Microbiome meta-analysis and cross-disease comparison enabled by the SIAMCAT machine-learning toolbox
Problem
Machine-learning models built on human microbiome data suffered from scarce metagenomics-specific software, overoptimistic evaluation, and poor generalization across studies.
Why it matters
Reliable cross-study microbiome models are needed to mine diagnostic and therapeutic biomarkers for human diseases.
What the AI did
SIAMCAT, an R toolbox for ML-based comparative metagenomics, enabled a meta-analysis of 10,803 fecal samples; a novel training set augmentation strategy restored accuracy lost when models were transferred across studies.
Results
Meta-analysis of 10,803 fecal metagenomic samples; naive cross-study transfer lost accuracy and disease specificity, resolved by training set augmentation; some biomarkers disease-specific, others shared.
Impact
Provides a freely available toolbox enabling cross-study microbiome ML analysis and biomarker comparison.
Record
| Venue | bioRxiv (Cold Spring Harbor Laboratory) |
|---|---|
| Published | 2020 |
| Work type | preprint |
| Authors | Jakob Wirbel, Konrad Zych, Morgan Essex, Nicolai Karcher, Ece Kartal, Guillem Salazar, Peer Bork B, Shinichi Sunagawa et al. |
| Citations | 32 (OpenAlex, at import time) |
| References | 112 |
| Model or method | SIAMCAT (ML toolbox; specific learner not stated) |
| Organization | EMBL (implied by siamcat.embl.de) |
| Primary source | https://doi.org/10.1101/2020.02.06.931808 |
| Open access | green |
| Keywords | microbiome, machine learning, meta-analysis, biomarkers, cross-study generalization |
Importance rating
real but narrower advance
Meta-analysis of 10,803 fecal samples addresses overoptimistic evaluation and poor cross-study generalization in microbiome ML; real but methodological/tooling advance for biomarker research.