End-to-End Artificial Learning of Protection Gradient from Raw eDNA Sequences

End-to-End Artificial Learning of Protection Gradient from Raw eDNA Sequences | Authorea try { document.documentElement.classList.add('js'); } catch (e) { } var _gaq = _gaq || []; _gaq.push(['_setAccount', 'G-8VDV14Y67G']); _gaq.push(['_trackPageview']); (function() { var ga = document.createElement('script'); ga.type = 'text/javascript'; ga.async = true; ga.src = ('https:' == document.location.protocol ? 'https://ssl' : 'http://www') + '.google-analytics.com/ga.js'; var s = document.getElementsByTagName('script')[0]; s.parentNode.insertBefore(ga, s); })(); Skip to main content Preprints Collections Wiley Open Research IET Open Research Ecological Society of Japan All Collections About About Authorea FAQs Contact Us Quick Search anywhere Search for preprint articles, keywords, etc. Search Search ADVANCED SEARCH SCROLL This is a preprint and has not been peer reviewed. Data may be preliminary. 17 September 2025 V1 Latest version Share on End-to-End Artificial Learning of Protection Gradient from Raw eDNA Sequences Authors : Letizia Lamperti 0000-0001-8059-1354 [email protected] , Theophile Sanchez , Steven Stalder 0009-0000-4568-8652 , David Mouillot 0000-0003-0402-2605 , Michele Volpi 0000-0003-2771-0750 , Camille Albouy , Morgane Bruno , Nino Molin , Olivier Francois 0000-0003-2402-2442 , Stephanie Manel 0000-0001-8902-6052 , and Loïc Pellissier 0000-0002-2289-8259 Authors Info & Affiliations https://doi.org/10.22541/au.175812518.84262613/v1 533 views 235 downloads Contents Abstract Supplementary Material Information & Authors Metrics & Citations View Options References Figures Tables Media Share Abstract To face the ever-increasing threats on ecosystems and their biodiversity, protected areas are expanding worldwide but assessing their effects requires advanced monitoring methods. Environmental DNA (eDNA) metabarcoding can reveal local biodiversity from genetic material released in the environment, offering ecosystem-health insights without invasive or destructive surveys. Yet, to harness the full potential of eDNA data in ecosystem assessment, robust and direct links between raw eDNA sequences and protection status remain to be learned with machine learning. Here, we present the first fully end-to-end framework learning and predicting a continuous protection gradient directly from raw eDNA sequences, bypassing taxonomic assignment or precomputed biodiversity indices. Our framework combines contrastive self-supervised pre-training to extract meaningful embeddings from eDNA sequences with a neural-network classifier to predict the level of protection. Applied to eDNA fish surveys on the Mediterranean coast, we show that raw sequences encode protection status along a gradient from unprotected to fully protected areas. This general framework can facilitate hypothesis testing on disturbance gradients and indicate whether a site is degraded or effectively protected, providing a scalable tool for ecosystem monitoring and adaptive management. 1 Introduction Most ecosystems on Earth face multiple threats such as climate change, pollution, resource exploitation and habitat destruction that endanger both biodiversity and the benefits nature provides to people (Sala et al., 2018; Duarte et al., 2020; Flandrin et al., 2024). So, the rapid decline of biodiversity is a critical global environmental issue that requires the establishment and reinforcement of protected areas to reach the ambitious target of 30% coverage on land and sea by 2030 (Hughes, 2023) but also advanced monitoring methodologies to assess the effectiveness of these conservation efforts (Kuempel et al., 2020; Wang et al., 2024). Yet, large-scale and accurate monitoring of biodiversity presents major logistical challenges and inherent constraints (Bianchi et al., 2022; Pauly & Zeller, 2016), particularly in protected areas where most invasive and destructive methods are forbidden (McGeady et al., 2023). Environmental DNA (eDNA) metabarcoding has emerged as a revolutionary method for detecting and monitoring biodiversity (Taberlet et al., 2012; Deiner et al., 2017; Polanco et al., 2021; Altermatt et al., 2025). This approach collects and analyzes genetic material shed into the environment by organisms and uses DNA barcode sequences to identify species present in the sampled environments (Miya, 2022). It provides detailed information on ecological communities without requiring direct observations or invasive capture methods (Beng & Corlett., 2020; Thomsen & Willerslev, 2015; Veron et al., 2023). eDNA metabarcoding offers numerous advantages, including enhanced detection of rare or cryptic species (Boussarie et al., 2018) but also standardized biodiversity assessments over large scales with minimal ecological impacts (e.g., Mathon et al., 2023). Moreover, various biodiversity indicators can be derived from eDNA metabarcoding, including taxonomic, phylogenetic or functional composition (Dalongeville et al., 2022; Condachou et al., 2023; He et al., 2024). Nonetheless, some biodiversity indicators derived from eDNA metabarcoding do not directly reflect conservation efforts as protection status (Dalongeville et al., 2022). Therefore, the vast amounts of data generated by eDNA metabarcoding, coupled with the inherent complexity of relationships between eDNA sequences and environmental factors (Harrison et al., 2019), require advanced analytical techniques to derive meaningful ecological indicators of threat or protection gradients (Boyer et al., 2016; Cordier et al., 2018; Blackman et al., 2024; Altermatt et al., 2025). The application of machine learning techniques to eDNA metabarcoding has shown potential for automating and improving the accuracy of ecological assessments (Hu et al., 2023). Machine learning, and particularly end-to-end deep learning, enables computational models to learn patterns directly from data and subsequently make predictions or decisions without explicit programming for specific tasks or data representation (Goodfellow et al., 2016). Both machine learning and deep learning have been successfully employed to ordinate eDNA data and predict ecological attributes, thereby increasing analytical efficiency and precision (Cordier et al., 2017; Lamperti et al., 2023; Keck et al., 2023; Zong et al., 2023; Lamperti et al., 2024; Sanchez et al., 2025). Yet, to date, there is no framework that allows the meaningful transfer of information contained in raw eDNA metabarcoding sequences directly into a reliable ecological indicator of threat or protection for ecosystem-scale monitoring and management. Associations between eDNA sequences and ecological properties are often context-dependent, with significant influence of temperature and habitat (Shea & Boehm, 2024; Perry et al., 2024; Lin et al., 2021; Rozanski et al., 2024). Consequently, eDNA-based indicators alone may be insufficient for fully assessing ecosystem health or conservation status, requiring additional contextual data—such as habitat characteristics, environmental factors, or human impacts— to develop accurate predictive models. However, existing frameworks for relating eDNA data to ecological indicators typically rely on traditional bioinformatics pipelines and classical machine learning algorithms (e.g., random forests), which are still affected by biases in PCR amplification, sequencing noise, and incomplete reference databases (Cordier et al., 2017; Creer et al., 2016; Bakker et al., 2019). While these frameworks are effective in specific contexts, they frequently fail to capture the complexity and multidimensionality inherent to eDNA datasets. End-to-end representation learning frameworks can overcome these limitations by modeling complexity directly from raw data, integrating environmental covariates, and mitigating potential biases in feature selection or interpretation. Here, we propose a novel end-to-end framework based on deep learning neural networks to analyze raw eDNA sequences and translate them into ecological indicators of threat or protection. In the first step, a self-supervised neural network learns meaningful data representations directly from raw eDNA sequences through an optimization process inspired by the Barlow Twins method (Zbontar et al., 2021). This approach does not need labels, so the pre-training step is de facto fully unsupervised. In the second step, a supervised neural network classifier is trained on those learned representations to predict a protection status, so the model can learn how labeled conditions map onto the latent representations. We hypothesize that training data accurately represent distinct protection statuses, so that the trained neural network can effectively indicate how closely eDNA raw sequences from previously unclassified sites match those of known levels of protection. To illustrate and validate our framework, we used 460 fish-eDNA samples from the western Mediterranean Sea (Boulanger et al., 2021; Dalongeville et al., 2022). We defined the protection status in two different ways: a binary scheme (protected vs. unprotected sites) and a five-level gradient (0 = no protection; 1–3 = increasing levels of protection; 4 = strict no-take reserve). Our approach relying on self-supervised pre-training was evaluated against a baseline model without pre-training. Environmental covariates, such as habitat, were integrated to further enhance predictive accuracy. Spatial autocorrelation has been explicitly accounted for when evaluating the model on the complete dataset, via a spatial-block cross-validation and related diagnostics, enhancing robustness and ensuring reliable predictions across spatial scales. This comprehensive validation demonstrates our framework’s capacity to link eDNA sequences to a gradient of protection, providing a rapid, accurate, and context‐aware assessment of ecosystem status across diverse environments. 2 Material and Methods 2.1 Case study 2.1.1 Mediterranean protection gradient We used a numerical protection gradient for Marine Protected Areas (MPAs), adapting the framework proposed by Grorud-Colvert et al. (2021). The gradient runs from 0 to 4: 0 denotes areas incompatible with conservation (i.e., unprotected; 300 samples), 1 minimally protected (61 samples), 2 lightly protected (40 samples), 3 highly protected (4 samples), and 4 fully protected or no-take reserves (55 samples) (Figure S1). This classification resulted in 160 samples within the protected classes (1–4). For analyses requiring a binary outcome, we dichotomized the gradient into unprotected (0) versus protected (1–4). This approach allowed us to clearly distinguish between protected and unprotected areas, thereby facilitating data analysis (Figure S1 a, c, d). 2.1.2 eDNA collection and sample processing We analyzed 460 eDNA samples collected in the French territorial waters of the Mediterranean Sea from May to October, from 2018 to 2023. The sampling covered the regions of Corsica (Corse; n = 377 samples), Provence-Alpes-Côte d’Azur (PACA; n = 55), and the Occitan coast (Occitanie; n = 28; Figure S1). Seawater samples were filtered following a sampling protocol optimized for monitoring coastal fish marine communities (Dalongeville et al., 2022; Boulanger et al., 2021). If a sampling transects overlapped multiple protection levels, we assigned the most restrictive category. eDNA extraction and amplification were performed at the SPYGEN facility. PCR amplification was performed using the teleo primer pair, targeting a 64 bp fragment of the mitochondrial DNA 12S rRNA gene specific for teleost fishes and elasmobranchs (Valentini et al., 2016). Data collection and sample processing are described in detail in Appendix S1. 2.1.3 Habitat environmental covariate The habitat variable captures the local benthic context at each sampling site. In this dataset, twelve habitat categories were identified from regional benthic cartography integrating substrate composition and biological cover. This categorical descriptor was one-hot encoded prior to modelling and used as an input covariate. 2.2 Self-supervised network and classifier for raw eDNA samples The framework consists of two components: a pre-trained self-supervised neural network for feature extraction from eDNA sequences and a classifier neural network for predicting the level of protection (Figure 1). The self-supervised network uses convolutional and self-attention layers to create fixed size embedding vectors that learn essential features directly from eDNA raw data without any labeled inputs. The classifier then takes these embedding vectors, along with covariates (e.g., the habitat), and predicts the protection level either on an ordinal scale from 0 (unprotected) to 4 (full no‐take) or as a binary outcome (protected vs. unprotected). 2.2.1 Self-supervised neural network for new representations of raw eDNA sequences This first model has two main components: a self-attention representation module and a multilayer perceptron (MLP) used for embedding generation (Figure 1a and b). Initially, raw eDNA sequences are processed through the representation module, designed to capture meaningful patterns within sequence data via self-attention mechanisms, producing a compact low-dimensional representation from raw eDNA sequences for each sample (Sanchez et al., 2025; Figure 1a). These representations are subsequently passed through a fully connected MLP consisting of three layers with ReLU activations and batch normalization, progressively reducing dimensionality from 2048 to the final embedding size (Figure 1b). The network is trained using a self-supervised learning framework inspired by the Barlow Twins (BT) method (Zbontar et al., 2021). Specifically, two non-overlapping subsets of N sequences each are extracted from every input (one eDNA sample), where each row corresponds to a single sequence. These subsets are obtained after a single initial random permutation of the rows. The corresponding embeddings \(z_{1}\)and \(z_{2}\)are then computed (Figure 1b) after being batch-standardised to zero mean and unit variance. A cross-correlation matrix \(C\ =\ \frac{1}{B}\ Z^{(1)T}Z^{(2)}\) , \(Z^{(k)}\in\ R^{B\ \times\ d}\) is calculated, where B is the batch size and d the embedding dimension, and \(Z^{(1)}\) and \(Z^{(2)}\ \)are the matrix of embeddings of dimension B x d . Optimization raises each diagonal element\(C_{\text{ii}}\) toward 1, making \(z_{1}\) and \(z_{2}\) as similar as possible, and drives every off-diagonal \(C_{\text{ij}}\)(i≠j) toward 0, keeping embedding dimensions independent. The resulting objective is \begin{equation} L_{\text{BT}}=\ \sum_{i=1}^{d}\left(1-C_{\text{ii}}\right)+\lambda\sum_{i=1}^{d}{\sum_{j=1\ j\neq i\ }^{d}C_{\text{ij}}^{2}}\nonumber \\ \end{equation} where λ modulates the decorrelation term. Parameters are optimized with AdamW, and the three loss components (total, diagonal, off-diagonal) are logged to monitor convergence. Hyper-parameter calibration on a held-out validation split yielded the following values, adopted for all experiments: subset size N = 500 sequences, batch size B = 32, embedding dimension d = 64, BT coefficient λ = 0.005, initial learning rate \(10^{-3}\), and weight decay\(10^{-4}\). Our choice of the Barlow-Twins objective departs fundamentally from the triplet-loss scheme adopted by Sanchez et al. (2025). Whereas triplet loss contrasts an anchor sample with an explicit positive and negative drawn from different labelled classes, Barlow Twins exploits the internal heterogeneity of each eDNA sample itself: two subsets from the same sample are treated as positive pairs, and no negative mining is required. This design better reflects the ecological premise that sequences within a single water sample capture the same latent community signal, while sequences from different samples can vary unpredictably. Consequently, optimization proceeds by maximizing within-sample similarity rather than between-sample contrast. 2.2.2 Ordinal Multiclass and Binary Classifiers To predict protection levels based on the embeddings generated by the self-supervised network, we implemented two supervised classifiers: a binary classifier (protected/unprotected), and an ordinal multiclass classifier (five-class ordinal gradient 0–4), both utilizing PyTorch (Paszke et al, 2017) and PyTorch Lightning (Falcon et al., 2019). Binary Classifier The binary classifier is used to distinguish protected from unprotected conditions. Its architecture consists of one hidden layer with 256 neurons, batch normalization, ReLU activation, and dropout (0.3). The final output is a single logit corresponding to the binary decision. The binary classification uses a BCEWithLogitsLoss , advantageous for numerical stability and handling class imbalance via class weighting. This binary framework provides a straightforward and rapid assessment of protection presence, which is valuable for initial screening. Ordinal Multiclass Classifier The ordinal classifier is designed to predict protection levels on a scale—in our case study, corresponding to protection status from unprotected (0) to no-take MPAs (4). The model architecture consists of a fully connected neural network with one hidden layer containing 256 neurons, batch normalization, ReLU activation, and dropout (0.3) for regularization. The final layer generates cumulative scores corresponding to the ordered nature of the target variable, following the CORAL framework for ordinal regression (Cao et al., 2020). The CORAL loss is chosen explicitly for our task as it inherently accounts for the ordinal structure of protection levels, treating the classification as a series of threshold-based binary decisions. This design helps the model capture relationships between adjacent categories, improving predictions’ interpretability and accuracy. To mitigate class imbalance, we applied class-specific weights to the CORAL loss, ensuring that under-represented classes contributed proportionally more to the optimization objective. The CORAL loss transforms the ordinal classification problem into multiple binary decisions, each corresponding to a cumulative threshold between adjacent classes. These decisions are evaluated using binary cross-entropy (BCE). For each sample i and each threshold k, the model outputs a scalar value \(z_{\text{ik}}\ \)which is passed through the sigmoid function σ(⋅) to produce a probability. The ground truth label\(y_{i}\ \)is converted into a binary target using the indicator function \(|(y_{i}\ >k)|\), which is 1 if the class of the sample exceeds the threshold \(k\), and 0 otherwise. The overall CORAL loss is computed as the average BCE across all thresholds and all samples: \begin{equation} \ L_{\text{CORAL}}=\ \frac{1}{N(K\ -\ 1)}\sum_{i=1}^{N}{\sum_{k=1}^{K-1}{BCE(I(y_{i}>k),\ \sigma(z_{\text{ik}}))}}\nonumber \\ \end{equation} where N is the batch size, K is the number of ordinal classes, and σ is the sigmoid activation function. This formulation ensures that the model learns a sequence of monotonic decision boundaries that respect the ordinal structure of the target variable. Both classifiers use the AdamW optimizer with an initial learning rate of \(10^{-3}\), coupled with a weight decay of \(10^{-4}\). Training was performed with a batch size of 10 and included early stopping based on validation loss to prevent overfitting. 2.3 Metrics and comparison 2.3.1 Spatial Partitioning for Cross-Validation To account for spatial autocorrelation and ensure robust unbiased generalization assessment (Roberts et al., 2016), we implemented a spatial k-fold cross-validation strategy using the spatialBlock function from the blockCV R package. Based on preliminary variogram analysis, a 30 km buffer was defined to control block size, outside of which no significant autocorrelation was found. After optimizing fold balance across 100 random iterations, we assigned each sampling location (n = 460) to one of five spatially contiguous folds (Figure 2a). During model training, for each cross-validation iteration, one-fold was held out as a validation set, while the remaining four folds were used to train the model (Figure 2b). This procedure ensured that each site served as a validation sample exactly once. Consequently, we obtained validation predictions for all 460 sites across the five folds. Model performance metrics—including accuracy, precision, recall, F1-score, MAE, and MSE—were computed within each fold and then averaged across folds, providing mean estimates that reflect the spatial structure of the data (Figure 2b). 2.3.2 Protection Ordinal Tolerance Accuracy We propose the Protection Ordinal Tolerance Accuracy (POTA), a novel performance measure developed for the ordinal multiclass framework, which accommodates the ordered nature of protection levels while preserving an absolute distinction for unprotected sites. Conventional accuracy counts only exact label matches as correct; by contrast, POTA enforces strict correctness for class 0 (unprotected), requiring that those sites be predicted as 0 without exception. For protected categories (classes 1–4), it then grants full credit (\(w_{i}\) = 1) when the prediction (pred) exactly matches the ground truth (gt), partial credit (\(w_{i}\) = α) when the prediction lies exactly one level away (∣pred − gt∣ = 1), and zero credit otherwise. In our implementation, we fixed α = 0.5 so that near misses receive half the credit of an exact match. Formally, for each sample i : \begin{equation} w_{i}=\left\{\begin{matrix}1,\ \ \&if\ \text{pred}_{i}=\text{gt}_{i}\\ \alpha,\ \ \&if\ {|pred}_{i}-\text{gt}_{i}|=1\ and\ \text{gt}_{i}\neq 0\\ 0,\ \ \&otherwhise\\ \end{matrix}\right.\ \nonumber \\ \end{equation} The POTA score is then the mean of all \(w_{i}\). By combining absolute fidelity for unprotected sites with graduated partial credit for near‐misses among protected levels, this metric provides a more nuanced and domain‐relevant assessment of ordinal classification performance than traditional accuracy alone. 2.3.3 Baseline for comparison To evaluate the utility of self-supervised network embeddings, we implemented a classification model that retains the architecture of the representation module (Figure 1a) but omits the Barlow Twins objective during training. In this configuration, the representation module transforms raw eDNA sequences into high-dimensional embeddings, which are passed directly to classifiers predicting conservation status. Unlike the self-supervised setting, the representation module is not pre-trained: it is trained jointly with the classifier, and the entire model is optimized exclusively using the classification loss, depending on whether the task is binary or ordinal multiclass. This design allows us to isolate and assess the contribution of self-supervised pretraining by comparing it against a model trained end-to-end on the supervised objective alone. 3 Results 3.1 Predictive performance of classifiers We evaluated the model’s ability to predict marine area protection status by comparing a binary classifier (protected vs. unprotected) and an ordinal multiclass classifier (five ordered protection levels) across spatial folds (Figure 3). In the absence of habitat information, the binary classifier exhibited superior nominal performance, with a mean accuracy of 94.0%, a precision of 0.83, a recall of 0.99, and a F1-score of 0.90 (POTA not applicable) across spatial folds. In contrast, the ordinal multiclass classifier attained a lower exact-match accuracy of 54.0% but achieved a POTA of 62.0%, indicating that many misclassifications fell within one adjacent protection level. We obtained a mean absolute error (MAE) of 0.68, with a macro-weighted precision of 0.75, a recall of 0.54, and a F1-score of 0.59. When habitat covariate was included, the binary model’s performance declined slightly—with a mean dropping to 92.0%, precision to 0.82, recall to 0.97, and F1-score to 0.87. Conversely, the ordinal multiclass classifier benefited substantially from habitat information with a mean exact-match accuracy rising to 72.0%, POTA increasing to 81.0%, MAE dropping to 0.40, while mean macro-weighted precision, recall, and F1-score reached up to 0.81, 0.72, and 0.78, respectively. 3.2 Effectiveness of the self-supervised network embeddings To benchmark the effectiveness of the self-supervised network embeddings, we compared the performance of the classifier model against a baseline model trained without embeddings from the self-supervised network (Table S1). Both models were trained on the same labeled dataset under comparable conditions until convergence and using early stopping to prevent overfitting. The binary classifier using self-supervised embeddings (mean accuracy: 94%, precision: 83%, recall: 99%, F1-score: 90%) performed better that the baseline model without pre-training, that reveals a mean accuracy of only 48%, with poor precision and recall (Table S1). These results highlight the importance of the self-supervised embeddings for model generalization when distinguishing between protected and unprotected sites. 3.3 Regional biases and the effect of habitat covariate on multiclass predictions Predicting protection levels across biogeographic regions revealed significant regional biases when habitat information was omitted (Figure 4). Without this covariate, Occitanie and PACA systematically under‐predicted the protection levels (median residuals +2.57 and +0.96 classes, respectively), with an exact‐match accuracy of only 35.7% and 18.2% and POTA of 35.7% and 55.5%, respectively. Corsica, by contrast, showed a slight over‐prediction trend (median residual - 0.28), achieving 59.7% accuracy and a POTA of 64.2%. The inclusion of habitat information helped to correct these biases. In Occitanie and Corsica, median residuals approached zero (+ 0.21 and + 0.28 classes, respectively), exact‐match accuracy rose to 85.7% and 80.1%, and POTA increased to 89.3% and 84.9%, respectively. In PACA, although an under-prediction of protection level persisted (median residual + 0.89), POTA held at 55.5% even as exact‐match accuracy improved slightly to 14.5%. Exact‐match accuracy remains very low despite the increased tolerance‐based performance. Supplementary analyses stratified by habitat type (Supplementary Figure 2B) confirmed these findings. In the model without habitat variables, prediction residuals varied widely across habitat categories and regions: in Corsica, mixed habitats such as coralligenous–meadow and rocks–sandy–bed were consistently over‐predicted (median residuals above + 1 class), while sandy beds showed slight under‐prediction of protection level. Similar patterns appeared in Occitanie and PACA, where some habitats were misclassified for their protection level by up to two classes on average. When habitat covariate was incorporated, all residuals tended to zero and variability across sites decreased. Across regions and habitat types, median residuals fell within ± 0.2 classes and interquartile ranges shrank to less than one class. Even the most difficult habitat (e.g., coralligenous meadow) showed a marked reduction in prediction error. Overall, including habitat covariate not only improved exact‐match accuracy and reduced systemic ordinal errors but also markedly increased tolerance‐based performance (POTA), resulting in more consistent and reliable protection level predictions across different regions and habitats. 3.4 Prediction errors across regions using habitat-augmented classifier A spatially explicit analysis of prediction errors highlighted distinct regional differences in model performance under the habitat-augmented ordinal multiclass model (Figure 5). Corsica and Occitanie exhibited high rates of exact-match classifications, reflecting strong model performance across their sampled areas. In Corsica, 302 out of 377 sites (80%) were accurately predicted, with the remaining errors distributed as +1 class for 37 sites, +2 classes for 22 sites, +3 classes for 14 sites, and +4 classes for 2 sites. Occitanie showed similar strong results, with 24 out of 28 sites (86%) correctly classified; the few errors consisted of two sites under-predicted by one protection class and two sites under-predicted by two classes. In contrast, PACA exhibited substantial under-prediction, with only 8 out of 55 sites (15%) accurately classified. Most errors in PACA involved under-prediction by one class (45 sites) and two classes (2 sites), corresponding to broader underestimation of protection level observed across the region. Analyzing the distribution of predicted protection levels further explained these patterns. In Corsica, most class-0 sites (286 observed) were over-represented in the predictions (313 predicted), indicating a slight tendency to under-predict high protection levels. Occitanie’s predicted distribution closely matched observations for class-0 sites but under-predicted the number of highly protected class-4 sites (14 predicted versus 18 observed). PACA showed the most pronounced bias, with a major shift towards predicting class-0 status (49 predicted class-0 sites versus only 4 observed) and a marked under-prediction of class-1 sites (2 predicted versus 43 observed). Together, these spatial and quantitative results demonstrate that adding habitat covariate markedly improves regional prediction accuracy, particularly in Corsica and Occitanie, effectively removing some systematic biases. However, the persistent under-prediction observed in PACA suggests that some regional heterogeneity in ecosystem features remains uncaptured by the current model (small-scale fisheries, sea currents). This combined spatial and quantitative perspective highlights both the strengths and the current limitations of the habitat-augmented network for geographically explicit ecosystem monitoring. 4 Discussion Global ecosystems are degrading at an alarming rate, faster than our current ability to monitor trends and implement conservation actions (Zhang et al., 2025; Keck et al., 2025; Pereira et al., 2024). Traditionally, the ecological integrity of ecosystems has been assessed through visual taxonomic identification and analysis of biological communities (Pawlowski et al., 2018). This process is labor-intensive and requires specialized expertise, limiting its applicability for large-scale monitoring. Our framework takes full advantage of the rich information embedded in raw environmental DNA samples. By bypassing the need for fully annotated genetic reference databases, which are notoriously incomplete (Marques et al., 2021; Schmid et al., 2025; Keck et al., 2023), it assesses ecosystem integrity directly from inherent sequence signals. The framework reliably reconstructs a protection gradient, from unprotected to fully protected sites. Moreover, the integration of habitat covariate further boosts predictive performance, particularly enhancing accuracy in intricate multiclass scenarios. We validate this framework on a Mediterranean eDNA survey, where protection levels can be accurately discriminated with spatial cross-validation procedures. Applying the model to all samples, we produced a general map of ecosystem integrity that highlights protection gradients and distinguishes highly protected areas from those impacted by anthropogenic stressors. This proof-of-concept demonstrates how an end-to-end deep-learning framework can transform raw eDNA into meaningful, reliable and interpretable indicators of conservation effort providing a scalable tool for scientists to test for ecological hypotheses and for managers and policymakers to monitor and anticipate ecosystem changes across diverse environments. End-to-end learning frameworks offer a powerful paradigm for biological data analysis, as they eliminate the need for manual features engineering and allow models to learn directly from raw, high-dimensional inputs (Eraslan et al., 2019; Zou et al., 2019). In combination with self-supervised pre-training, these approaches can extract informative representations from unlabeled data—an essential advantage in empirical field studies where annotated datasets are limited or costly to obtain. This is particularly relevant in ecological genomics, where environmental DNA (eDNA) provides rich but unstructured signals of community composition and ecosystem change. Recent studies show that self-supervised learning can uncover latent ecological structure from eDNA and improve generalization across complex environmental contexts (Sanchez et al., 2025). In our case, we observed a substantial performance improvement when comparing a classifier trained on self-supervised embeddings to a baseline model trained from scratch (94% vs. 48% accuracy). These findings support the utility of self-supervised end-to-end approaches as scalable and data-efficient solutions for ecological hypothesis testing and biodiversity monitoring. Communities in ecosystems are shaped by intricate interactions among species and socio-environmental contexts that define their composition and structure (Godoy et al., 2024; Mathon et al., 2023). In our models, incorporating habitat covariate substantially reduced prediction errors, particularly in the more challenging ordinal multiclass classifications. Key environmental variables such as water temperature, salinity, nutrient availability, and anthropogenic pressure are critical in determining which species can thrive and how they interact (Levesque, 2019). These socio-environmental factors also influence ecological dynamics such as predation and competition for resources (Pinault et al., 2014). Thus, the flexibility of machine learning algorithms—especially deep learning—allows for automatic optimization analysis of massive datasets, making them highly effective in ecological applications where relationships between species and socio-environmental conditions are non-linear and multidimensional (Pichler & Hartig, 2023). On the other hand, the ability of such models to learn targets fully depends on the quality and representativeness of the inputs. Our model’s capacity to integrate detailed habitat features illustrates how deep learning can be leveraged for comprehensive and accurate assessments of marine ecosystem status and offers a simple and flexible way to include more relevant covariates under a self-supervised learning pretraining. In our marine case study of 460 fish eDNA samples spanning protection gradients along the Mediterranean coast, the contrast between protected and unprotected areas provided a strong ecological signal that can be effectively captured by the binary neural network model. The binary model achieved 94% accuracy, confirming that protection status is associated with measurable differences in eDNA signal (Figure 3). This aligns with previous findings showing that MPAs significantly modify fish community composition relative to surrounding areas without changing the biodiversity level (Boulanger et al., 2021; Duarte et al., 2020; Lamperti et al., 2023). In the ordinal classification setting, most misclassifications occurred between adjacent protection levels, indicating that the model captures well the ordinal structure of protection levels and reflects an ecologically coherent representation of the protection gradient (Figures 3 and 5). Nonetheless, fine-scale distinctions between intermediate categories remain challenging, potentially due to ecological overlaps between classes or inconsistencies in the scoring system itself. The inclusion of habitat covariate substantially improved model accuracy, particularly in the ordinal multiclass scenario. To contextualize these results, we also evaluated a model trained exclusively on habitat (Appendix 2). Our results also demonstrate clear regional variability in model accuracy and predictive performance, possibly due to differences in levels of anthropogenic pressure, or specific local ecological characteristics such as substrate type, artisanal fisheries, tourism, sea currents or regional species composition. For example, the model predictions for Corsica and Occitanie were highly accurate, while predictions for the PACA region highlighted persistent underestimation, suggesting that unique ecological or anthropogenic conditions remain inadequately captured (Figure 4). Even when habitat covariate was included, 78% of misclassifications in this region correspond to an error of only one protection level, indicating that the model recovered the ordinal structure of the gradient, but with a consistent bias. Notably, 41 out of 43 samples labeled as minimally protected were misclassified as not protected, revealing a downward bias. Only the highest protected sites were classified correctly, suggesting a strong biological signal at high protection, but weaker signals at intermediate protection levels. Results remained similar when habitat covariate was excluded, indicating that this pattern is not solely attributable to missing environmental descriptors. The potential causes of this phenomenon may include limitations in the coverage of the training dataset, the detectability of biological gradients through eDNA analysis alone in this region, and the possible consequences of inadequate management practices in the context of protection (Seguin et al., 2025). Despite robust performance in our case study, some residual errors in converting eDNA signals into ecological indicators remain unexplained. The potential causes of these variations may include label uncertainty, temporal dynamics not captured by single time‐point sampling, or unmodeled environmental heterogeneity. Incorporating additional temporal and local covariates (e.g., seasonal time-series data, local anthropogenic pressures etc.) could assist in the reduction of this residual variance. Moreover, integrating interpretable machine learning approaches could enhance both predictive accuracy and the transparency of ecological inference, helping to identify key drivers of community structure and ecosystem change (Natarajan et al., 2024). Above all, our study highlights the transformative potential of combining eDNA metabarcoding with end-to-end deep learning to enhance monitoring and assessment of ecosystems. By simplifying the traditional post-processing workflow—bypassing the need for taxonomic identification or extensive reference databases—our approach enables a more direct and scalable evaluation of biodiversity and protection status. Expanding the geographic scope of sampling across diverse ecosystems will further improve model generalizability and ecological relevance. Despite remaining uncertainties, this integrative framework—uniting ecological genomics, environmental characterization, and machine learning—offers a high-resolution, transferable approach to biodiversity monitoring and delivering actionable insights for ecosystem management across biomes in the context of increasing anthropogenic pressures but also conservation efforts. Acknowledgments We would like to thank the SDSC project DNAi, as in our previous study, as well as the SNF-ANR shifteDNA project for their support. This research was funded by the IA-Biodiv ANR FISH-PREDICT project (ANR-21-AAFI-0001-01), by the Biodiversa+ BioBoost+ project (EU grant agreement 101052342), by the ANR ShifteDNA, and by the Predictive Ecological Genomics project (PEG2, ANR-22-CE45-0033). The eDNA data were collected during field campaigns supported by the Agence de l’Eau Rhône Méditerranée. We are also grateful to GRICAD for providing intensive computing resources, and to all MPA managers and volunteers who contributed to the fieldwork. Bibliography Altermatt, F., Couton, M., Carraro, L., Keck, F., Lawson-Handley, L., Leese, F., Zhang, X., Zhang, Y., Blackman, R.C., 2025. Utilizing aquatic environmental DNA to address global biodiversity targets. Nat. Rev. Biodivers. 1, 332–346. https://doi.org/10.1038/s44358-025-00044-xBakker, J., Wangensteen, O.S., Baillie, C., Buddo, D., Chapman, D.D., Gallagher, A.J., Guttridge, T.L., Hertler, H., Mariani, S., 2019. Biodiversity assessment of tropical shelf eukaryotic communities via pelagic eDNA metabarcoding. Ecoloy and Evolution 9, 14341–14355. https://doi.org/10.1002/ece3.5871Beng, K.C., Corlett, R.T., 2020. Applications of environmental DNA (eDNA) in ecology and conservation: opportunities, challenges and prospects. Biodivers Conserv 29, 2089–2121. https://doi.org/10.1007/s10531-020-01980-0Bianchi, C.N., Azzola, A., Cocito, S., Morri, C., Oprandi, A., Peirano, A., Sgorbini, S., Montefalcone, M., 2022. Biodiversity Monitoring in Mediterranean Marine Protected Areas: Scientific and Methodological Challenges. Diversity 14, 43. https://doi.org/10.3390/d14010043Blackman, R., Couton, M., Keck, F., Kirschner, D., Carraro, L., Cereghetti, E., Perrelet, K., Bossart, R., Brantschen, J., Zhang, Y., Altermatt, F., 2024. Environmental DNA: The next chapter. Molecular Ecology 33, e17355. https://doi.org/10.1111/mec.17355Boulanger, E., Loiseau, N., Valentini, A., Arnal, V., Boissery, P., Dejean, T., Deter, J., Guellati, N., Holon, F., Juhel, J.-B., Lenfant, P., Manel, S., Mouillot, D., 2021. Environmental DNA metabarcoding reveals and unpacks a biodiversity conservation paradox in Mediterranean marine reserves. Proc. R. Soc. B. 288, rspb.2021.0112, 20210112. https://doi.org/10.1098/rspb.2021.0112Boussarie et al., 2018 - Cerca con Google [WWW Document], n.d. URL https://www.science.org/doi/10.1126/sciadv.aap9661 (accessed 9.3.25).Boussarie, G., Bakker, J., Wangensteen, O.S., Mariani, S., Bonnin, L., Juhel, J.-B., Kiszka, J.J., Kulbicki, M., Manel, S., Robbins, W.D., Vigliola, L., Mouillot, D., 2018. Environmental DNA illuminates the dark diversity of sharks. Science Advances 4, eaap9661. https://doi.org/10.1126/sciadv.aap9661Boyer, F., Mercier, C., Bonin, A., Le Bras, Y., Taberlet, P., Coissac, E., 2016. obitools: a unix-inspired software package for DNA metabarcoding. Mol Ecol Resour 16, 176–182. https://doi.org/10.1111/1755-0998.12428Cao, W., Mirjalili, V., Raschka, S., 2020. Rank consistent ordinal regression for neural networks with application to age estimation. Pattern Recognition Letters 140, 325–331. https://doi.org/10.1016/j.patrec.2020.11.008Condachou, C., Milhau, T., Murienne, J., Brosse, S., Villéger, S., Valentini, A., Dejean, T., Mouillot, D., 2023. Inferring functional diversity from environmental DNA metabarcoding. Environmental DNA 5, 934–944. https://doi.org/10.1002/edn3.391Cordier, T., Esling, P., Lejzerowicz, F., Visco, J., Ouadahi, A., Martins, C., Cedhagen, T., Pawlowski, J., 2017. Predicting the Ecological Quality Status of Marine Environments from eDNA Metabarcoding Data Using Supervised Machine Learning. Environ. Sci. Technol. 51, 9118–9126. https://doi.org/10.1021/acs.est.7b01518Cordier, T., Forster, D., Dufresne, Y., Martins, C.I.M., Stoeck, T., Pawlowski, J., 2018. Supervised machine learning outperforms taxonomy-based environmental DNA metabarcoding applied to biomonitoring. Mol Ecol Resour 18, 1381–1391. https://doi.org/10.1111/1755-0998.12926Cowman, P.F., Bellwood, D.R., 2013. The historical biogeography of coral reef fishes: global patterns of origination and dispersal. Journal of Biogeography 40, 209–224. https://doi.org/10.1111/jbi.12003Creer, S., Deiner, K., Frey, S., Porazinska, D., Taberlet, P., Thomas, W.K., Potter, C., Bik, H.M., 2016. The ecologist’s field guide to sequence-based identification of biodiversity. Methods in Ecology and Evolution 7, 1008–1018. https://doi.org/10.1111/2041-210X.12574Dalongeville, A., Boulanger, E., Marques, V., Charbonnel, E., Hartmann, V., Santoni, M.C., Deter, J., Valentini, A., Lenfant, P., Boissery, P., Dejean, T., Velez, L., Pichot, F., Sanchez, L., Arnal, V., Bockel, T., Delaruelle, G., Holon, F., Milhau, T., Romant, L., Manel, S., Mouillot, D., 2022. Benchmarking eleven biodiversity indicators based on environmental DNA surveys: More diverse functional traits and evolutionary lineages inside marine reserves. Journal of Applied Ecology 59, 2803–2813. https://doi.org/10.1111/1365-2664.14276Deiner, K., Bik, H.M., Mächler, E., Seymour, M., Lacoursière-Roussel, A., Altermatt, F., Creer, S., Bista, I., Lodge, D.M., de Vere, N., Pfrender, M.E., Bernatchez, L., 2017. Environmental DNA metabarcoding: Transforming how we survey animal and plant communities. Molecular Ecology 26, 5872–5895. https://doi.org/10.1111/mec.14350Duarte, C.M., Agusti, S., Barbier, E., Britten, G.L., Castilla, J.C., Gattuso, J.-P., Fulweiler, R.W., Hughes, T.P., Knowlton, N., Lovelock, C.E., Lotze, H.K., Predragovic, M., Poloczanska, E., Roberts, C., Worm, B., 2020. Rebuilding marine life. Nature 580, 39–51. https://doi.org/10.1038/s41586-020-2146-7Eraslan, G., Avsec, Ž., Gagneur, J., Theis, F.J., 2019. Deep learning: new computational modelling techniques for genomics. Nat Rev Genet 20, 389–403. https://doi.org/10.1038/s41576-019-0122-6Falcon, W., The PyTorch Lightning team, 2019. PyTorch Lightning. https://doi.org/10.5281/zenodo.3828935Flandrin, U., Mouillot, D., Albouy, C., Bejarano, S., Casajus, N., Cinner, J., Edgar, G., Ghilardi, M., Leprieur, F., Loiseau, N., MacNeil, A., Maire, E., McLean, M., Parravicini, V., Pellissier, L., Schiettekatte, N., Stuart-Smith, R.D., Villéger, S., Mouquet, N., 2024. Fish communities can simultaneously contribute to nature and people across the world’s tropical reefs. One Earth 7, 1772–1785. https://doi.org/10.1016/j.oneear.2024.09.011Godoy, O., Soler-Toscano, F., Portillo, J.R., Langa, J.A., 2024. The assembly and dynamics of ecological communities in an ever-changing world. Ecological Monographs 94, e1633. https://doi.org/10.1002/ecm.1633Goodfellow, I., Yoshua Bengio, Aaron Courville, 2016. Deep Learning. MIT press.Grorud-Colvert, K., Sullivan-Stack, J., Roberts, C., Constant, V., Horta e Costa, B., Pike, E.P., Kingston, N., Laffoley, D., Sala, E., Claudet, J., Friedlander, A.M., Gill, D.A., Lester, S.E., Day, J.C., Gonçalves, E.J., Ahmadia, G.N., Rand, M., Villagomez, A., Ban, N.C., Gurney, G.G., Spalding, A.K., Bennett, N.J., Briggs, J., Morgan, L.E., Moffitt, R., Deguignet, M., Pikitch, E.K., Darling, E.S., Jessen, S., Hameed, S.O., Di Carlo, G., Guidetti, P., Harris, J.M., Torre, J., Kizilkaya, Z., Agardy, T., Cury, P., Shah, N.J., Sack, K., Cao, L., Fernandez, M., Lubchenco, J., 2021. The MPA Guide: A framework to achieve global goals for the ocean. Science 373, eabf0861. https://doi.org/10.1126/science.abf0861Harrison, J.B., Sunday, J.M., Rogers, S.M., 2019. Predicting the fate of eDNA in the environment and implications for studying biodiversity. Proc Biol Sci 286, 20191409. https://doi.org/10.1098/rspb.2019.1409He, Y., Zhao, X., Shi, C., Peng, K., Wang, Z., Jiang, Z., 2024. Fish community monitoring in floodplain lakes: eDNA metabarcoding and traditional sampling revealed inconsistent fish community composition. Ecological Indicators 166, 112467. https://doi.org/10.1016/j.ecolind.2024.112467Hu, H., Wei, X.-Y., Liu, L., Wang, Y.-B., Jia, H.-J., Bu, L.-K., Pei, D.-S., 2023. Supervised machine learning improves general applicability of eDNA metabarcoding for reservoir health monitoring. Water Research 246, 120686. https://doi.org/10.1016/j.watres.2023.120686Hughes, A.C., 2023. The Post-2020 Global Biodiversity Framework: How did we get here, and where do we go next? Integrative Conservation 2, 1–9. https://doi.org/10.1002/inc3.16Keck, F., Brantschen, J., Altermatt, F., 2023a. A combination of machine-learning and eDNA reveals the genetic signature of environmental change at the landscape levels. Mol Ecol 32, 4791–4800. https://doi.org/10.1111/mec.17073Keck, F., Couton, M., Altermatt, F., 2023b. Navigating the seven challenges of taxonomic reference databases in metabarcoding analyses. Molecular Ecology Resources 23, 742–755. https://doi.org/10.1111/1755-0998.13746Keck, F., Peller, T., Alther, R., Barouillet, C., Blackman, R., Capo, E., Chonova, T., Couton, M., Fehlinger, L., Kirschner, D., Knüsel, M., Muneret, L., Oester, R., Tapolczai, K., Zhang, H., Altermatt, F., 2025. The global human impact on biodiversity. Nature 641, 395–400. https://doi.org/10.1038/s41586-025-08752-2Kim, K.G., 2016. Book Review: Deep Learning. Healthc Inform Res 22, 351. https://doi.org/10.4258/hir.2016.22.4.351Kuempel, C.D., Chauvenet, A.L.M., Possingham, H.P., Adams, V.M., 2020. Evidence-Based Guidelines for Prioritizing Investments to Meet International Conservation Objectives. One Earth 2, 55–63. https://doi.org/10.1016/j.oneear.2019.12.013Lamperti, L., François, O., Mouillot, D., Mathon, L., Sanchez, T., Albouy, C., Pellissier, L., Manel, S., 2024. A spatial matrix factorization method to characterize ecological assemblages as a mixture of unobserved sources: An application to fish eDNA surveys. Methods in Ecology and Evolution 15, 2301–2315. https://doi.org/10.1111/2041-210X.14430Lamperti, L., Sanchez, T., Si Moussi, S., Mouillot, D., Albouy, C., Flück, B., Bruno, M., Valentini, A., Pellissier, L., Manel, S., 2023. New deep learning‐based methods for visualizing ecosystem properties using environmental DNA metabarcoding data. Molecular Ecology Resources 23, 1946–1958. https://doi.org/10.1111/1755-0998.13861Levesque, J.C., 2019. Spatio-temporal patterns of the oceanic conditions and nearshore marine community in the Mid-Atlantic Bight (New Jersey, USA). PeerJ 7, e7927. https://doi.org/10.7717/peerj.7927Lin, M., Simons, A.L., Harrigan, R.J., Curd, E.E., Schneider, F.D., Ruiz-Ramos, D.V., Gold, Z., Osborne, M.G., Shirazi, S., Schweizer, T.M., Moore, T.N., Fox, E.A., Turba, R., Garcia-Vedrenne, A.E., Helman, S.K., Rutledge, K., Mejia, M.P., Marwayana, O., Munguia Ramos, M.N., Wetzer, R., Pentcheff, N.D., McTavish, E.J., Dawson, M.N., Shapiro, B., Wayne, R.K., Meyer, R.S., 2021. Landscape analyses using eDNA metabarcoding and Earth observation predict community biodiversity in California. Ecological Applications 31, e02379. https://doi.org/10.1002/eap.2379Marques, V., Milhau, T., Albouy, C., Dejean, T., Manel, S., Mouillot, D., Juhel, J.-B., 2021. GAPeDNA: Assessing and mapping global species gaps in genetic databases for eDNA metabarcoding. Diversity and Distributions 27, 1880–1892. https://doi.org/10.1111/ddi.13142Mathon, L., Marques, V., Manel, S., Albouy, C., Andrello, M., Boulanger, E., Deter, J., Hocdé, R., Leprieur, F., Letessier, T.B., Loiseau, N., Maire, E., Valentini, A., Vigliola, L., Baletaud, F., Bessudo, S., Dejean, T., Faure, N., Guerin, P.-E., Jucker, M., Juhel, J.-B., Kadarusman, Polanco F., A., Pouyaud, L., Schwörer, D., Thompson, K.F., Troussellier, M., Sugeha, H.Y., Velez, L., Zhang, X., Zhong, W., Pellissier, L., Mouillot, D., 2023. The distribution of coastal fish eDNA sequences in the Anthropocene. Global Ecology and Biogeography 32, 1336–1352. https://doi.org/10.1111/geb.13698McGeady, R., Runya, R.M., Dooley, J.S.G., Howe, J.A., Fox, C.J., Wheeler, A.J., Summers, G., Callaway, A., Beck, S., Brown, L.S., Dooly, G., McGonigle, C., 2023. A review of new and existing non-extractive techniques for monitoring marine protected areas. Front. Mar. Sci. 10. https://doi.org/10.3389/fmars.2023.1126301Miya, M., 2022. Environmental DNA Metabarcoding: A Novel Method for Biodiversity Monitoring of Marine Fish Communities. Ann Rev Mar Sci 14, 161–185. https://doi.org/10.1146/annurev-marine-041421-082251Natarajan, G., Elango, E., Gnanasekaran, R., Soman, S., 2024. Explainable Artificial Intelligence for Ocean Health: Applications and Challenges, in: De, D., Sengupta, D., Tran, T.A. (Eds.), Artificial Intelligence and Edge Computing for Sustainable Ocean Health. Springer Nature Switzerland, Cham, pp. 241–270. https://doi.org/10.1007/978-3-031-64642-3_11Paszke, A., et al. (2017) Automatic Differentiation in PyTorch. 31st Conference on Neural Information Processing Systems (NIPS 2017), Long Beach, CA, USA. - References - Scientific Research Publishing [WWW Document], n.d. URL https://www.scirp.org/reference/referencespapers?referenceid=3657867 (accessed 9.3.25).Pauly, D., Zeller, D., 2016. Catch reconstructions reveal that global marine fisheries catches are higher than reported and declining. Nat Commun 7, 10244. https://doi.org/10.1038/ncomms10244Pawlowski, J., Kelly-Quinn, M., Altermatt, F., Apothéloz-Perret-Gentil, L., Beja, P., Boggero, A., Borja, A., Bouchez, A., Cordier, T., Domaizon, I., Feio, M.J., Filipe, A.F., Fornaroli, R., Graf, W., Herder, J., van der Hoorn, B., Iwan Jones, J., Sagova-Mareckova, M., Moritz, C., Barquín, J., Piggott, J.J., Pinna, M., Rimet, F., Rinkevich, B., Sousa-Santos, C., Specchia, V., Trobajo, R., Vasselon, V., Vitecek, S., Zimmerman, J., Weigand, A., Leese, F., Kahlert, M., 2018. The future of biotic indices in the ecogenomic era: Integrating (e)DNA metabarcoding in biological assessment of aquatic ecosystems. Science of The Total Environment 637–638, 1295–1310. https://doi.org/10.1016/j.scitotenv.2018.05.002Pereira, H.M., Martins, I.S., Rosa, I.M.D., Kim, H., Leadley, P., Popp, A., van Vuuren, D.P., Hurtt, G., Quoss, L., Arneth, A., Baisero, D., Bakkenes, M., Chaplin-Kramer, R., Chini, L., Di Marco, M., Ferrier, S., Fujimori, S., Guerra, C.A., Harfoot, M., Harwood, T.D., Hasegawa, T., Haverd, V., Havlík, P., Hellweg, S., Hilbers, J.P., Hill, S.L.L., Hirata, A., Hoskins, A.J., Humpenöder, F., Janse, J.H., Jetz, W., Johnson, J.A., Krause, A., Leclère, D., Matsui, T., Meijer, J.R., Merow, C., Obersteiner, M., Ohashi, H., De Palma, A., Poulter, B., Purvis, A., Quesada, B., Rondinini, C., Schipper, A.M., Settele, J., Sharp, R., Stehfest, E., Strassburg, B.B.N., Takahashi, K., Talluto, L., Thuiller, W., Titeux, N., Visconti, P., Ware, C., Wolf, F., Alkemade, R., 2024. Global trends and scenarios for terrestrial biodiversity and ecosystem services from 1900 to 2050. Science 384, 458–465. https://doi.org/10.1126/science.adn3441Perry, W.B., Seymour, M., Orsini, L., Jâms, I.B., Milner, N., Edwards, F., Harvey, R., de Bruyn, M., Bista, I., Walsh, K., Emmett, B., Blackman, R., Altermatt, F., Lawson Handley, L., Mächler, E., Deiner, K., Bik, H.M., Carvalho, G., Colbourne, J., Cosby, B.J., Durance, I., Creer, S., 2024. An integrated spatio-temporal view of riverine biodiversity using environmental DNA metabarcoding. Nat Commun 15, 4372. https://doi.org/10.1038/s41467-024-48640-3Pichler, M., Hartig, F., 2023. Machine learning and deep learning—A review for ecologists. Methods in Ecology and Evolution 14, 994–1016. https://doi.org/10.1111/2041-210X.14061Pinault, M., Bissery, C., Gassiole, G., Magalon, H., Quod, J.-P., Galzin, R., 2014. Fish community structure in relation to environmental variation in coastal volcanic habitats. Journal of Experimental Marine Biology and Ecology 460, 62–71. https://doi.org/10.1016/j.jembe.2014.06.005Polanco F., A., Mutis Martinezguerra, M., Marques, V., Villa-Navarro, F., Borrero Pérez, G.H., Cheutin, M.-C., Dejean, T., Hocdé, R., Juhel, J.-B., Maire, E., Manel, S., Spescha, M., Valentini, A., Mouillot, D., Albouy, C., Pellissier, L., 2021. Detecting aquatic and terrestrial biodiversity in a tropical estuary using environmental DNA. Biotropica 53, 1606–1619. https://doi.org/10.1111/btp.13009Roberts, D.R., Bahn, V., Ciuti, S., Boyce, M.S., Elith, J., Guillera-Arroita, G., Hauenstein, S., Lahoz-Monfort, J.J., Schröder, B., Thuiller, W., Warton, D.I., Wintle, B.A., Hartig, F., Dormann, C.F., 2017. Cross-validation strategies for data with temporal, spatial, hierarchical, or phylogenetic structure. Ecography 40, 913–929. https://doi.org/10.1111/ecog.02881Rozanski, R., Velez, L., Hocdé, R., Duhamet, A., Waldock, C., Mouillot, D., Pellissier, L., Albouy, C., 2024. Seasonal dynamics of Mediterranean fish communities revealed by eDNA: Contrasting compositions across depths and Marine Fully Protected Area boundaries. Ecological Indicators 166, 112290. https://doi.org/10.1016/j.ecolind.2024.112290Sala, E., Mayorga, J., Costello, C., Kroodsma, D., Palomares, M.L.D., Pauly, D., Sumaila, U.R., Zeller, D., 2018. The economics of fishing the high seas. Science Advances 4, eaat2504. https://doi.org/10.1126/sciadv.aat2504Sanchez, T., Stalder, S., Lamperti, L., Brosse, S., Frossard, A., Leugger, F., Rozanski, R., Zong, S., Manel, S., Medici, L., Kuhn, F., Han, X., Mestrot, A., Albouy, C., Volpi, M., Pellissier, L., n.d. ORDNA: Deep-learning-based ordination for raw environmental DNA samples. Methods in Ecology and Evolution n/a. https://doi.org/10.1111/2041-210X.70033Schmid, S., Straube, N., Albouy, C., Delling, B., Maclaine, J., Matschiner, M., Møller, P.R., Nocita, A., Palandačić, A., Rüber, L., Sonnewald, M., Alvarez, N., Manel, S., Pellissier, L., 2025. Combining eDNA and Museomics to Enhance Biodiversity Monitoring.Seguin, R., Le Manach, F., Devillers, R., Velez, L., Mouillot, D., 2025. Global patterns and drivers of untracked industrial fishing in coastal marine protected areas. Science 389, 396–401. https://doi.org/10.1126/science.ado9468Shea, M.M., Boehm, A.B., 2024. Environmental DNA metabarcoding differentiates between micro-habitats within the rocky intertidal. Environmental DNA 6, e521. https://doi.org/10.1002/edn3.521Taberlet, P., Coissac, E., Hajibabaei, M., Rieseberg, L.H., 2012. Environmental DNA. Molecular Ecology 21, 1789–1793. https://doi.org/10.1111/j.1365-294X.2012.05542.xThomsen, P.F., Willerslev, E., 2015. Environmental DNA – An emerging tool in conservation for monitoring past and present biodiversity. Biological Conservation, Special Issue: Environmental DNA: A powerful new tool for biological conservation 183, 4–18. https://doi.org/10.1016/j.biocon.2014.11.019Valentini, A., Taberlet, P., Miaud, C., Civade, R., Herder, J., Thomsen, P.F., Bellemain, E., Besnard, A., Coissac, E., Boyer, F., Gaboriaud, C., Jean, P., Poulet, N., Roset, N., Copp, G.H., Geniez, P., Pont, D., Argillier, C., Baudoin, J.-M., Peroux, T., Crivelli, A.J., Olivier, A., Acqueberge, M., Le Brun, M., Møller, P.R., Willerslev, E., Dejean, T., 2016. Next-generation monitoring of aquatic biodiversity using environmental DNA metabarcoding. Mol Ecol 25, 929–942. https://doi.org/10.1111/mec.13428Veron, P., Rozanski, R., Marques, V., Joost, S., Deschez, M.E., Trenkel, V.M., Lorance, P., Valentini, A., Polanco F., A., Pellissier, L., Eme, D., Albouy, C., 2023. Environmental DNA complements scientific trawling in surveys of marine fish biodiversity. ICES J Mar Sci 80, 2150–2165. https://doi.org/10.1093/icesjms/fsad139Wang, J., Bai, Y., Huang, Z., Ashraf, A., Ali, M., Fang, Z., Lu, X., 2024. Identifying ecological security patterns to prioritize conservation and restoration：A case study in Xishuangbanna tropical region, China. Journal of Cleaner Production 444, 141222. https://doi.org/10.1016/j.jclepro.2024.141222Zbontar, J., Jing, L., Misra, I., LeCun, Y., Deny, S., 2021. Barlow Twins: Self-Supervised Learning via Redundancy Reduction, in: Proceedings of the 38th International Conference on Machine Learning. Presented at the International Conference on Machine Learning, PMLR, pp. 12310–12320.Zhang, M., Sun, J., Wang, Y., Li, Y., Duo, J., 2025. State-of-the-art and challenges in global grassland degradation studies. Geography and Sustainability 6, 100229. https://doi.org/10.1016/j.geosus.2024.08.008Zong, S., Brantschen, J., Zhang, X., Albouy, C., Valentini, A., Zhang, H., Altermatt, F., Pellissier, L., 2024. Combining environmental DNA with remote sensing variables to map fish species distributions along a large river. Remote Sensing in Ecology and Conservation 10, 220–235. https://doi.org/10.1002/rse2.366Zou, Y., Lan, R., 2020. An End-to-End Calibration Method for Welding Robot Laser Vision Systems With Deep Reinforcement Learning. IEEE Transactions on Instrumentation and Measurement 69, 4270–4280. https://doi.org/10.1109/TIM.2019.2942533 Supplementary Material File (figures_paper_ele_def.pdf) Download 1.78 MB Information & Authors Information Version history V1 Version 1 17 September 2025 Copyright This work is licensed under a Non Exclusive No Reuse License. Keywords biodiversity monitoring deep learning ecological disturbances environmental dna neural networks self-supervised learning Authors Affiliations Letizia Lamperti 0000-0001-8059-1354 [email protected] EPHE PSL View all articles by this author Theophile Sanchez ETH Domain Federal Institutes of Technology View all articles by this author Steven Stalder 0009-0000-4568-8652 Swiss Data Science Center View all articles by this author David Mouillot 0000-0003-0402-2605 Université Montpellier-CNRS-IFREMER View all articles by this author Michele Volpi 0000-0003-2771-0750 Swiss Data Science Center View all articles by this author Camille Albouy ETH Zurich View all articles by this author Morgane Bruno CEFE View all articles by this author Nino Molin CEFE View all articles by this author Olivier Francois 0000-0003-2402-2442 Université Grenoble Alpes Grenoble INP Institut d’Ingénierie et de Management View all articles by this author Stephanie Manel 0000-0001-8902-6052 Ecole Pratique des Hautes Etudes Section des Sciences de la Vie et de la Terre View all articles by this author Loïc Pellissier 0000-0002-2289-8259 ETH Zurich View all articles by this author Metrics & Citations Metrics Article Usage 533 views 235 downloads .FvxKWukQNSOunydq8rnd { width: 100px; } Citations Download citation Letizia Lamperti, Theophile Sanchez, Steven Stalder, et al. End-to-End Artificial Learning of Protection Gradient from Raw eDNA Sequences. Authorea . 17 September 2025. DOI: https://doi.org/10.22541/au.175812518.84262613/v1 If you have the appropriate software installed, you can download article citation data to the citation manager of your choice. Simply select your manager software from the list below and click Download. For more information or tips please see 'Downloading to a citation manager' in the Help menu . Format Please select one from the list RIS (ProCite, Reference Manager) EndNote BibTex Medlars RefWorks Direct import Tips for downloading citations document.getElementById('citMgrHelpLink').addEventListener('click', function() { popupHelp(this.href); return false; }); $(".js__slcInclude").on("change", function(e){ if ($(this).val() == 'refworks') $('#direct').prop("checked", false); $('#direct').prop("disabled", ($(this).val() == 'refworks')); }); View Options View options PDF View PDF Figures Tables Media Share Share Share article link Copy Link Copied! Copying failed. Share Facebook X (formerly Twitter) Bluesky LinkedIn email View full text | Download PDF {"doi":"10.22541/au.175812518.84262613/v1","type":"Article"} Now Reading: Share Figures Tables Close figure viewer Back to article Figure title goes here Change zoom level Go to figure location within the article Download figure Toggle share panel Toggle share panel Share Toggle information panel Toggle information panel Go to previous graphic Go to next graphic Go to previous table Go to next table All figures All tables View all material View all material xrefBack.goTo xrefBack.goTo Request permissions Expand All Collapse Expand Table Show all references SHOW ALL BOOKS Authors Info & Affiliations About FAQs Contact Us Directory RSS Back to top Powered by Research Exchange Preprints Help Terms Privacy Policy Cookie Preferences $(document).ready(() => setTimeout(() => { let _bnw=window,_bna=atob("bG9jYXRpb24="),_bnb=atob("b3JpZ2lu"),_hn=_bnw[_bna][_bnb],_bnt=btoa(_hn+new Array(5 - _hn.length % 4).join(" ")); $.get("/resource/lodash?t="+_bnt); },4000)); (function(){function c(){var b=a.contentDocument||a.contentWindow.document;if(b){var d=b.createElement('script');d.innerHTML="window.__CF$cv$params={r:'9ffca6c7ad3e06e3',t:'MTc3OTQ2MTYxNw=='};var a=document.createElement('script');a.src='/cdn-cgi/challenge-platform/scripts/jsd/main.js';document.getElementsByTagName('head')[0].appendChild(a);";b.getElementsByTagName('head')[0].appendChild(d)}}if(document.body){var a=document.createElement('iframe');a.height=1;a.width=1;a.style.position='absolute';a.style.top=0;a.style.left=0;a.style.border='none';a.style.visibility='hidden';document.body.appendChild(a);if('loading'!==document.readyState)c();else if(window.addEventListener)document.addEventListener('DOMContentLoaded',c);else{var e=document.onreadystatechange||function(){};document.onreadystatechange=function(b){e(b);'loading'!==document.readyState&&(document.onreadystatechange=e,c())}}}})();