Chronological age estimation from human microbiomes with transformer-based Robust Principal Component Analysis.

Area of Science:

• Microbiome analysis
• Computational biology
• Machine learning

Background:

• Deep learning models show promise for analyzing microbiome data to understand human phenotypes.
• Existing methods may lack interpretability or optimal performance for complex microbiome datasets.

Purpose of the Study:

• To introduce Transformer-based Robust Principal Component Analysis (TRPCA) for microbiome analysis.
• To evaluate TRPCA's performance against conventional machine learning models for age prediction.
• To explore TRPCA's utility in multi-task learning (MTL) for combined prediction tasks.

Main Methods:

• TRPCA combines transformer architectures with Robust Principal Component Analysis.
• Benchmarking was performed on age prediction using 16S rRNA gene amplicon (16S) and whole-genome sequencing (WGS) data from skin, oral, and gut body sites.
• Multi-task learning was employed for simultaneous classification and regression tasks.

Main Results:

• TRPCA significantly improved age prediction accuracy, showing the largest Mean Absolute Error (MAE) reduction for WGS skin (28%) and 16S skin (14%) samples.
• The TRPCA MTL approach achieved 89% accuracy for birth country prediction and enhanced age prediction from WGS stool samples.
• Residual analysis revealed connections between subjects and prediction errors across sequencing methods and body sites.

Conclusions:

• TRPCA offers improved age prediction accuracy in human microbiome samples.
• The method maintains feature-level interpretability, aiding in understanding microbiome-phenotype relationships.
• TRPCA demonstrates potential for complex tasks like multi-site and multi-modal microbiome data analysis.