MetAML

MetAML performs machine-learning-based prediction and biomarker discovery from shotgun metagenomic quantitative microbiome profiles to assess microbiome–phenotype associations.

Key Features:

• Machine Learning Classifiers: Employs machine learning classifiers to build predictive models from complex microbial data.
• Automatic Model and Feature Selection: Implements automatic model selection and feature optimization to improve predictive accuracy.
• Cross-Validation and Cross-Study Analysis: Supports rigorous cross-validation within studies and cross-study analysis to evaluate model generalizability across cohorts.
• Quantitative Microbiome Profiles: Uses species-level relative abundances and strain-specific markers derived from shotgun metagenomic analysis as input features.
• Meta-analysis of Public Samples: Enables meta-analysis of publicly available metagenomic samples for aggregated evaluation.

Scientific Applications:

• Biomarker Discovery: Facilitates identification of microbial indicators linked to specific health conditions in human-associated microbiomes.
• Disease Prediction: Enables disease prediction across cohorts by leveraging high-resolution microbial features.
• Dysbiosis Assessment: Assists in distinguishing general dysbiotic states from disease-specific signals, with examples such as Streptococcus anginosus indicating general dysbiosis.
• Control Integration for Improved Prediction: Incorporates healthy control samples from multiple studies to enhance disease prediction and help define microbial dysbiosis more generally.

Methodology:

Performs a comprehensive meta-analysis of publicly available shotgun metagenomic samples; evaluated on 2,424 samples from eight large-scale studies using cross-validation and feature selection, and compares models using strain-specific markers versus species-level taxonomic abundance, with strain-specific marker models often outperforming species-level models.