ORDNA: Deep-learning-based ordination for raw environmental DNA samples

Abstract

Environmental DNA (eDNA) metabarcoding has revolutionized biodiversity monitoring, offering non-invasive tools to assess ecosystem health. The complexity of eDNA metabarcoding data poses major challenges for conventional ordination methods in understanding assemblage similarities and assessing biodiversity patterns. Here, we introduce ORDNA (ORDination via Deep Neural Algorithm), a new deep learning method tailored for eDNA sample ordination. Leveraging artificial neural networks, ORDNA processes raw sequences from eDNA samples directly, bypassing potentially biased and cumbersome expert-based bioinformatic steps. The method is trained with a contrastive self-supervised learning approach, the triplet loss, to derive a two-dimensional representation of eDNA samples based on their read composition. We apply ORDNA to four distinct eDNA datasets, demonstrating its robustness and superiority over traditional ordination techniques in capturing and visualizing ecological patterns. Our results underline the potential of deep learning in advancing eDNA analysis, with ORDNA serving as a promising tool for more accurate and efficient biodiversity assessments.