Decision Tree Ensembles Utilizing Multivariate Splits Are Effective at Investigating Beta Diversity in Medically...

Area of Science:

• Microbiome research
• Computational biology
• Human health and disease

Background:

• Understanding microbial community variation is crucial for analyzing health and disease states.
• 16S rRNA sequencing data provides insights into bacterial community composition.
• Previous methods may not fully capture complex high-throughput sequencing data structures.

Purpose of the Study:

• To investigate the use of learned dissimilarities for analyzing bacterial community composition in human stool samples.
• To develop and apply a workflow (TreeOrdination) for learning dissimilarities, dimensionality reduction, and feature identification.
• To improve the analysis of microbial communities in conditions like Crohn's disease and colorectal cancer.

Main Methods:

• Utilized 16S rRNA sequencing data from human stool samples.
• Developed an unsupervised decision tree ensemble approach to learn dissimilarities.
• Implemented a workflow for learning dissimilarities, projecting data into lower dimensions, and identifying impactful features.
• Applied centered log ratio transformation within the TreeOrdination workflow.

Main Results:

• TreeOrdination successfully identified differences in microbial communities between Crohn's disease patients and healthy controls.
• The workflow elucidated the impact of amplicon sequence variants (ASVs) on sample projections.
• Models demonstrated good generalization to unseen data and facilitated integration of patient data.
• Multivariate splits in models enhanced the analysis of complex sequencing data.

Conclusions:

• Learned representations can create informative ordinations for microbiome analysis.
• Model introspection algorithms can quantify the impact of specific taxa in ordinations.
• Identified taxa are associated with immune-mediated inflammatory diseases and colorectal cancer, highlighting their role in disease.