Life in the surface ocean produces biogenic material that is constantly exported to depth, fueling deep-sea ecosystems with nutrients and minerals. Exported particulate organic carbon (POC) is the basis of the biological carbon pump (BCP), a key process in regulating atmospheric CO2 levels1. This export results from various pathways, including transport by gravitational settling of particles, vertically migrating organisms, and lateral transport2. Just below the epipelagic layer where light becomes insufficient for photosynthesis (typically around 200 m), the mesopelagic layer receives a rain of organic material punctuated by episodic inputs from the surface ocean3. This mesopelagic realm is a vast transition zone where ecological interactions determine the fate and amount of material that will ultimately reach the deep ocean.

Because of inherent sampling constraints, our knowledge of stocks and processes is scarcer in the deep than in the upper ocean, leaving severe uncertainties in the response of the BCP to global changes4. Settling particles are sources of food for heterotrophic organisms, which consume ~90% of POC exported from surface waters before it reaches 1000 m depth3. Therefore, conducting a census of the mesopelagic biota and quantifying its contribution to biogeochemical cycling in regard to its trophic role is essential to understand the fate of sinking material3,5. Most previous studies on mesopelagic biota focused on morphologically robust metazoan taxa, which are more easily accessible using nets or active acoustics6,7. In contrast, the role of unicellular zooplankton such as Rhizaria, known to be abundant in this layer8,9,10, has been overlooked.

Rhizaria are a 515-Ma ancient eukaryotic lineage11, among which planktonic taxa populate modern oceans from the surface to the abyss8,9,12 and from equatorial to polar waters8,9,10. Their size range spans from a few μm to several mm, with some taxa being able to form colonies of up to one meter9. Planktonic Rhizaria include Phaeodaria, Radiolaria, and Foraminifera8,9. Phaeodaria and Radiolaria (including the orders Acantharia, Orodaria, and Collodaria) mostly biomineralize siliceous skeletons, while many Collodaria are naked and Acantharia build strontium sulfate skeletons8,9. These groups are often constrained to a narrow depth range12, according to their trophic mode and ecological niche.

Many Radiolaria and Foraminifera are mixotrophs, harboring photosynthetic algal endosymbionts which sustain energetic requirements of the host cell13. They contribute to atmospheric CO2 uptake and can thrive in the epipelagic layer of oligotrophic oceans, where organic food resources are scarce9. Phaeodaria, on the other hand, are strictly heterotrophic and mostly found in the mesopelagic layer, where they feed upon sinking particles14,15. Many large planktonic Rhizaria have adapted their lifestyle to food-depleted environments through a low cellular carbon density16. In contrast, siliceous Rhizaria are the most silicified pelagic organisms known to date17,18. As they are abundant down to the bathypelagic ocean, they contribute to silica uptake where no other planktonic organism takes up dissolved silica18. Siliceous skeletons of Phaeodaria can act as ballasting minerals upon organisms’ death and increase the settling velocity of incorporated and attached POC towards the deep ocean. However, these organisms are flux-feeders, i.e., they feed on sinking particles rather than suspended ones14,15. As a result, they can attenuate a substantial amount of sinking flux especially in the upper mesopelagic, where they are most abundant19.

Because traditional sampling techniques damage fragile representatives especially of large Rhizaria, these organisms have often been neglected in biogeochemical studies20. Nevertheless, in situ imaging tools have revealed their substantial contribution to elemental stocks20,21 and fluxes17, as well as their role as gatekeepers of the BCP19,22. Yet of interest for biogeochemical modeling and our understanding of the biological pump, their capacity to contribute to flux attenuation has never been assessed to a larger extent.

In this study, we assess the global organic carbon and silica biomasses of large Rhizaria as well as the role of large Phaeodaria in mediating the fluxes of these elements. While we revise downward the total carbon biomass of Rhizaria, we show that these protists are widely distributed at high latitudes, where their abundance was previously shown to decline20. Furthermore, we highlight the importance of large Phaeodaria in the mesopelagic layer and show that they can attenuate 3.8–9.2% of gravitational carbon export. We find that these silicifying protists co-dominate the silicon cycle, along with diatoms and sponges. In particular, they constitute a unique source of biogenic silica (bSi) production and stock in the mesopelagic layer, where no other silicifiers are found. With expected changes in oceanic conditions and in global carbon export, we discuss the role of these protists in future oceans.

Results and discussion

Collecting Rhizaria images and modeling their distribution

Here, we present a global dataset of large Rhizaria distribution and abundance collected in situ throughout all oceans between 2008 and 2021. The dataset consists of 4252 vertical profiles acquired with the Underwater Vision Profiler 523 (UVP5), of which 1959 extend down to 1000 m depth (Fig. 1 and Supplementary Table 1). The UVP5 recorded 74,157 images in the epipelagic layer (0–200 m) and 93,394 images in the mesopelagic layer (200–1000 m), which have been all manually validated. With Equivalent Spherical Diameters (ESD) ranging from 0.6 to 20 mm (therefore excluding more abundant smaller Rhizaria; Supplementary Table 2 and Supplementary Fig. 1), Rhizaria images cover 18 taxonomic categories12 included in Radiolaria, Foraminifera, Phaeodaria and other Rhizaria (Fig. 2 and Supplementary Table 2). We apply the most recent allometric volume-to-elemental content relationships16,18 to obtain carbon content of all Rhizaria groups, as well as the silica content for Phaeodaria only. Indeed, all Phaeodaria are known as silicifying, while silicified Collodaria cannot be distinguished from naked ones in UVP5 images. Besides their contribution to element stocks, we further investigate their role in biogeochemical processes by estimating Phaeodaria carbon demand22 and bSi production18.

Fig. 1: Spatial and temporal coverage of the 4252 vertical in situ profiles.
figure 1

a Geographic location of Underwater Vision Profiler 5 profiles used in this study. Left and top histograms show the latitudinal and longitudinal distribution of profiles along 3° bins. b Seasonal distribution of sampling points according to latitude. Squares are colored proportionally to the sampling effort. Details about sampling are summarized in Supplementary Table 1. The map was created using the R software version 4.0.3 (ref. 63).

Fig. 2: Overview of sampled Rhizaria specimens.
figure 2

a Equivalent spherical diameter distribution, counts, and characteristics of each taxon considered. Characteristic mixotrophic, flux-feeding (feeding on sinking rather than suspended particles), and silicifying (the capacity to build a silica skeleton) lifestyles are indicated on the right-hand side (dark gray indicates a proven attribute, light gray an evident attribute for at least parts of the species within this taxon, and white its proven absence). b Example images for analyzed Rhizaria taxa numbered according to (a) (all images are on the same scale).

Associating each of the 167,551 Rhizaria images with a set of environmental variables, we quantify these biomasses and processes on a global 1° × 1° grid by using boosted regression trees, with both random and spatial cross-validation (CV) for model assessment. We used spatial CV to account for the fact that spatial data are not independent and to assess the performance of our models on new environmental conditions. For each model and CV method, we provide a R2 value (Supplementary Tables 3 and 4), calculated as the squared Pearson correlation coefficient between the observed biomass and the mean predicted biomass (see “Methods”).

Global Rhizaria biomass from the epipelagic to the mesopelagic

Our models predict the global carbon biomass of large Rhizaria (>0.6 mm) within the upper 1000 m to be 0.012 Pg C (0.11–0.13 Pg C; see “Methods” for uncertainty estimation; Supplementary Table 3) and 0.007 Pg C within the top 500 m, obtained by integrating biomass within the epipelagic layer and the top 300 m of the mesopelagic layer. This is about one to two orders of magnitude lower than prior estimates (0.204 Pg C and 0.061 Pg C in the top 500 m)20,21. In contrast to these studies, our estimates rely on dedicated volume-to-elemental content allometric relationships established from carbon and volume measurements on living specimens16, which showed lower carbon densities of large Rhizaria and the inadequacy of previous conversion factors. Yet, large discrepancies can be observed in model assessment (R2 metrics; Supplementary Table 3) depending on the use of random or spatial CV. By using a spatial CV, we purposely excluded parts of the observed environmental conditions, leading to highly conservative R2 values. Low observed R2 values when using spatial CV show the difficulty for our models to predict on new data, but it should be kept in mind that the final model is trained on the entire dataset and therefore includes a larger range of environmental conditions.

Based on total mesozooplankton biomass estimates21, we now assess the contribution of large Rhizaria to total mesozooplankton biomass in the top 500 m to be 1.7%. Although we revise downward their contribution to biomass, the statement that these low carbon density organisms could contribute globally to 31% of mesozooplankton abundance from UVP datasets20 still holds true.

All taxa considered, large Rhizaria are distributed worldwide in both the epipelagic and the mesopelagic layers (Fig. 3a–c). By extending our spatial coverage poleward compared to other studies, we reveal the prevalence of these organisms at high latitudes, where their biomass was previously shown to be lower10,20,24. However, it should be kept in mind that high-latitude sampling mostly occurred in boreal and austral summers, likely during periods of highest biomass. Most importantly, the worldwide carbon concentration of large Rhizaria is of the same order of magnitude between the epipelagic and the mesopelagic layer in the tropical ocean and at high latitudes (Fig. 3a, b). Integrated biomass values are similar between layers around 20°N and S, but get consistently higher in the mesopelagic near the equator and above 30°N and S (Fig. 3c). The two peaks at 50°N and 60°S follow recent biomass pattern delineations in subpolar regions and around the equator21,24, likely due to the presence of fronts at these high latitudes. Nonetheless, these global patterns hide taxon-driven differences.

Fig. 3: Predicted global organic carbon biomass in planktonic Rhizaria, their impact on carbon flux attenuation, and their carbon demand.
figure 3

a, b Maps of the predicted average 1° × 1° carbon concentration in the epipelagic (0–200 m; a) and mesopelagic layer (200–1000 m; b). c Integrated carbon biomass as a function of latitude for both layers. Regression curves were derived using Generalized Additive Models. Note the logarithmic scaling for carbon biomass. d Total carbon biomass for Rhizaria groups in both layers. Only groups whose model calculated by random cross-validation is >0.05 are shown (see also Supplementary Table 3). e Global and regional contribution of large mesopelagic Phaeodaria to gravitational particulate organic carbon (POC) flux attenuation, based on the ratio of their annual carbon demand to the respective median annual carbon export (i.e., transport out of the euphotic zone) as reported elsewhere32,33,34 and summarized in Supplementary Table 5. Bars show the ratio calculated with the single export value available or the median value when more than one export value was available. In the latter case, dots show export values used to calculate the median. Error bars are represented according to the range of observed export values. No POC export measures were available for the subarctic Pacific and Atlantic. f Global and regional integrated daily carbon demand of large mesopelagic Phaeodaria. Dots show mean values, error bars show regional mean ± standard deviation as presented in Supplementary Table 5. The number of 1° × 1° cells used to derive statistics are displayed below for each region. a, b Maps were created using the R software version 4.0.3 (ref. 63).

Contrasting patterns between mixotrophic and heterotrophic taxa

Given their different trophic modes and lifestyles, we observe distinct patterns between mixotrophic and heterotrophic Rhizaria groups (Fig. 3d and Supplementary Fig. 2). Globally, Collodaria, Acantharia, and, to a lesser extent, Foraminifera, make up most of the Rhizaria biomass in the epipelagic layer of inter-tropical regions (45°N–45°S), particularly within subtropical oligotrophic gyres (Supplementary Fig. 3). This pattern is expected given the mixotrophic nature of these radiolarian orders13,20,25. Through nutrient retention, mixotrophs can enhance primary production13, besides shortcutting energy pathways to higher trophic levels13,20. As these environments are expected to expand as a consequence of global changes26, the importance of these protists in ecosystem functioning is likely to increase in warmer and more oligotrophic oceans. More investigations about their role in CO2 uptake and carbon export is yet needed to provide more detailed predictions about their role in future oceans.

In contrast to mixotrophic taxa, Phaeodaria dominate Rhizaria biomasses in the mesopelagic layer (Fig. 3d and Supplementary Fig. 2). Unlike Radiolaria which prey upon various organisms ranging from bacteria to small Metazoa25, these nonmotile organisms are floating in particle-rich zones, where they intercept aggregates by extending cytoplasmic strands (i.e., pseudopodia8). Often thought to be restricted to deep waters8, our observations reveal important epipelagic biomasses of Phaeodaria in several high-latitude areas, where they could feed on the sinking of large particles. This is in agreement with previous observations in the Southern Ocean27,28,29, in particular the Weddell Sea29, the North Pacific20,30, but also the Sea of Japan31. Globally, their total carbon biomass is tenfold higher in the mesopelagic than in the epipelagic zone (Fig. 3d and Supplementary Table 3), with an overall contribution of 81% to total Rhizaria biomass. The prevalence of Phaeodaria in deep waters and in cold high-latitude areas can be explained by their adaptation to cold-water environments, due to their gigantic size and low cellular carbon density. Despite their important abundance and biovolume20 in many regions and depths, a result of their low carbon density is that their average contribution to total mesozooplankton carbon biomass is low (0.9% in the top 500 m). However, it is also highly variable in different ocean regions: their proportion on the total C biomass ranges from 2.7 to 13.7% between 150 and 1000 m in the North Pacific30 up to 22.3% between 250 and 3000 m in the Sea of Japan31. Since these abundant mesopelagic detritivores are flux-feeders14,15, we further investigated their metabolic requirements and potential impact on carbon fluxes in the mesopelagic layer.

Role of mesopelagic Phaeodaria in carbon flux attenuation

We estimated the individual carbon demand of mesopelagic Phaeodaria based on their carbon content and the temperature at their depth, using a range of turnover time values (the time it takes for one individual to be replaced) and a gross growth efficiency of 40% (see “Methods”). When predicted at global scale, we estimated an annual carbon demand of 0.46 Pg C y−1 by mesopelagic Phaeodaria alone (Supplementary Table 4). Scaled to recent global gravitational POC export estimates, ranging from 5 to 12 Pg C y−1 (see ref. 32 and references therein), they would intercept between 3.8 and 9.2% of the gravitational POC flux exported out of the euphotic zone (Fig. 3e). By modulating their turnover times over the full span of observed values22 (see “Methods”), the global flux attenuation by Phaeodaria could range from 2.4–5.8% to 5.0–12.0%. Considering a gross growth efficiency of 20% (see “Methods”), Phaeodaria-mediated attenuation could double. Attenuation from these large heterotrophic protists has never been taken into account in previous assessments of zooplankton carbon flux attenuation, which was thought to be driven by Metazoa3,6,7. Integrating this number into the global carbon budget is therefore necessary to refine biogeochemical models and to improve predictions of the future ocean state.

Maximum potential Phaeodaria-driven attenuation is found in the Southern Ocean, where it ranges between 11.2 and 23.4% of the gravitational POC export (0.62–1.3 Pg C y−1, refs. 33, 34). In contrast, it approximates 3.8–6.7% in equatorial and upwelling areas32,34 (Fig. 3e), where large Phaeodaria are also abundant. These observations align with previous findings showing more important zooplankton carbon demands in high latitudes and productive areas7. Their daily integrated carbon demand averages 3.9 ± 3.4 mg C m−2 d−1 worldwide (Fig. 3f and Supplementary Table 5), with the highest values in the subarctic Pacific (13.0 ± 6.5 mg C m−2 d−1; Fig. 3f and Supplementary Table 5). In subtropical gyres, where they are least abundant, they could account for 18.9% of the total zooplankton carbon demand (3.4 ± 3.1 mg C m−2 d−1 from 17.9 mg C m−2 d−1 reported for the North Pacific Subtropical Gyre35). In contrast, in the subarctic Pacific where their demand is maximum, their contribution to total zooplankton demand (133.1 mg C m−2 d−1, ref. 7) drops to 9.7%, as metazoan zooplankton may outcompete them for food.

As Phaeodaria are known to consume preferably sinking particles, rather than suspended ones14,15, they can exert a substantial influence on rapidly sinking particles, which are expected to be preferentially transferred to the deep ocean due to their high sinking speed. In the California Current, the Phaeodaria family Aulosphaeridae alone can be responsible for an average 10% of total flux attenuation at the depth of their maximum abundance19 in the upper mesopelagic zone where also food supply is highest. Due to the fact that a substantial portion of Phaeodaria also feed in the lower epipelagic above 200 m12,19 (Supplementary Fig. 4), our estimates are likely conservative as we only estimated their carbon demand for depths between 200 and 1000 m. Ingested material is used for Phaeodaria growth but is ultimately processed into mini-pellets, ejected into the water column, and thought to play an important role in carbon export36. Indeed, mini-pellet abundance can be almost five orders of magnitude higher than that of krill fecal pellets in the Weddell Sea29. This leaves an open question regarding whether mesopelagic Phaeodaria ultimately slow down or accelerate downward fluxes. To answer it, the nature of the exported material—either fecal pellets or aggregates37,38—and the associated differences in sinking velocities and carbon contents, as well as the resulting effects on global export must be further explored. As Phaeodaria aggregates can be observed with in situ imaging38, this tool can be used to estimate Phaeodaria-mediated fluxes. Their silicified skeleton, often included in aggregates37, is likely increasing their sinking velocities; thus, we further investigated the role of Phaeodaria in the silicon cycle and potential impacts for the carbon cycle.

Biogenic silica stocks and production mediated by Phaeodaria

Globally, our models predict that Phaeodaria account for 4.25 (3.8–4.7) Tg of bSi standing stocks in the upper 1000 m of the oceans, among which 0.34 (0.28–0.39) Tg are in the epipelagic and 3.91 (3.51–4.31) Tg in the mesopelagic layer (Fig. 4 and Supplementary Table 4). The mean integrated value for the epipelagic layer is 1.1 ± 3.3 mg Si m−2 (range 0–204.4 mg Si m−2; Fig. 4c), which is in the range of values measured on small Rhizaria, not within the size spectrum of the UVP518. The whole population of Rhizaria would thereby account for 7–18% of the total integrated bSi pool in the epipelagic layer39 while most of the remaining stock is being attributed to diatoms40. In the mesopelagic layer, the mean integrated value is 11.6 ± 12.1 mg Si m−2 (range 0–378.3 mg Si m−2; Fig. 4c) and would account for ~0–15% of the total bSi integrated pool, based on average estimates from other studies41,42.

Fig. 4: Predicted biogenic silica (bSi) biomass of planktonic Rhizaria.
figure 4

a, b Maps of the predicted average 1° × 1° biogenic silica concentration in the epipelagic (0–200 m; a) and mesopelagic layer (200–1000 m; b). c Integrated bSi biomass as a function of latitude for both layers. Regression curves were derived using Generalized Additive Models. Note the logarithmic scaling for bSi biomass. d Total bSi biomass for Phaeodaria families in both layers. Only groups whose model calculated by random cross-validation is >0.05 are shown. a, b Maps were created using the R software version 4.0.3 (ref. 63).

We further quantify the role of Phaeodaria in the silicon cycle by estimating total bSi production (or dissolved silica uptake). Average global annual bSi production by Phaeodaria is 0.70 Tg Si y−1 (range 0.22–2.33) in the epipelagic and 3.96 Tg Si y−1 (range 1.25–13.1) in the mesopelagic zone (Supplementary Table 4) while daily integrated production rates averages 0.005 ± 0.016 mg Si m−2 d−1 in the epipelagic and 0.030 ± 0.031 in the mesopelagic layer. In the Southern Ocean, the mean production rate within the top 200 m (0.006 ± 0.004 mg Si m−2 d−1) is about three orders of magnitude lower compared to previous estimates that included small Rhizaria only28. Although biomasses and production rates in the epipelagic layer are seemingly low compared to diatoms28,40, the overwhelming majority of Phaeodaria is located in the mesopelagic zone (Fig. 4). Our results complete previous bSi production rates estimated in the global ocean18 from plankton net samples. While this first estimate (56–1600 Tg Si y−1) was based mainly on small organisms, our 4.66 Tg Si y−1 estimate focuses on the largest specimens (>600 µm) and is therefore difficult to compare. With the most recent estimates40 only focusing on diatom- and sponge-mediated bulk bSi production in the sunlit epipelagic and benthic environments, less is known in the deep ocean. Therefore, the role of Phaeodaria in mediating silica through production, export, and dissolution is unique, leaving the fate of deeply produced bSi still not sufficiently assessed.

Due to the porous nature of Phaeodaria skeletons8, the skeleton of epipelagic populations dissolves in the upper part of the ocean43. As a result, dissolved silica is spread throughout the water column, with the potential to resurface through physical mixing, and export to the deep ocean is minimal. In contrast, as bSi production by mesopelagic Phaeodaria occurs in deeper waters, we can expect that their carcasses also reach greater depths before undergoing total dissolution, exporting and spreading bSi more efficiently to the deep ocean. Furthermore, as Phaeodaria accumulate siliceous material within their phaeodium, with quantities often reaching the same magnitude as their skeletal structure, it likely enhances the sinking velocity of both the cell and the organic and inorganic materials it carries44. Nevertheless, dissolution of bSi before it reaches the seafloor is likely, preventing preservation and fossilization of fragile large Phaeodaria skeletons in deep-sea sediments. The contribution of deep-living Phaeodaria to fluxes, along with the susceptibility of their skeleton to dissolution, has been hypothesized to be an important factor of silica recycling in the North Pacific45.

Our results suggest that mesopelagic Phaeodaria are important producers and recyclers of bSi, besides organic matter, particularly in the deep ocean. Diatoms have been considered to be the main driver of the silicon cycle globally since the Mesozoic era, when they took over the control from Rhizaria and marine sponges46. Our results therefore suggest that Rhizaria, diatoms and sponges still co-dominate the silicon cycle globally, and also provide crucial elements for the future changing oceans. Indeed, diatom populations are thought to decline in the future ocean due to increased water column stratification and thus a decreased nutrient supply into the euphotic zone47, or a pH-driven decrease in silica dissolution of sinking minerals causing a reduced recycling of silicic acid into the surface ocean48. Given the present results, we further explore the consequences of such changes also for Rhizaria populations and their impact on biogeochemical cycles.

Implications for future ocean biogeochemistry and limitations

Rhizaria have populated the global ocean since 515 Ma and survived all major extinctions11. Because of climate change, significant impacts on all oceanic regions are expected, such as seawater warming from the surface to the deep or increased stratification47. In future more oligotrophic oceans26, we can expect mixotrophic Rhizaria populations, thriving in oligotrophic gyres (Supplementary Fig. 3), to remain stable or to expand their habitat range globally. These organisms create microenvironments of enhanced primary production to meet their carbon needs49. As such, they will likely not be impacted by changes in primary productivity in surrounding waters and may be favored by elevated temperatures (Supplementary Fig. 5). In addition, they may also profit from a shift towards smaller phytoplanktonic prey, as mixotrophic Radiolaria are known to feed on small organisms25. Consequently, their importance on food webs may increase in the future.

In contrast, the fate of Phaeodaria is less certain. These protists are influenced by surface chlorophyll a concentrations (Supplementary Fig. 5) and their distribution areas may decrease because of expanding oligotrophic oceans26. However, they are generalists, feeding on diverse types of particles14,15. Therefore, they may adapt to changes in the upper phytoplankton community composition. Nevertheless, the response of export flux to global changes is highly uncertain as fluxes may increase at high latitudes, while decreasing in equatorial and upwelling areas4. Consequently, we can expect Antarctic or subarctic Phaeodaria populations to remain stable, or even become more important. In contrast, upwelling and equatorial populations could decline following such changes. These organisms primarily inhabit the deep ocean, where silica concentrations exceed their requirements and where they are its exclusive users40. We can therefore expect them to remain unaffected by changes in the silicon cycle and likely increase their control, leading to more bSi dissolution and recycling at depth. Altogether, their ubiquity as well as the variety of their trophic modes make Rhizaria persistent key organisms also in future oceans.

Our findings advocate for the inclusion of planktonic Rhizaria as a separate compartment in biogeochemical models due to their role in silica and carbon cycling. Yet, in such models, the zooplankton compartment is still inadequately represented. While zooplankton encompass both mixotrophs and heterotrophs, their diversity is often limited to size classes, such as microzooplankton and mesozooplankton50. This is partly due to the lack of information about its various components. By refining our knowledge regarding the contrasting distribution and role of mixotrophic and heterotrophic Rhizaria, we provide synoptic information to better represent them in biogeochemical models. However, their impact on element turnover is likely to change, calling repeated assessments of their role in biogeochemical cycles in the future.

Our study provides a comprehensive assessment of the role of giant planktonic Rhizaria in the biogeochemical cycles of carbon and silicon, offering a perspective on their significance for ocean biogeochemistry. They are the sole drivers of biogenic silica production in the mesopelagic layer, and likely throughout the deep ocean, co-dominating the silicon cycle with diatoms and sponges worldwide. Despite relatively low carbon biomasses compared to metazoan zooplankton, they could consume up to 9% of the global gravitational carbon flux worldwide, diminishing the transfer efficiency through the mesopelagic layer. However, their ability to export material through both ballasted dead bodies and mini-pellets may counterbalance this effect. The fate of these substantial biomasses in vertical or silica dissolution fluxes therefore remain to be investigated. However, the present study did not account for smaller, yet more abundant38, Rhizaria taxa. These may process matter more quickly due to shorter turnover times and could increase recycling of organic matter and bSi throughout the water column. While we expect a limited impact of climate change on overall Rhizaria populations, uncertainties remain regarding the evolution of biogeochemical processes in the mesopelagic zone and the role of Rhizaria in them. Future research should therefore focus on including the full size spectrum of Rhizaria and their multiple roles within mesopelagic food webs and pathways through which they mediate carbon and silica.


Global underwater vision profiler 5 dataset

We used a global UVP5 dataset from 64 oceanographic cruises covering a 13-year period (2008–2021; Supplementary Table 1), which took place across all oceans and across a large range of oceanic structures (Fig. 1a). Among all profiles collected, 4252 covered the first 200 m and 1959 the first 1000 m (Supplementary Table 1). Sampling occurred throughout the year, except at high latitudes where access is limited to boreal or austral summers (Fig. 1b).

The UVP5 images a water volume of ~1 L every 5–20 cm of the water column during the descent part of a vertical profile. The onboard computer measures all particles larger than ~0.1 mm, but stores vignettes for particles >0.6 mm only23. Upon recovery, vertical profiles are processed to extract images, which are associated with a set of metadata and morphological measurements.

Rhizaria images were classified by supervised machine learning algorithms and validated by taxonomy experts on the EcoTaxa web application51. In total, 167,551 Rhizaria images were validated, and classified into 18 subgroups belonging to Radiolaria (i.e., Acantharia, Collodaria, and Orodaria), Foraminifera, Phaeodaria and unidentified Rhizaria, following the latest classification for in situ Rhizaria images12 (Fig. 2 and Supplementary Table 2). Phaeodaria were represented by the two abundant families Aulacanthidae and Aulosphaeridae, and by the less abundant families Cannosphaeridae, Castanellidae, Coelodendridae, Tuscaroridae, plus an additional category “Phaeodaria_unknown”. At the lower range of the detection threshold of the UVP5, Acantharia, distinguished by their symmetric spines surrounded by a black center20, were divided into Acantharia and “Acantharia_like”. Collodaria were further classified into colonial specimens, solitary black, and solitary globule20. The radiolarian order Orodaria was split between the genus Cytocladus and other Orodaria. Foraminifera and other Rhizaria were all classified as such. Our Rhizaria specimens covered a size spectrum ranging from 0.6 mm to 20 mm, the smallest specimens belonging to Acantharia and the largest to colonial Collodaria (Fig. 2).

Biomass, carbon demand, and biogenic silica production estimates

For each individual, volume was determined by first computing the Equivalent Spherical Diameter (ESD, in µm) from the surface area of the organism in square pixels converted in µm2 (area), including all pixels above a given threshold, extracted by ZooProcess using Eq. (1):

$${ESD}=\sqrt{\frac{4\times {area}}{\pi }}$$

Then, the volume V (in µm3 cell−1) of the associated sphere was calculated following Eq. (2):

$$V=\frac{4}{3}\pi {\times {ESD}}^{3}$$

Volumes were also calculated by fitting a prolate ellipse to the object using the major and minor axes length, to ensure that it would not lead to significant differences. Although no significant differences were observed (Supplementary Fig. 6), the sphere method was chosen to estimate volumes as the ellipse method may inflate volumes due to the inclusion of spines in the ellipse.

A volume-to-carbon allometric relationship combining all available carbon content measurements16,52 was applied to individual volumes V (in µm3 cell−1) to obtain individual carbon contents QC (in µg C cell−1) as described by Eq. (3):

$${Q}_{C}={10}^{0.958}\times {V}^{0.455}$$

For siliceous Phaeodaria, biogenic Si contents QbSi (in µg Si cell−1) were computed from volumes V (in mm3 cell−1) using the volume-to-biogenic Si allometric relationship reported elsewhere18 (Eq. 4):

$${Q}_{{bSi}}={10}^{-4.05}\times {V}^{0.52}$$

Individual carbon demand (CD, in µg C cell−1 d−1) for mesopelagic flux-feeders Phaeodaria was calculated from individual carbon content QC following Eq. (5)22:

$${CD}=\frac{{Q}_{C}}{{GGE}\times {\tau }_{T}}$$

with GGE the gross growth efficiency (unitless) and τT the turnover time (in d). Assuming that turnover times are temperature-dependent with a Q10 of 2 (i.e., increasing the temperature by 10 °C divides the turnover time by 2), we used a median reference turnover time τ of 10.9 d at a temperature of 10 °C, calculated over a range of turnover time values estimated between 0 and 150 m22. This value was used for all Phaeodaria taxa considered. We also considered the 1st (τ = 8.4 d) and 3rd (τ = 17.4 d) quartiles of the same range of turnover times. Each mesopelagic Phaeodaria specimen was assigned a temperature value coming from the World Ocean Atlas53 according to its depth, location and month of sampling. The reference turnover time was adjusted to the local temperature using Eq. (6)20:

$${\tau }_{T}=\tau \times {2}^{\left(\frac{{T}_{{ref}}-{T}_{{obs}}}{10}\right)}$$

with Tref = 10 °C and Tobs the observed temperature. GGE, the ratio of prey carbon to predator carbon, ranges between 0.1 and 0.4 and averages 0.2–0.3 for a broad range of protist and metazoan zooplankton54. A high value indicates an optimized carbon uptake from food, resulting in minimized energy expenditures. As we expect Phaeodaria, living in the food-depleted mesopelagic environment, to reduce their energy expenditures, we applied a GGE of 0.4 as reported elsewhere22 for all considered taxa.

To propose a range of biogenic Si production (ρSi, in µg Si cell−1 d−1), we considered the minimum, maximum, and median from literature values (0.17, 0.54, and 1.78 nmol Si cell−1 d−1)18 and applied a Q10 of 2 as described above for all specimens by Eq. (7):

$${\rho }_{{Si,T}}={\rho }_{{Si}}{\left[{2}^{(\frac{{T}_{{ref}}-{T}_{{obs}}}{10})}\right]}^{-1}$$

with Tref and Tobs as described above. Daily values are then multiplied by 365 to obtain yearly estimates.

Environmental data

Temperature, salinity, oxygen and nutrient concentrations (i.e., silicate, phosphate and nitrate) data, from 2008 to 2019, were extracted from the World Ocean Atlas database53,55,56,57. These data were chosen due to their reliance on climatologies calculated on the basis of actual observations, besides their standard use in marine species distribution models (e.g., refs. 58, 59) which enables direct comparability with similar approaches.

They were delineated on a 1° × 1° horizontal grid over a 0–800 m depth range (as silicate, phosphate, and nitrate data were not available deeper) with a monthly temporal resolution covering the years 2008–2019. They were averaged throughout both layer’s temporal coverage and depth range. Monthly averaged surface chlorophyll a data, extracted from the Copernicus database, and bathymetric data, extracted from the NOAA database60, were used for the corresponding time period. These datasets were standardized to a grid resolution of 1° × 1°. UVP5 data were spatially matched to this environmental data on the global 1° × 1° grid. Due to the fact that the World Ocean Atlas does not exactly follow the coastline, profiles that could not be matched to the environmental grid were included into the nearest cell, if any. Finally, calculated individual biomass, C demand, and bSi production were averaged for each layer and each cell21.

Predicting global distributions

To model the relationship between environmental variables and Rhizaria distributions, and to ultimately predict global Rhizaria biomass distributions, we used boosted regression trees (BRTs), following a methodology developed recently21. Briefly, BRTs function as classical regression trees linking a response (i.e., biomass, carbon demand and biogenic Si production) to predictors (i.e., environmental variables) by performing recursive binary splits61. Boosting allows the combination of successive short regression trees, which are adjusted to improve performance on observations poorly modeled by existing trees62. They do not produce a single relationship, but instead combine relatively simple successive tree models and are thus well adapted to fit complex, nonlinear relationships between sparse species datasets and their environment, yet being robust against overfitting62. BRTs were implemented for each layer and each taxa or group of taxa (i.e., Phaeodaria, each Radiolaria order, Foraminifera, Rhizaria_other, and all Rhizaria). We implemented BRTs using the R software version 4.0.363 and the xgboost package version

To account for spatial correlation between data points, BRTs’ performance was tested using random and spatial cross-validation procedures, the latter to improve independence between the training and test sets65. For the spatial cross-validation, data were split into five spatial folds according to geographical distances using the R package blockCV66. Each set consisting of 4 spatial folds was used to train the model, while the remaining fold was used for testing, repeated 20 times. To evaluate the models, we calculated for each test fold (spatial and random) the mean predicted biomass for all repetitions. Then, we calculated the two-sided Pearson correlation coefficient between the observed biomass and the mean predicted biomass. We calibrated BRTs by trying various combinations of hyperparameters, including the learning rate per tree, the maximum depth of a tree, the minimum number of elements per leaf, and the number of trees, to minimize predictive deviation from the test set (error reduction) and avoid overfitting (complexity minimization)21,67. For all models, we chose a learning rate of 0.08, a maximum depth of tree of 2, a minimum number of elements per leaf of 1, and a maximal number of trees of 500. For each layer, we applied models to environmental predictors for all cells of the 1° × 1° world grid, repeated 20 times to obtain maps of the mean and standard deviation of predicted biomass concentrations. We represented the maps of coefficients of variation (standard deviation divided by mean) for the model predicting the carbon biomass of all Rhizaria and the model predicting the bSi biomass of Phaeodaria (Supplementary Fig. 7).

Univariate partial dependence plots were represented to show the effect of each variable on biomass prediction (Supplementary Fig. 5). They were computed by averaging the marginal effect of the variable across model resamples (predicted changes in fitted biomass values for one unit change in the variable). Predicted values were integrated for each layer (in µg m−2) by multiplying the mean concentration (in µg m−3) for this layer by its thickness (in m). Biomass in the top 500 m was estimated by integrating the mean concentration in the epipelagic and the mesopelagic over 300 m. To derive global values, integrated values were multiplied by the area of each 1° × 1° grid cell and then summed over the world as reported elsewhere21. To provide an uncertainty range, we derived global values using the mean prediction ± its standard deviation for each cell.

Regional values were obtained by partitioning world predictions using Longhurst’s provinces68. To estimate the percentage of flux attenuation, we used carbon export (i.e., out of the euphotic zone layer) values from the literature32,33,34,69 and computed the ratios of Phaeodaria carbon demand to carbon export globally and for each oceanic region.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.