DBFE: distribution-based feature extraction from structural variants in whole-genome data.

Area of Science:

• Genomics and Bioinformatics
• Computational Biology
• Machine Learning in Biosciences

Background:

• Whole-genome sequencing provides comprehensive DNA data for biosciences.
• Machine learning is crucial for analyzing genomic variants in cancer research.
• High-volume genomic data presents challenges in feature engineering for machine learning.

Purpose of the Study:

• To propose novel feature extraction methods for genomic variants.
• To facilitate classifier learning from large-scale whole-genome data.
• To automate feature extraction for machine learning pipelines in genomics.

Main Methods:

• Developed three feature extraction strategies: variant length binning, clustering, and density estimation.
• Created a programming library for automated distribution-based feature extraction.
• Validated methods on five real-world datasets using four classification and one clustering algorithm.

Main Results:

• Experiments on ovarian, lung, and breast cancer patient genomes demonstrated successful identification of genomic biomarkers.
• The proposed methods correlate biomarkers with cancer subtypes and clinical treatment response.
• Extracted features are applicable to unsupervised learning for genomic sample analysis.

Conclusions:

• The developed feature extraction methods effectively address challenges in analyzing whole-genome data.
• Automated feature engineering aids in discovering cancer-related genomic biomarkers.
• These methods enhance machine learning applications in precision oncology and genomic research.