Feature Selection Strategies for Deep Learning-Based Classification in Ultra-High-Dimensional Genomic Data.

Area of Science:

• Genomics
• Bioinformatics
• Machine Learning

Background:

• High-throughput sequencing generates vast genomic data, posing statistical challenges like the p >> n problem in Whole-Genome Sequencing.
• Efficient feature selection is crucial for analyzing ultra-high-dimensional genomic datasets.

Purpose of the Study:

• To address the need for efficient computational and statistical tools for feature selection in high-dimensional genomic data.
• To evaluate the performance of different feature selection algorithms for breed classification using genomic data.

Main Methods:

• Applied three feature selection algorithms: SNP-tagging, one-dimensional Supervised Rank Aggregation (1D-SRA), and multidimensional Supervised Rank Aggregation (MD-SRA).
• Classified 1825 individuals into five breeds using 11,915,233 Single Nucleotide Polymorphisms (SNPs).
• Utilized a deep learning classifier (Convolutional Neural Networks) for breed classification.

Main Results:

• SNP-tagging achieved an F1-score of 86.87% with rapid computation.
• 1D-SRA offered the best classification quality (96.81%) but faced computational, memory, and storage limitations.
• MD-SRA demonstrated a balance between classification quality (95.12%) and computational efficiency (17x faster, 14x less storage).

Conclusions:

• MD-SRA is a suitable and efficient approach for classifying high-dimensional data, offering a balance between accuracy and computational resources.
• SRA-based methods are versatile and applicable beyond genomic data analysis.