TaxaHFE: a machine learning approach to collapse microbiome datasets using taxonomic structure.

Area of Science:

• Bioinformatics
• Computational Biology
• Machine Learning in Biology

Background:

• Biologists utilize machine learning for prediction and explanation.
• Feature reduction enhances model performance and interpretability.
• Hierarchical structures in biological data (e.g., microbiome taxonomy) are often unexploited in feature reduction.

Purpose of the Study:

• To design a feature engineering algorithm that exploits hierarchical relationships in biological data.
• To improve the interpretability and performance of machine learning models using taxonomic information.

Main Methods:

• Developed TaxaHFE, an algorithm to collapse information-poor features into higher taxonomic levels.
• Applied TaxaHFE to six biological datasets.
• Compared machine learning models using species-level features versus TaxaHFE-preprocessed features.

Main Results:

• Achieved an average 90% reduction in features across datasets.
• Models using TaxaHFE features showed an average 3.47% increase in receiver operator curve area under the curve.
• TaxaHFE uniquely accommodates both categorical and continuous response variables for feature collapse.

Conclusions:

• TaxaHFE effectively reduces hierarchically organized features into a more informative subset.
• This reduction enhances the interpretability of machine learning models in biological studies.
• The algorithm offers a novel approach to feature engineering for taxonomic data.