Key Expoitable Results (KERs)

This collection of global maps provides contemporary and future projections of marine planktonic species (n = 843) based on the Habitat Suitability Index retrieved from an ensemble across Species Distribution Models accounting for several sources of uncertainty (algorithm complexity and climate models). Each species file can be opened and contains 5 dimensions (rcp, statistic, timesteps, latitude and longitude). The data covers the time period from 2012 until 2100 in monthly timesteps for three different RCPs (RCP2.6, RCP4.5, RCP8.5). The maps cover the time period from 2012 until 2100 in monthly timesteps for three different RCPs (RCP2.6, RCP4.5, RCP8.5). The statistics (mean, standard deviation, minimum and maximum) are calculated across three different Earth System Models (CNRM-CM5, GFDL-ESM2M, IPSL-CM5A-LR) and three different algorithms (General Linear Model, General Additive Model and an Artificial Neural Network). The resolution of the data is 1° by 1° monthly resolution.

This collection of global maps provides the projected presence probabilities of 374 Single-cell and Metagenome Assembled Genomes (SMAGs) from Delmont et al. (2022). They are projected using 10 year average ensemble climatologies (CMIP5) in the present day (2006-15) and in the future (2090-99, RCP8.5).

This collection of global maps provides global projections of habitat suitability index (HSI) estimated for each of 53 different picozoa operational taxonomic units (pOTUs), as described in Huber et al. 2024. For each pOTU-level map, mean, max, min, stdev, and coefficient of variation of HSI values computed across several algorithms are provided. We developed the single pOTU HSI maps from 2,394 eDNA samples relative to the pico-size fraction (i.e. 0.2 to 5 μm), which were retrieved from the EukBank database. Picozoa samples were classified based on their depth into epipelagic (0-200 m), mesopelagic (200-1000 m) and bathypelagic (>1000 m) samples. To estimate the HSI, we used species distribution models based on a presence/absence design; among all retrieved samples of pico-eukaryotes, georeferenced samples where a specific pOTU was detected were treated as presences, while sampling locations where that pOTU was not detected were treated as (pseudo-) absences. For each pOTU, we contrasted average multi-year depth-specific environmental conditions at presence vs absence points using an ensemble of modelling algorithms, including Generalized Linear Models (GLM), Artificial Neural Network (ANN), Boosted Regression Trees (BRT) and Random Forest (RF). Models were calibrated using multi-year average conditions across epipelagic, mesopelagic and bathypelagic layers. HSI maps were projected over epipelagic conditions only, due to lack of sufficient data coverage for deeper layers. Environmental predictors were retrieved from the World Ocean Atlas 2018.

This collection of global maps provides global marine prokaryotic diversity using a standardized ensemble pipeline that integrates metagenomic profiles with environmental predictors. Metagenomic samples from the Ocean Microbiomics Database were processed into domain- and class-level taxonomic profiles, and diversity indices (richness, Shannon, Chao1) were estimated from rarified mOTU counts. Multiple algorithms—Generalized Linear and Additive Models, Random Forests, Boosted Regression Trees, Support Vector Machines, and shallow neural networks—were trained and tuned using five-fold cross-validation. Models were evaluated using root mean square error and R², with only well-performing models (R² ≥ 0.25) retained. Ensemble predictions were derived as the mean across all successful models, while associated uncertainty was quantified as the standard deviation among them. The resulting dataset contains annual, global projections of domain- and class-level prokaryotic diversity indices (richness, Shannon, Chao1). Each netCDF file provides both the ensemble average and standard deviation, offering spatially explicit estimates alongside model-based uncertainty.

This collection of global maps provides global predictions of surface monthly microplastics concentrations (in #items.km-2, log10 transformed) modelled through five distribution models (GLM, GAM, ANN, RF, GBM). The models were trained on an observational dataset that aggregated measurements from multiple microplastic sampling campaigns compiled by van Sebille et al. (2015) and completed by Schmiz (2021). The original dataset of Schmiz (2021) synthetised no less than 13 159 microplastic concentrations observations spanning from 1979 to 2019. These microplastic concentrations measurements were collected through diverse sampling methods (~83% come from plankton nets) and covered plastics particles ranging from 32µm to 20cm. Giacomo Poli (MSc student, ETH Zürich) analysed this dataset and adapted the modelling pipeline of Knecht et al., (2023) to model global surface microplastics concentrations based on five types of statistical models and an exhaustive suite of environmental and anthropogenic predictors.

This collection of global maps is based on protein data from 1,379 metagenomes (from AtlantECO-BASEv2), we define Operational Protein Units (OPUs) through a sensitive heuristic clustering strategy that captures both known and unknown amino acid sequences, including those associated with "functional dark matter" often missed by traditional analyses. Based on abundance and occurrence, 20 representative OPUs were selected for the modelling process. These OPUs were assigned to different functional pathways. For each of 20 OPUs we ran species distribution models to predict their global distribution. We modeled using four models- Generalized Linear Models (GLM), Artificial Neural Network (ANN), Boosted Regression Trees (BRT), Random Forest (RF), resulting in 80 maps. With this, with the ensemble approach, we used the mean of the suitability among the four models, resulting in the potential distribution maps. The OPUs were modeled to sunlit ocean regions ( 200 meters). The models showed different OPUs distribution patterns: widely, polar, no polar, tropical, temperate and sub polar distributions. These resulting maps are stored in the folder “SDM OPUs”. We used species distribution models to predict the global distribution of 20 OPUs well represented across all the ocean basins.

This collection of global maps provides monthly global biogeography of 14 diazotroph species including cyanobacterial and non-cyanobacterial diazotrophs. Each netCDF file contains the ensemble average, standard deviation, maximum and minimum across all ensemble members. We used total, group-specific and cruise-specific target group approaches.

This collection of global maps is based on Association Distribution Modeling (ADM), which was conceived as a parallel to well-established Species Distribution Models, but dedicated to the modeling and prediction of pairwise ecological associations, statistical proxies for potential biotic interactions, with the aim of further modelling community biogeography with better reliability. Here, only co-occurrence projections are reported, with 3 possible states: 1 for co-presence of both species, 0 for co-absence, and -1 for exclusion (i.e., one species is present but not the other). On the one hand, 57,082 ADMs were projected in the contemporary era using the World Ocean Atlas 18 dataset (Boyer et al. 2018). On the other hand, 17,517 ADMs were projected throughout the century using IPSL (Institut Pierre-Simon Laplace) climate model under a Representation Concentration Pathway 4.5 climate scenario (Boucher et al. 2020).

This collection of global maps provides the projected probabilities of presence of the 27 “climato-genomic provinces” described in Fremont et al. (2022) and originally defined in Richter et al (2022). They are projected using 10 year average ensemble climatologies (CMIP5) in the present day (2006-15) and in the future (2090-99, RCP8.5). Provinces encompass 6 size fractions of plankton that were sampled during the Tara Oceans expeditions (2009-2013): 0-0.2 µm enriched in viruses, 0.22-3 µm enriched in bacteria, 0.8-5 µm enriched in protists, 5-20 µm, 20-180 µm and 180-2000 µm enriched in small metazoans (mainly copepods).

This collection of global maps provides two NetCDF files that record the ensemble predictions of global monthly (12 time levels) biomass concentrations of calcifying Pteropoda (in mg C m-3) and Foraminifera (in µg C m-3). The monthly biomass concentration fields were obtained by training a suite of five distribution models tailored for predicting plankton biomass data obtained from traditional sampling techniques (see Knecht et al., 2023 for details). The observational data and the environmental layers used to train the biomass models were integrated over the epipelagic layer (0-200m) of the global ocean. Therefore, the predicted values are representative of the conditions prevailing in the epipelagic layer. The two NetCDF files thus record the: minimum (Min), maximum (Max), mean, median and the standard deviation (Stdev) of the monthly predictions issued from these five models.

This collection of global maps provides a global Bray-Curtis dissimilarity index (accounting for projected changes in the different size fraction) comparing present day and end of century projected provinces is included. This index is also projected on EEZ and main fisheries (from Watson, 2017).

This collection of global maps provides a global ensemble modeling framework allying cell abundance dataset derived from empirical quantification of phytoplankton groups in metagenomic samples with global environmental variables. Cell abundances of phytoplankton groups were estimated from the quantification of the psbO-gene in metagenomics samples from the Tara Oceans and Tara Pacific expeditions. Phytoplankton communities were resolved into 16 taxonomic groups. Separate models were trained for each of 17 responses (16 groups + total phytoplankton), using only samples with complete predictors and targets. Hyperparameters (tree number, minimum leaf size) were tuned by Bayesian optimization. Optimal configurations were determined via 5-fold cross-validation, balancing predictive accuracy and data use. Performance was quantified using RMSE and R². Final models were retrained with all valid data and applied to the global prediction grid. The ensemble models relied on a suite of key environmental predictors, sea surface temperature (SST), sea surface salinity (SSS), chlorophyll-a, iron, nitrate, phosphate, and silicate. Model predictions were made over monthly global environmental grids, with phytoplankton maps generated at 1° × 1° resolution. We computed Mahalanobis distances between environmental conditions at each grid point and the training dataset to identify regions outside the empirical environmental envelope and quantify prediction confidence and extrapolation risk. In parallel, we conducted sensitivity analyses by varying subsets of the data, quantifying how training data composition influences global predictions.

This collection of global maps provides global picozoa richness for the epipelagic layer, obtained as the sum of mean pOTU-level HSI values across all 53 pOTUs, and the coefficient of variation over all pOTUs-level HSI maps.

This collection of global maps provides global gridded estimates of the functional diversity (FD) of marine copepod assemblages. FD indices were estimated by combining species distribution models (SDMs) with species-level functional traits (i.e., body size, trophic group, feeding mode, myelination and spawning mode) for > 300 copepod species. The dataset includes two complementary products: a contemporary baseline representing average copepod FD patterns for the present-day ocean and a future projection representing expected end-of-century (2081–2100) average changes in FD under a high-emission climate scenario (RCP8.5). Both products are provided as monthly climatologies at 1° × 1° spatial resolution for the surface ocean (0–10 m). Multiple FD indices were computed to cover multiple aspects of functional diversity including: functional richness, evenness, divergence, dispersion, and trait dissimilarity. These data allow researchers to investigate how copepod trait composition is structured across the global ocean today, and how it may be reshuffled in response to anthropogenic climate change. This dataset offers a valuable resource to assess potential impacts of changing zooplankton communities on marine productivity, carbon cycling, and climate feedbacks.

This collection of global maps is based on metagenomic data collected during multiple oceanographic expeditions, including Tara Oceans, Malaspina, GO-SHIP, bioGEOTRACES, and OSD2014. Sample metadata, including geographic coordinates and sampling depth, were curated and merged with the environmental raster stack using bilinear extraction (terra::extract). Only samples with complete environmental metadata and valid coordinates were retained, and with samples from the 0.22-3 µm size fraction. The metagenomic data is from AtlantECO-BASEv2, a database developed within the AtlantECO project. All the metagenomic data were analyzed using the version 5.0 of MGnify’s pipeline. Taxonomic profiles were generated by running mOTUs on MGnify. For each sample (n = 1,210), tables containing mOTUs taxonomic assignments were downloaded and compiled into a single table. Diversity was then calculated from this dataset using the Shannon and Simpson indices in R (vegan package). We conducted our analysis separately, using either data from all expeditions pooled together or using only GO-SHIP data, as this was the only dataset covering the latitudinal gradient. We focused on the sunlit zone due to its ecological relevance for primary production and microbial functional diversity. To predict Simpson and Shannon diversity across global marine expeditions, we applied a species distribution modeling (SDM) framework using four algorithms within R, using the package “h2o” (v3.42.0.2). All models were trained to predict Simpson or Shannon functional/taxonomic diversity based on a selected set of non-collinear environmental predictors: mixed layer depth (MLD), chlorophyll (Chl), salinity (Sal), silicate (Si), temperature (T), and the excess of nitrates over phosphates (N*) according to Redfield ratio [NO3−] − 16[PO43−]. Models were trained on the final dataset using 5-fold cross-validation, for four algorithms (Artificial Neural Network, Random Forest, Generalized Linear Model, Gradient Boosting Machine). Predictions were projected using each trained model. Extracted predictions at sampling locations were compared against observed Simpson/Shannon diversity values using Pearson correlation, RMSE, and R². An ensemble prediction was calculated as the pixelwise mean across the four models.