PhyloMix: enhancing microbiome-trait association prediction through phylogeny-mixing augmentation

Area of Science:

• Microbiome research
• Computational biology
• Machine learning applications

Background:

• Understanding trait-microbe associations is key in microbiome research.
• Machine learning (ML) models show promise but struggle with high-dimensional, compositional, and imbalanced microbiome data.
• Data augmentation is crucial to improve ML model performance on such data.

Purpose of the Study:

• Introduce PhyloMix, a novel data augmentation method for microbiome data.
• Enhance predictive analyses by leveraging phylogenetic relationships among taxa.
• Address the compositional nature of microbiome data for improved ML.

Main Methods:

• PhyloMix generates synthetic microbial samples by combining phylogenetic subtrees from existing samples.
• The method is designed to handle both raw counts and relative abundances, addressing data compositionality.
• Phylogenetic relationships serve as an informative prior to guide synthetic sample generation.

Main Results:

• PhyloMix significantly improves predictive performance across six real microbiome datasets and five ML models.
• Outperforms baseline methods including vanilla mixup, compositional cutmix, and the phylogeny-based method TADA.
• Demonstrates wide applicability in both supervised learning and contrastive representation learning.

Conclusions:

• PhyloMix effectively enhances ML model performance for microbiome data analysis.
• Leveraging phylogenetic information is a powerful strategy for microbiome data augmentation.
• The method offers a robust solution for analyzing complex microbiome datasets.