Microbiome meta-analysis and cross-disease comparison enabled by the SIAMCAT machine learning toolbox

Area of Science:

• Microbiome Research
• Bioinformatics
• Machine Learning Applications

Background:

• The human microbiome is a valuable source for diagnostic and therapeutic biomarkers.
• Existing machine learning (ML) software for metagenomics is limited, leading to evaluation issues and poor cross-study generalization.

Purpose of the Study:

• To develop a versatile R toolbox, SIAMCAT, for ML-based comparative metagenomics.
• To address the scarcity of metagenomics-specific software and improve the reliability of ML models in microbiome research.

Main Methods:

• Developed SIAMCAT, an R toolbox for machine learning-based comparative metagenomics.
• Conducted a meta-analysis of 10,803 fecal metagenomic samples.
• Implemented a novel training set augmentation strategy to improve model generalization.

Main Results:

• Machine learning models showed reduced accuracy and disease specificity when transferred naively across studies.
• The novel training set augmentation strategy in SIAMCAT successfully resolved these issues.
• Identified both disease-specific and condition-shared microbial biomarkers.

Conclusions:

• SIAMCAT provides a robust solution for ML-based comparative metagenomics, enhancing biomarker discovery.
• The developed strategy improves the cross-study generalization and reliability of microbiome ML models.
• SIAMCAT is freely available, promoting wider adoption in microbiome research.