Information-theoretic approaches to SVM feature selection for metagenome read classification

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Metagenomics involves analyzing environmental DNA, yielding numerous genome fragments requiring taxonomic classification.
• Current methods often utilize all features, potentially missing opportunities to maximize classifier accuracy.

Purpose of the Study:

• To investigate the impact of feature selection on the performance of taxonomic classifiers for metagenomic data.
• To propose and evaluate novel filter-based feature selection methods derived from information theory.

Main Methods:

• Three filter-based feature selection techniques were proposed: a combination of Kullback-Leibler, Mutual Information, and distance information; TF-IDF (Term Frequency-Inverse Document Frequency); and mRMR (minimum redundancy-maximum-relevance).
• Support vector machine (SVM) classification of genomic reads was used to compare the effectiveness of these feature selection methods.

Main Results:

• Feature selection methods demonstrably boost the performance of taxonomic classifiers.
• The 6mer mRMR method exhibited strong performance, particularly for phyla-level classification.
• A trade-off exists between feature set size and the optimal feature selection method for classification performance.

Conclusions:

• Feature selection is crucial for optimizing taxonomic classification in metagenomics.
• mRMR and TF-IDF offer effective strategies for improving classification accuracy, with mRMR excelling at all taxonomic levels for N=6 and TF-IDF performing better for larger feature sets at finer resolutions.