Gene-based microbiome representation enhances host phenotype classification.

Area of Science:

• Microbiome research
• Machine learning
• Bioinformatics

Background:

• The gut microbiome's role in host health is increasingly recognized, with potential for biomarker discovery.
• Shotgun metagenomics generates high-dimensional data, posing challenges for modeling host-microbiome interactions.
• Developing effective data representations is crucial for machine learning in microbiome studies.

Purpose of the Study:

• To compare the predictive performance of different data representations from shotgun metagenomics for host health classification.
• To evaluate gene-based approaches against traditional taxonomic and functional profiles.
• To explore the utility of gene families and functional categories for improved classification.

Main Methods:

• Comparison of machine learning model performance using various data representations: taxonomic profiles, functional profiles, and gene clusters.
• Utilized five case-control datasets: Type 2 diabetes, obesity, liver cirrhosis, colorectal cancer, and inflammatory bowel disease.
• Investigated the impact of using subsets of gene families from specific functional categories.

Main Results:

• Gene-based data representations, alone or combined with reference-based data, achieved improved or similar classification performance compared to taxonomic and functional profiles.
• Subsets of gene families from specific functional categories highlighted their importance in host phenotype.
• Both reference-free microbiome representations and curated annotations proved valuable for machine learning.

Conclusions:

• Data representation significantly impacts machine learning performance in metagenomic studies.
• Untargeted microbiome gene content can be as effective or superior to taxonomic profiling for classification.
• Function-based feature selection can enhance classification for specific pathologies, potentially generating new mechanistic hypotheses.