Supervised Relevance-Redundancy assessments for feature selection in omics-based classification scenarios.

Area of Science:

• Translational bioinformatics
• Genomics
• Machine learning

Background:

• High-dimensional data and imbalanced classes challenge classification in bioinformatics.
• This impacts classifier robustness, leading to overfitting and hindering precision medicine for diseases like cancer.
• Effective feature selection is crucial for removing irrelevant, redundant, and noisy features.

Purpose of the Study:

• To introduce ReRa, a novel supervised Relevance-Redundancy feature selection approach.
• To improve patient stratification for complex diseases by enhancing feature selection.
• To address limitations of existing methods in handling high dimensionality and class imbalance.

Main Methods:

• ReRa employs a two-step process: relevance-based filtering followed by similarity-based redundancy reduction.
• It uses a combination of global and class-specific similarity assessments to preserve class-differentiating features.
• The method is designed to be efficient and scalable for high-dimensional datasets.

Main Results:

• ReRa significantly improved classification performance for breast cancer patient subtyping compared to other methods like LASSO and MRmr.
• The method demonstrated effectiveness in two use cases: gene expression and transcript isoform expression.
• ReRa-selected feature spaces enhanced classifier performance, particularly in imbalanced scenarios.

Conclusions:

• The ReRa approach offers a robust solution for feature selection in imbalanced classification tasks.
• It outperforms existing Relevance-Redundancy methods like MRmr by not requiring feature number tuning and allowing feature re-evaluation.
• ReRa's scalability and ability to preserve class-differentiated features make it valuable for translational applications in precision medicine.