A benchmark analysis of feature selection and machine learning methods for environmental metabarcoding datasets

Next-Generation Sequencing methods like DNA metabarcoding enable the generation of large community composition datasets and have grown instrumental in many branches of ecology in recent years. However, the sparsity, compositionality, and high dimensionality of metabarcoding datasets pose challenges in data analysis. In theory, feature selection methods improve the analyzability of eDNA metabarcoding datasets by identifying a subset of informative taxa that are relevant for a certain task and discarding those that are redundant or irrelevant. However, general guidelines on selecting a feature selection method for application to a given setting are lacking. Here, we report a comparison of feature selection methods in a supervised machine learning setup across 13 environmental metabarcoding datasets with differing characteristics. We evaluate workflows that consist of data preprocessing, feature selection and a machine learning model by their ability to capture the ecological relationship between the microbial community composition and environmental parameters. Our results demonstrate that, while the optimal feature selection approach depends on dataset characteristics, feature selection is more likely to impair model performance than to improve it for tree ensemble models like Random Forests. Furthermore, our results show that calculating relative counts impairs model performance, which suggests that novel methods to combat the compositionality of metabarcoding data are required.