(PTFS08-03-23) Not All Taxa Matter Equally: A Machine Learning Approach To Collapse Microbiome Datasets Using Taxonomic Structure

Objectives: Many research questions require an understanding of data types that have hierarchical relationships (e.g., microbial taxonomy or food trees). While machine learning identifies informative features, it still struggles with a problem present in many human -omic studies: too many features describing relatively few samples. Our objective was to design a tool that enables users to leverage known hierarchical relationships and choose the maximally informative subset of features for input to machine learning models.

Methods: We developed software, called taxaHFE (Hierarchical Feature Engineering), which works by first considering the pairwise correlation structure between a taxon and its descendants to prune descendants above a correlation threshold. Next it permutes a random forest on the taxon and remaining descendants to determine how important each is at explaining an intervention or clinical covariate. If, on average, the taxon is the most important feature in the model, the descendants are dropped, otherwise only the descendants more important than the taxon are kept. Last, an optional final filter step considers all features remaining, and again permutes a random forest. Any features which are either below the average importance of all remaining features or have a negative or zero average importance are dropped.  

Results: Using three previously published datasets, we show that taxaHFE results in a 90% reduction in number of features (s.d = 4.4%) compared to using the most complete taxonomy. Moreover, machine learning models comparing the taxaHFE output against the complete taxonomy achieved an increase in receiver operator curve area under the curve (ROC-AUC) of 4.9%.

Conclusions: Here we present a tool for dynamically collapsing hierarchical data (such as taxonomies in microbiome data) to reduce feature rich datasets. The primary strengths of this method are threefold: 1) a dramatic decrease in the number of features, 2) an increase in the performance of machine learning models after preprocessing with taxaHFE, and 3) the ability to use both discrete and continuous covariates. Future work should examine the utility of taxaHFE on other hierarchically represented data, such as dietary data represented by food trees.

Funding Sources: United States Department of Agriculture, Agricultural Research Service
Oak Ridge Institute for Science and Education