Differential Compositional Variation FeatureSelection: A Machine Learning Framework with LogRatios for Compositional Metagenomic Data

Abstract

The demand for tight integration of compositional data analysis and machine learning methodologies for predictive modeling in high-dimensional settings has increased dramatically with the increasing availability of metagenomics data. We develop the differential compositional variation machine learning framework (DiCoVarML) with robust multi-level log ratio bio-marker discovery for metagenomic datasets. Our framework makes use of the full set of pairwise log ratios, scoring ratios according to their variation between classes and then selecting out a small subset of log ratios to accurately predict classes. Importantly, DiCoVarML supports a targeted feature selection mode enabling researchers to define the number of predictors used to develop models. We demonstrate the performance of our framework for binary classification tasks using both synthetic and real datasets. Selecting from all pairwise log ratios within the DiCoVarML framework provides greater flexibility that can in demonstrated cases lead to higher accuracy and enhanced biological insight.