The Unsupervised Feature Selection Algorithms Based on Standard Deviation and Cosine Similarity for Genomic Data...

Area of Science:

• Genomics
• Bioinformatics
• Machine Learning

Background:

• Genomic data analysis faces challenges due to high dimensionality, limited samples, and class imbalance.
• Unsupervised feature selection is crucial for identifying relevant biological markers without prior labels.

Purpose of the Study:

• To propose novel unsupervised feature selection techniques for genomic data.
• To address the challenges of high dimensionality and class imbalance in cancer genomics.
• To identify robust biomarkers for cancer classification and research.

Main Methods:

• Developed Standard deviation and Cosine similarity based Feature Selection (SCFS) using discernibility and independence.
• Derived three algorithms: SCEFS, SCRFS, and SCAFS, based on different cosine similarity measures.
• Utilized KNN and SVM classifiers with selected features on 18 cancer genomic datasets.

Main Results:

• The proposed algorithms (SCEFS, SCRFS, SCAFS) successfully identified stable biomarkers with strong classification capabilities across 18 cancer datasets.
• Feature importance was determined by a 2D space plotting discernibility against independence.
• Functional analysis linked identified biomarkers to gene regulation levels relevant to cancer occurrence.

Conclusions:

• The SCFS-derived algorithms offer a powerful approach for unsupervised feature selection in high-dimensional genomic data.
• Identified biomarkers provide insights into cancer pathology, aiding drug development and early diagnosis.
• This method has significant implications for advancing cancer research and clinical applications.