The primary dataset is based on 34 core-top samples (Table 1, Fig. 1), collected from the North Atlantic Ocean and Mediterranean Sea, with a secondary dataset based on pre-industrial sample material spanning ∼ 0–1500 common era (CE) from ENAM9606 for the North Atlantic (Richter et al., 2009). The core-top samples encompass a substantial natural pCO2 (aq) gradient because the rate of cooling of the North Atlantic Current exceeds that of CO2 uptake, whereas pCO2 in the Mediterranean is close to or slightly above equilibrium with the atmosphere. Sample selection is further based on the dinocyst occurrence maps of Zonneveld et al. (2013), including only samples with an expected relative abundance of at least 10 %–20 % of the target species. Similarly, the coverage of environmental parameters such as sea surface temperature (SST), pCO2 and difference in environmental settings was maximized during sample selection. Existing ocean databases are used for obtaining the relevant environmental parameters (Table 1).

Using standard palynological techniques (see, e.g., Brinkhuis et al., 2003), ca. 5–10 g freeze-dried sediment of the upper 1–2 cm of core material was processed for each sample. This involved dissolving carbonates and silicate components using strong acids (HCl 30 % and HF 38 %–40 %). After acid steps, residues were pH-neutralized and sieved using an ultrasonic bath and 250 and 10 µm nylon mesh sieves to remove large and small particles, respectively. Subsequently, samples were transferred to glass test tubes with demineralized water and centrifuged at 3200 rpm for 10 min to obtain an optimum concentrate of the sample material. Prior to dinocyst selection, samples were stored in 4 mL glass vials in demineralized water.

A micro-manipulator consisting of a Leica inverted microscope and a Narishige IM-9B micro-injector connected to a strung-out pipette was used to manually select individual dinocysts from a water droplet on a glass petri dish. Dinocyst selection followed a strict protocol, in which cyst morphology (primarily cyst shape and process length) was kept constant and the contribution of other organic particles was minimized. Specimens with darker coloration or amorphous organic matter adhering to the cyst or processes were avoided. In the case of O. centrocarpum, the morphological selection primarily involved selecting specimens of equal size and process length to avoid cysts that may be derived from different environments (e.g., Mertens et al., 2009). For Spiniferites, we were able to distinguish and separate three distinct morphological species in sufficient numbers: S. ramosus, S. elongatus and S. mirabilis. For all dinocyst species, the selected diameter excluding processes was in the order of ∼ 30–40 µm, except for S. mirabilis (∼ 60 µm), although constraining the exact size of each individual specimen was not feasible within the current analytical procedures. Stable carbon isotope analyses for individual samples are based on replicating the analyses of single of dinocysts, with ∼ 30 individual measurements being conducted to obtain a reasonably precise (∼ 0.3 ‰–0.4 ‰) sample average (van Roij et al., 2017). Given the size of the dinocysts used here (∼ 30–40 µm cyst diameter), three to seven specimens were required for each measurement, and hence ∼ 150 cysts were required to obtain sample averages (Table 1).

Dinocysts were placed on a 6 mm ∅ nickel sample tray, after which an identical second tray is added on top and compressed to fixate the dinocysts. Before placement in the ablation chamber, approximately ∼ 1 mm2 of International Atomic Energy Agency CH-7 (IAEA-CH7) polyethylene standard (PE; certified δ13C value -32.15 ‰ ± 0.05 ‰; 1σ) was added to the sample tray. Stable carbon isotope analyses of the dinocysts followed the procedures described in previous work (van Roij et al., 2017; Sluijs et al., 2018), utilizing the recently developed ablation–nano-combustion–gas-chromatography–isotope ratio mass spectrometry (LA/nC/GC-IRMS) method. Fragments resulting from deep-ultraviolet LA were carried using a continuous helium flow in 0.32 mm capillaries and oxidized in a combustion oven at 940 ∘C. The resultant CO2 was transported to a GC combustion interface, dried in a Nafion tube using a He counterflow and subsequently into a Thermo Fisher DeltaV Advantage IRMS for isotope analysis. Each analytical run included five standards with signal intensity above 4 Vs (ca. 40 ng C; δ13C precision better than 0.5 ‰) to allow calibrating to the Vienna Peedee Belemnite (VPDB) scale. Direct visual monitoring of the ablation process was used as initial quality assessment of each individual measurement.

To calculate the fractionation factor εp of the dinocysts relative to dissolved inorganic carbon (DIC) from which the dinocyst was produced, we take the δ13CDIC from the modeled grid of Tagliabue and Bopp (2008). As many dinoflagellate species, including those that produce O. centrocarpum and S. ramosus cysts, are able to utilize both HCO3-, which makes up the majority of DIC, and CO2 for carbon fixation (Hoins et al., 2016a), we also compare the δ13CDINO data to δ13CCO2 and to overall seawater carbon partitioning.

εp-DIC is calculated as δ13CDIC–δ13CDINO, and εp-CO2 is calculated as δ13CCO2–δ13CDINO. For the latter the δ13C of dissolved CO2 is calculated from δ13CDIC using the temperature-dependent fractionation between DIC and CO2(aq) (Mook et al., 1974). To evaluate the dominant contributions to 13C fractionation in dinocysts, we compare the εp-DIC and εp-CO2 values to measured and interpolated physicochemical parameters. We test a suite of parameters, [NO3-], [PO43-], [Si], alkalinity, pCO2, SST and sea surface salinity (SSS), which are extracted using Ocean Data View (https://odv.awi.de/, last access: 20 July 2020) from existing (gridded) datasets (Gouretski and Koltermann, 2004; Takahashi et al., 2014, 2016) (Supplement). Where possible, data are averaged over a grid of 4∘ longitude and latitude around the sample position. This is both to reduce errors introduced by data scarcity and to account for potential lateral transport of dinocysts during sinking (Nooteboom et al., 2019). Carbonate chemistry is calculated using the R package seacarb (Gattuso et al., 2021), with alkalinity and pCO2 as input variables to calculate the other relevant parameters of the carbonate system: the relative contributions of CO2 (aq), HCO3- and CO32-, i.e., carbon speciation.

Ideally, all environmental parameters would be known for the different locations as would the time the dinoflagellates lived and encysted. This is, however, unfeasible because the dinocysts assemblage in core-top sediments (typically the upper 2 cm of sediment) integrates conditions over several centuries, assuming moderate to low average sediment accumulation rates (< 10 cm kyr-1) that generally characterize open-ocean settings such as examined here. We therefore apply a rough correction for pCO2, based on the assumption that local air–sea gas exchange has remained similar, that equals the atmospheric pCO2 rise between the sampling date and 1850 CE. The correction from actual measurements to “pre-industrial” conditions for pCO2 yields a substantial offset due to the ∼ 90 ppmv atmospheric pCO2 rise from 1850 CE to the average sampling date (ca. 2000 CE). As this correction is broadly similar for all sample localities, it has only a small impact on the overall pattern in the CO2 data (Fig. 3). We employed a Monte Carlo simulation to assess the potential impact of the pCO2 correction by propagating (1) the 5 % analytical error in pCO2 values and adding 45 ± 15 ppm to reflect a normally distributed mixture of modern and pre-industrial conditions and (2) a resampled uncertainty derived from the pCO2 rise since pre-industrial times (1800–2000 CE). Both scenarios are set up to represent worst-case scenarios; a single error drawn from the error distribution is imposed on a sample basis and not a resampled average of the number of δ13CDINO measurements within a sample, as that would reduce the error through averaging.

Changes in SST, SSS and nutrient concentrations are also expected, partly also by anthropogenic activity, but offsets in these parameters are generally subtle and more local compared to the changes in pCO2 and hence would require site-specific reconstructions. Still, recent widespread eutrophication and enhanced productivity may impact the carbon isotope results through increased DIC uptake in algal blooms (i.e., counteracting the impact of enhanced pCO2). However, as eutrophication has mainly affected coastal areas (Hallegraeff, 1993; Anderson et al., 2002), this is expected to play a minor factor at our, mostly open marine, sample localities (Fig. 1). Lastly, long-term natural changes in nutrients, SSS and SST also occur, and it is currently not possible to fully filter out the various anthropogenic offsets. With the exception of pCO2, we hence assume all parameters (SST, SSS, nutrients) to have remained constant over the period the core-top samples represent.

Despite our pre-screening to include only samples with high relative abundances of the target species, some of the selected samples contained too few dinocysts or at too low an abundance relative to other organic particles (amorphous organic matter, plant debris, pollen, non-dinocyst marine palynomorphs etc.) to be used in our study. Ultimately, out of the initial core-top sample set of 34 samples, 19 were found suitable for species-specific dinocyst stable carbon isotope analyses (Table 1).

Typically, ∼ 150 individual cysts were picked and analyzed for a total of 20–50 measurements, amounting to three to seven cysts per carbon isotope measurement. We calculate an average signal size of 0.2 Vs for a single cyst, which amounts to a carbon yield of ∼ 6–7 ng C per cyst (van Roij et al., 2017). Although the variability in signal intensity from individual measurements suggests there is substantial intra-sample (cyst–cyst) variability, no significant offsets in average carbon content per cyst were observed between samples, suggesting the average carbon content of the cysts within each of the analyzed populations is similar. Spiniferites mirabilis is the notable exception to this rule, as far fewer specimens of this species are needed for a single δ13Ccyst measurement. Based on the signal intensity per specimen, we estimate that this larger cyst species contains twice the amount of C compared to S. ramosus, S. elongatus and O. centrocarpum.

The range of measured δ13CDINO values (∼ 5 ‰) far exceeds the variability in surface ocean δ13CDIC (∼ 1 ‰) and δ13CCO2 (∼ 2.5 ‰), implying the observed range likely reflects differences in fractionation related to changing uptake or leakage of different inorganic carbon phases (CO2 and HCO3-; Sharkey and Berry, 1985; Hoins et al., 2016a), and this hence determines most of the variability in the δ13CDINO data. Here, we quantitatively assess fractionation as a function of several environmental parameters.

The simple (non-weighted) linear regressions show poor correlations between εp-DIC and δ13CDIC for all environmental parameters, and the correlations slightly improve when compared to εp-CO2 (Tables 2, 3). However, none of the tested parameters individually explain the majority of the observed variance in εp-DIC (maximum R2; ∼ 0.1) or εp-CO2 (maximum R2 with pCO2; ∼ 0.38), despite high significance (low p values) of the regressions. The explained variance increases when polynomial regressions are applied. Several controlled growth experiments indeed show a non-linear response of εp as a function of pCO2 of the growth medium (Hoins et al., 2015), although the number of data points in such experiments limits full mathematical descriptions of the trends within the pCO2 range of this field study. Here, a second-order polynomial (quadratic) regression achieves an R2 of ∼ 0.74 and ∼ 0.79 for the non-weighted and weighted versions, respectively.

It is conceivable that other environmental parameters also significantly contribute to εp variability (Fig. 6). For example, [PO43-], [NO3] and pCO2 contribute significantly to a (linear) multiple-regression model, which takes the form of εp-CO2=c+xCO2+yPO4+zNO3, where c, x, y and z are numerical constants. The multiple-regression model using these three parameters covers ∼ 58 % of the variance in O. centrocarpum εp-CO2 (not weighted) and 67 % when weighted to number of measurements per sample. Including more parameters, such as SST, oxygen concentrations or other carbonate system parameters, explains slightly more of the observed variance but does not significantly improve the model. The residual mean standard error (RSME) of the carbon–nitrogen–phosphorus (CNP)-εp multiple-regression model is ∼ 1.45 ‰ while a linear regression with only pCO2 yields 1.7 ‰. Only weighted regressions are given here and reported ranges of the constants represent 1 standard error or equivalent percentiles in the case of Monte Carlo-simulated errors. These models have the following optimal formats:

Equation (1), linear.1εp-CO2=6.6±2.1+0.031±0.008pCO2(adjusted R2=0.48, p=0.001, RSME = 1.7 ‰) (Fig. 6b).

The two linear multiple-regression models are offset (Eqs. 3a and 4), primarily because of the carbon isotope fractionation between HCO3- and CO2. The slope with respect to pCO2 also varies slightly between the models for εp-DIC and εp-CO2 due to the temperature-dependent fractionation between HCO3- and CO2, but the slopes with NO3- and PO43- are indistinguishable. The quadratic regression seemingly better fits the variability observed in εp-CO2 compared to other (multiple) linear regressions and removes any structure in the residuals, potentially signaling a non-linear response in εp-CO2 to pCO2. The quadratic regression also indicates insensitivity to pCO2≤240 µatm and should not be used below this value (Fig. 6b). The Monte Carlo simulations of scenarios where an additional pCO2 uncertainty is imposed as a normally distributed mixture of pre-industrial and modern values, offsetting pCO2 by +45 ± 15 ppm (Fig. 7c, Eq. 2c), and a resampled uncertainty derived from the pCO2 rise since pre-industrial times (Fig. 7d, Eq. 2d) shows that the parameters of the quadratic regression are fairly robust to these uncertainties (i.e., none of the parameters become insignificant at p>0.05), although the absolute pCO2 values and errors increase.

The recorded δ13CDINO range and absolute values (∼ -18 ‰ to -35 ‰) correspond well with global δ13C values previously reported for marine particulate organic matter (δ13CPOC) (e.g., Freeman and Hayes, 1992; Goericke and Fry, 1994) and modeled phytoplankton biomass (e.g., Magozzi et al., 2017; Tagliabue and Bopp, 2008). Consequently, εp-DIC and εp-CO2 are also within the expected range for general marine particulate organic matter. However, the intra-sample variance of δ13CDINO appears to be substantial, often spanning most of the full range (∼ 10 ‰) observed for δ13CPOC. Some of the observed variability might be related to the limited analytical precision during measurements of the extremely small numbers of carbon of individual dinocysts. Fully constraining the contribution of this analytical uncertainty to the observed variance is, however, not possible because of unresolvable micrometer-scale heterogeneity in the PE standard (van Roij et al., 2017; Sluijs et al., 2018). In most cases, the variance in δ13CDINO is similar to that of the standard. Still, it is likely that some of the seasonal δ13CDIC differences are also recorded in the δ13CDINO and that some additional inter-specimen δ13C variance is present. This is to be expected since the δ13C populations from our integrated core-top samples span seasons and decades and thus also considerable variability in seawater properties and population change. In addition, growth-induced randomness and changes in δ13C and DIC in the cell's microenvironment likely contributed to inter-specimen variability. Note that in our data inter-specimen variability is still underestimated because we analyzed three to seven specimens per ablation event, as single-cyst carbon yield (∼ 7 ng C) from these cyst-sizes approached the limit for reliable measurements (van Roij et al., 2016). We minimized potential influence of differences in cell size or shape through manual selection. We thus analyzed a population where the pre-selection of similar-sized cysts restricts the variance in cell surface area and volume, unlike biomarker-based proxies for which the cell size has to be reconstructed independently (Henderiks and Pagani, 2007; Stoll et al., 2019). This approach could reduce scatter in the relation of εp to environmental variables (Popp et al., 1998).

One of the striking differences between the here generated data and the existing culture experiments is that carbon isotope fractionation of dinocysts in the natural environment appears to be much larger than that of motile cells from controlled growth (dilute batch) experiments (Hoins et al., 2015, 2016b). We find average εp values ranging between 13 ‰–20 ‰ and 23 ‰–29 ‰ with respect to CO2 and DIC. Cultured cells of O. centrocarpum yielded not only a smaller overall εp but also a smaller range (∼ 8 ‰–12 ‰ and 18.5 ‰–20 ‰) across a larger CO2 gradient, implying the cysts have a much steeper fractionation slope with CO2 compared to the motile cells. Despite these differences, the average εp for Spiniferites species (S. ramosus, elongatus and mirabilis) is often somewhat larger than for O. centrocarpum (Fig. 4). This is consistent with culture experiments that showed larger CO2 dependency and overall slightly larger εp in the motile species G. spinifera compared to P. reticulatum (Hoins et al., 2015).

While the cultured single strains and dinoflagellate populations in nature may behave somewhat differently, we do not expect that this alone underlies such a marked offset between the motile cultured cells and natural cysts. Natural cysts and cultured cells seem consistently offset in δ13C, although at present the exact amplitude of this offset cannot be determined. However, such an offset is in line with certain compounds in dinoflagellate cells being depleted in 13C relative to the bulk biomass (Schouten et al., 1998; Wilkes et al., 2017). The organic-walled dinocysts consist of mostly aliphatic and aromatic compounds, forming a complex biopolymer referred to as dinosporin (de Leeuw et al., 2006; Versteegh et al., 2007, 2012). Depending on the biosynthetic pathway of the cyst material and the derivation or degradation of the original compounds, this may result in offsets in δ13C values between cysts and the motile cells. A potential additional fractionation might be introduced during taphonomy and also later by the processing of sediments to concentrate the dinocysts from sediment samples. The sediment processing involves hydrochloric and hydrofluoric acids, which affects the more labile organic compounds. Last, it is conceivable that fractionation in the dilute batch experiments may be reduced by, e.g., higher-than-natural growth rates. This may be supported by chemostat culture experiments on Alexandrium tamarense (Wilkes et al., 2017), which show a (much) greater fractionation compared to the dilute batch experiments (Hoins et al., 2015). However, the enhanced fractionation recorded in chemostat experiments is likely an artifact of isotope equilibration times exceeding CO2 uptake rates (Brandenburg et al., 2022; Zhang et al., 2022). The range of options cannot be narrowed down until cultured cysts are compared to their motile cells harvested from the same culture and treated with similar techniques as used for the sediments. Until these data become available, inferences on the origin and amplitude of the offsets between the cells and cysts of O. centrocarpum and Spiniferites remain speculative.

Carbon isotope fractionation is determined by RuBisCO and several environmental parameters, dominantly pCO2 but also cell size and shape, growth rates, and nutrient or light regimes (e.g., Freeman and Hayes, 1992; Pagani, 2013; Popp et al., 1998; Stoll et al., 2019, and many others). In most cases, fractionation is pCO2-dependent, and consequently εp in various groups and genera has been used as a paleo-pCO2 proxy (e.g., Freeman and Hayes, 1992; Pagani, 2013; Rae et al., 2021). We performed a broad-spectrum multiple-regression analysis to identify environmental factors that contribute significantly to dinocyst εp. The most important parameter is pCO2, in line with previous studies (Freeman and Hayes, 1992). A large part of the remaining variability can be attributed to growth rate and ultimately nutrient content, specifically nitrate and phosphate of the surface waters, in line with previous studies on phytoplankton and dinoflagellate carbon isotope fractionation (Popp et al., 1998; Hoins et al., 2016b; Wilkes et al., 2017; Wilkes and Pearson, 2019). We find that pCO2 and εp are positively correlated, while NO3- and PO43- are negatively correlated with εp (Fig. 6c, d).

The inclusion of nutrient levels as environmental factors in the carbon isotope fractionation reduces the error in the fit between the measured and modeled fractionation. In the rare cases where paleo-nutrient concentrations are estimated, reconstructions usually cover only [PO43-], implying the [NO3-] in Eqs. (3a) and (4) has to be approximated from [PO43-]. As [PO43-] and [NO3] are generally well-correlated (typically ∼ 1:10 [PO43-] : [NO3]; here ∼ 1:8) in ocean waters, this is relatively straightforward. For application in the paleo-domain, however, nutrient ([PO43-]) reconstructions may not be available or may be considered too unreliable to provide meaningful constraints. In specific cases where accurate paleo-[PO4] estimates or large changes are reconstructed, Eq. (3b) may be applied, in which [NO3] is substituted for [PO43-] at a 1:8 ratio. However, unless there are clear changes in the nutrient regime or sufficient proxy constraints on the nutrient concentrations, we prefer a calibration based exclusively on carbon isotope fractionation and carbonate system parameters, including sea surface temperature to calculate δ13CCO2 from δ13CDIC to reconstruct pCO2.

As many phototrophs, including dinoflagellates, have mechanisms for concentrating CO2 near the cell membrane, the sensitivity of carbon isotope fractionation to ambient CO2 is expected to diminish below a certain concentration (M. R. Badger, 2003; M. P. S. Badger, 2021; Stoll et al., 2019). This is particularly the case as dinoflagellates utilize type II RuBisCO, which has generally poorer performance compared to type I RuBisCO at low CO2 concentrations (Giordano et al., 2005). Indeed, substantial activity of the carboxyl anhydrase (CA) enzyme, which facilitates the conversion of HCO3- to CO2 inside the cell or near the membrane, was observed in numerous dinoflagellate species, including Lingulodinium (Lapointe et al., 2008), Symbiodinium (Leggat et al., 1999) and the here analyzed Operculodinium (Ratti et al., 2007). Also, here we find a relatively low sensitivity at the lower end of the CO2 scale. The lower CO2 values correspond to the northern-most locations, with trends below 240 µatm becoming somewhat obscured, at a minimum εp-CO2 around 13 ‰ (Fig. 6b). However, part of this leveling of the proxy relationship may reflect the locally higher nutrient concentrations offsetting the higher CO2. Though growth rates have a clear influence on εp in algal groups (Burkhardt et al., 1999), including dinoflagellates (Wilkes et al., 2017), the dilute batch culturing experiments conducted with P. reticulatum showed no clear influence of growth rates on εp (see also Sect. 4.2). It is also conceivable that higher growth rates influence εp indirectly through, for example, seasonally enhanced CO2 drawdown, resulting in higher δ13C values in the remaining DIC. This effect may be enhanced by the relatively short growing season at the high latitudes. However, in culture experiments at low CO2 concentrations with other dinoflagellate species, 13C fractionation was higher under nutrient-limiting conditions than under replete conditions (Hoins et al., 2016b). Because of these confounding factors, the influence of carbon-concentrating mechanisms on εp in O. centrocarpum is difficult to gauge with the presently available data and would ideally be tested using high nutrient or very low CO2 concentrations.

Still, also in the relatively limited range the current ocean offers for testing pCO2 proxies we have established a robust, albeit not overly sensitive, relation between pCO2 and dinocyst δ13C. Our cyst-based calibration yields more conservative and arguably more realistic absolute CO2 estimates and variability compared to available culture-based calibrations as it is based on the same compounds as will be analyzed in the paleo-domain. However, the low sensitivity at low CO2 implies that, until better constraints become available, the proposed calibration is potentially less suitable for application across, for example, the Pleistocene glacial periods. Further, it is important to realize that the value of 240 µatm is based on the assumption that the εp-CO2–pCO2 relation originated from cysts that have not been affected by the Suess effect and thus represent a lower limit for CO2 (in)sensitivity. While our data do not preclude fractionation smaller than the here observed minimum (∼ 13 ‰) during low-pCO2 periods, increased sensitivity at higher CO2 suggests that CO2 above (minimum) 240 µatm and CO2 variability can be reconstructed with reasonable precision.

A topic that warrants specific attention is the potential impact of anthropogenic carbon emissions on shaping the relation between δ13CDINO and εp. Here, we assume that our outlier analyses preferentially excluded samples that were significantly affected by anthropogenic CO2 and that, on average, the remainder of our core-top δ13CDINO was not appreciably affected by either decreasing δ13CDIC or elevated pCO2. The similarity in pre-industrial down-core δ13CDINO and that of three core-top localities in the North Atlantic corroborate the validity of this assumption locally, but these observations cannot be extrapolated to other regions. Unfortunately, sedimentation rates or other constraints for cyst production datums are not available. Further, if sedimentation rates were available for core-top localities, that would constitute an imperfect solution as it cannot provide the required dinocyst-specific age distribution needed to obtain an appropriate local pCO2 correction to the datum of cyst production. This challenge may be unique to data such as those presented here, as studies on other substrates with pCO2-proxy potential either could not generate data for individual single-celled organisms or have avoided the issue through other means, such as culture experiments (e.g., Pagani et al., 2002; Henehan et al., 2013), approaches that have other drawbacks.

Although we have no constraints on the ages of the cysts analyzed here, we can provide a meaningful test of the potential uncertainty added by our assumption that cysts are representative of pre-industrial conditions (Eq. 2a, b). The two scenarios that we have explored through Monte Carlo simulations show that, depending on the error distribution imposed on the assumed pCO2 used in the quadratic regressions (Eq. 2a, b), the fitted regression shifts towards higher values (Fig. 7c) and may be steeper (Fig. 7d). These error-propagation case studies illustrate that our proposed pre-industrial transfer function (Eq. 2a, b), if it does indeed contain a substantial proportion of very recent dinocysts, is likely to lead to underestimated pCO2 and perhaps pCO2 variability when applied in the paleo-domain. Consequently, we recommend future studies target, for example, sediment trap collections and culture-derived dinocysts to validate the results obtained here.