TaxaHFE

TaxaHFE enhances machine learning on microbiome datasets by leveraging taxonomic hierarchies to collapse information-poor features into higher taxonomic levels for feature reduction while preserving interpretability.

Key Features:

• Hierarchical Feature Engineering: Exploits taxonomic organization (e.g., species to genus to family) to consolidate lower-level features into higher-level features.
• Significant Feature Reduction: Produces an average feature count reduction of 90% (standard deviation 5.1%) across six published datasets.
• Improved Model Performance: Yields an average increase of 3.47% in AUC-ROC when compared to models using the most resolved taxonomic level (species).
• Versatility with Response Variables: Supports both categorical and continuous response variables to inform the feature-collapse decisions.
• Enhanced Interpretability: Reduces hierarchically organized features to a more information-rich subset, improving interpretability of downstream models.

Scientific Applications:

• Microbiome machine-learning preprocessing: Streamlines taxonomically structured microbial feature tables prior to model training.
• Association studies: Facilitates detection of relationships between microbial taxa and biological or health-related outcomes.
• Hypothesis generation and validation: Improves interpretability to support hypothesis development and testing in ecological and medical studies.

Methodology:

Algorithmically collapses lower-level taxonomic features into higher taxonomic levels based on information richness, with the collapse guided by categorical or continuous response variables.