endoR

endoR interprets tree ensemble machine learning models to extract and visualize feature importance and interactions in compositional, high-dimensional, and sparse microbiome and metagenomic datasets.

Key Features:

• Model simplification: Simplifies fitted tree ensemble models into decision ensembles to improve interpretability.
• Feature importance and interactions: Extracts importance of individual features and their pairwise interactions from the model.
• Visual network representation: Represents features and their interactions as an interpretable network visualization.
• Adjustable regularization and bootstrapping: Applies adjustable regularization and bootstrapping to reduce model complexity and retain essential components.
• Support for compositional, high-dimensional, sparse data: Operates on compositional, high-dimensional, and sparse microbial and genomic feature datasets.

Scientific Applications:

• Method evaluation: Validated on simulated and real metagenomic datasets to infer associations with accuracy comparable or superior to other commonly used methods.
• Cirrhosis microbiome analysis: Used to confirm known microbiome differences between cirrhotic and healthy individuals.
• Global metagenome analysis: Applied to a global metagenome dataset of 2,203 individuals to investigate components predicting the presence of human gut methanogens.
• Methanogen-associated networks: Confirmed associations between Methanobacteriaceae and Christensenellales and identified Methanobacteriaceae association with a network of hydrogen-producing bacteria.

Methodology:

endoR simplifies fitted tree ensemble models into decision ensembles, extracts feature importance and pairwise interactions, applies adjustable regularization and bootstrapping to retain essential components, and visualizes relationships as an interpretable network.