EPheClass: ensemble-based phenotype classifier from 16S rRNA gene sequences.

Area of Science:

• Bioinformatics and computational biology
• Microbiome research
• Machine learning applications in health

Background:

• Classifying polymicrobial diseases from high-dimensional 16S rRNA gene sequencing data is challenging due to data complexity and microbial community imbalance.
• Existing methods struggle with the heterogeneity and high dimensionality inherent in microbiome datasets.

Purpose of the Study:

• To develop and evaluate a curated bioinformatics pipeline for binary phenotype classification of microbiomes using 16S rRNA gene amplicon data.
• To assess the performance of machine learning models, specifically Dynamic Ensemble Selection (DES) techniques, for disease classification.

Main Methods:

• A pipeline was developed using feature selection and machine learning models on 16S rRNA gene amplicon count tables.
• Dynamic Ensemble Selection (DES) models were tuned and evaluated, with DES-P identified as the top-performing technique.
• The pipeline was validated on datasets for periodontal disease, inflammatory bowel disease (IBD), and antibiotic exposure.

Main Results:

• DES models showed similar and more robust performance compared to individual models.
• DES-P achieved high diagnostic metrics for periodontal disease, including an F1 score of 0.913 and AUC of 0.973 with only 13 features.
• The EPheClass tool, based on this pipeline, outperformed existing methods for IBD classification and showed competitive results for antibiotic exposure detection.

Conclusions:

• The proposed classification pipeline demonstrates strong generalization capabilities across different phenotypes, sample types, and study niches.
• Dynamic Ensemble Selection (DES) techniques offer a robust approach for microbiome-based disease classification.
• This approach holds significant potential for advancing the diagnosis and understanding of polymicrobial diseases.