features from the initial high-dimensional dataset.
• Supervised Redundancy Minimization: ReRa uses a supervised similarity-based procedure combining global and class-specific similarity assessments to remove redundant features while preserving class-differentiated ones.
• Iterative Re-Evaluation: ReRa re-evaluates previously selected features at each iteration of the redundancy assessment to retain the most relevant and class-differentiated features.
• No Feature Number Tuning: ReRa does not require predefining the number of features to preserve.

Scientific Applications:

• Breast cancer patient subtyping: ReRa-selected feature spaces improved model performance for breast cancer patient subtyping using gene or transcript isoform expression data.
• Comparative benchmarking: ReRa demonstrated performance improvements over simple filtering, LASSO regularization, and the MRmr feature selection method.
• Precision medicine and patient stratification: ReRa supports development of predictive models for precision medicine and patient stratification in complex diseases such as cancer.
• Unbalanced classification scenarios: ReRa addresses high-dimensional datasets with unbalanced class distributions common in translational bioinformatics and genomics.

Methodology:

Two computational stages: relevance-based filtering to identify predictive features, followed by supervised redundancy minimization using global and class-specific similarity assessments with iterative re-evaluation of selected features.