Abstract
Environmental DNA metabarcoding is a powerful approach for use in biomonitoring and impact assessments. Amplicon-based eDNA sequence data are characteristically highly divergent in sequencing depth (total reads per sample) as influenced inter alia by the number of samples simultaneously analyzed per sequencing run. The random forest (RF) machine learning algorithm has been successfully employed to accurately classify unknown samples into monitoring categories. To employ RF to eDNA data, and avoid sequencing-depth artifacts, sequence data across samples are normalized using rarefaction, a process that inherently loses information. The aim of this study was to inform future sampling designs in terms of the relationship between sampling depth and RF accuracy. We analyzed three published and one new bacterial amplicon datasets, using a RF, based initially on the maximal rarefied data available (minimum mean of > 30,000 reads across all datasets) to give our baseline performance. We then evaluated the RF classification success based on increasingly rarefied datasets. We found that extreme to moderate rarefaction (50-5000 sequences per sample) was sufficient to achieve prediction performance commensurate to the full data, depending on the classification task. We did not find that the number of classification classes, data balance across classes, or the total number of sequences or samples, were associated with predictive accuracy. We identified the ability of the training data to adequately characterize the classes being mapped as the most important criterion and discuss how this finding can inform future sampling design for eDNA based biomonitoring to reduce costs and computation time.