Relevant and Non-Redundant Feature Selection for Cancer Classification and Subtype Detection.

Area of Science:

• Genomics and Bioinformatics
• Computational Biology
• Translational Medicine

Background:

• Identifying significant, non-redundant features from omics data is crucial for biological research.
• Existing statistical and machine learning methods for gene selection often yield highly co-regulated features, leading to inconsistent performance.

Purpose of the Study:

• To develop a novel feature selection algorithm for identifying disease-related and non-redundant features from diverse omics datasets.
• To evaluate the algorithm's performance in disease classification and subtyping across multiple cancer types.

Main Methods:

• A new feature selection algorithm was developed to extract salient, non-redundant features from omics data.
• The algorithm was applied to three biological problems: disease-normal classification, multiclass disease classification, and disease subtype detection.
• Performance was evaluated using metrics such as ROC-AUC, false-positive, and false-negative rates on TCGA cancer datasets.

Main Results:

• The proposed algorithm outperformed existing gene selection and differential expression methods in binary and multiclass cancer classification.
• Selected genes enhanced disease subtyping accuracy for four cancer types compared to state-of-the-art approaches.
• The algorithm demonstrated robust performance across six different cancer datasets from TCGA.

Conclusions:

• The novel feature reduction method effectively identifies disease-specific biomarkers and relevant features from omics data.
• This approach supports advancements in precision medicine design and disease subtyping.
• The algorithm offers a valuable tool for biologists and researchers working with complex omics datasets.