Abstract
Microbiome composition profiles generated from 16S rRNA sequencing have been extensively studied for their usefulness in phenotype trait prediction, including for complex diseases such as diabetes and obesity. These microbiome compositions have typically been quantified in the form of Operational Taxonomic Unit (OTU) count matrices. However, alternate approaches such as Amplicon Sequence Variants (ASV) have been used, as well as the direct use of k-mer sequence counts. The overall effect of these different types of predictors when used in concert with various machine learning methods has been difficult to assess, due to varied combinations described in the literature. Here we provide an in-depth investigation of more than 1,000 combinations of these three clustering/counting methods, in combination with varied choices for normalization and filtering, grouping at various taxonomic levels, and the use of more than ten commonly used machine learning methods for phenotype prediction. The use of short k-mers, which have computational advantages and conceptual simplicity, is shown to be effective as a source for microbiome-based prediction. Among machine-learning approaches, tree-based methods show consistent, though modest, advantages in prediction accuracy. We describe the various advantages and disadvantages of combinations in analysis approaches, and provide general observations to serve as a useful guide for future trait-prediction explorations using microbiome data.