• Open Access

Long-term evolutionary and ecological responses of calcifying phytoplankton to changes in atmospheric CO2



Calcifying phytoplankton play an important role in marine ecosystems and global biogeochemical cycles, affecting the transfer of both organic and inorganic carbon from the surface to the deep ocean. Coccolithophores are the most prominent members of this group, being well adapted to low-nutrients environments (e.g., subtropical gyres). Despite urgent concerns, their response to rising atmospheric carbon dioxide levels (pCO2) and ocean acidification is still poorly understood, and short-term experiments may not extrapolate into longer-term climatic adaptation. Current atmospheric pCO2 (~390 ppmv) is unprecedented since at least 3 million years ago (Ma), and levels projected for the next century were last seen more than 34 Ma. Hence, a deep-time perspective is needed to understand the long-term effects of high pCO2 on the biosphere. Here we combine a comprehensive fossil data set on coccolithophore cell size with a novel measure of ecological prominence: Summed Common Species Occurrence Rate (SCOR). The SCOR is decoupled from species richness, and captures changes in the extent to which coccolithophores were common and widespread, based on global occurrences in deep-sea sediments. The size and SCOR records are compared to state-of-the-art data on climatic and environmental changes from 50 to 5 Ma. We advance beyond simple correlations and trends to quantify the relative strength and directionality of information transfer among these records. Coccolithophores were globally more common and widespread, larger, and more heavily calcified in the pre-34 Ma greenhouse world, and declined along with pCO2 during the Oligocene (34–23 Ma). Our results suggest that atmospheric pCO2 has exerted an important long-term control on coccolithophores, directly through its availability for photosynthesis or indirectly via weathering supply of resources for growth and calcification.


Marine phytoplankton are responsible for about half of the global annual net primary productivity, fixing ~45–50 Pg carbon per year (Longhurst et al., 1995; Field et al., 1998). A third of this organic carbon is exported to the deep ocean (Falkowski et al., 1998), effectively removing carbon from the atmosphere (a process termed the biological pump). Diatoms are very efficient at inorganic carbon fixation as long as nutrients and silica are available, but in stratified ocean gyres and other low-nutrient settings they are outcompeted by other, smaller-sized groups, such as coccolithophores (calcifying haptophyte algae; Winter & Siesser, 1994) and cyanobacteria (Raven, 1998).

Calcification by coccolithophores is an additional component of biogenic carbon fixation, affecting ocean-atmosphere CO2 exchange on time scales ranging from seasonal blooms to glacial-interglacial cycles (Holligan et al., 1993; Rost & Riebesell, 2004; Barker et al., 2006; Sarmiento & Gruber, 2006). Today, planktonic calcification is estimated to be ~3% of pelagic primary production (Balch et al., 2007; Poulton et al., 2007) with an annual global production of 1.1–1.6 ± 0.3 Pg calcite (Feely et al., 2004; Balch et al., 2007). Coccolithophores rose to global prominence during the Cretaceous and have been the main pelagic producers of carbonate since ~100 million years ago (Ma), which significantly altered global carbon cycling by shifting the main locus of carbonate burial from continental shelves to the deep sea (Zeebe & Westbroek, 2003; Hay, 2004).

Conversely, the level of CO2 in the atmosphere affects calcifying algae and other carbonate-secreting marine organisms via its coupling with ocean carbonate chemistry. Substantial efforts are currently made to predict the response of coccolithophores to ocean acidification, a lowering of seawater pH caused by the rapid anthropogenic rise in atmospheric CO2, with results so far being inconclusive (Riebesell et al., 2000; Langer et al., 2006; Iglesias-Rodriguez et al., 2008; Ridgwell et al., 2009; Irie et al., 2010; Beaufort et al., 2011; Findlay et al., 2011; Lohbeck et al., 2012). Model calculations reveal that it will take tens of thousands of years before the atmospheric partial pressure of CO2 (pCO2) reaches steady state and ocean pH returns to pre-industrial values (Archer, 2005; Uchikawa & Zeebe, 2008).

The present pCO2 is unprecedented since at least 3 Ma (Seki et al., 2010), while the last period of sustained high-CO2 (greenhouse) conditions in Earth history ended ~34 Ma. Researchers are therefore turning to past episodes of greenhouse warming in the Cenozoic era (the past 65 million years) for insights into the coupling of climate and the carbon cycle and the consequences of continued anthropogenic carbon emissions (Hansen et al., 2008; Zachos et al., 2008). How did coccolithophores respond to large changes in atmospheric pCO2 in the deep past?

Here we address this question by combining data on global occurrences of common coccolithophore species from deep-sea sedimentary records with a comprehensive fossil data set on coccolithophore cell size to infer evolutionary and ecological changes over the interval 50–5 Ma. We use occurrences of common species to quantify relative changes in the extent to which coccolithophores were globally common and widespread. This quantity is decoupled from species richness. Size is a key trait in life history and physiology (Peters, 1983; Calder, 1984) that influences rates of photosynthesis, respiration, and growth in algae (Raven, 1998); rates of calcification (Stoll et al., 2002) and sinking rates (Raven & Waite, 2004); as well as genome size and evolution (Oliver et al., 2007). We analyze the relationship of coccolithophore size and occurrences to geological records of coupled environmental variables that are relevant to planktonic ecosystems, including changes in temperature, sea level, ocean carbon chemistry, and carbon cycling.

Disentangling the relationships among these variables is not trivial, because the plankton and paleoenvironmental records capture components of the Earth system that are inextricably linked, with potentially nonlinear, scale-dependent, and time-varying interactions. For example, the evolving relationship between pCO2, temperature, and terrestrial rock weathering rates in the Cenozoic may have involved biotically mediated negative feedbacks (Li et al., 2009; Pagani et al., 2009). Inferring causal interactions between plankton evolution and climatic changes may ultimately require multi-scale Earth system modeling, however, sufficient knowledge regarding the relevant mechanisms and time scales is lacking. On the other hand, simplistic matching of observed wiggles and trends does not distinguish correlations from drive-response relationships. We therefore take an alternative, information-theoretic approach to detecting potential drive-response relations directly from observed records without making mechanistic assumptions (Schreiber, 2000; Verdes, 2005; Vejmelka & Paluš, 2008; Hannisdal, 2011a). We use this tool to test the hypothesis of a causal relationship between Cenozoic climate variability and coccolithophore macroevolution.

Materials and methods

Paleoenvironmental records

Oxygen isotope records (δ18O) from epibenthic foraminiferal calcite (Zachos et al., 2001, 2008) represent a composite signal of deep-sea (and high-latitude surface ocean) temperatures and terrestrial ice volume, relating to large-scale changes in ocean thermal structure and circulation, which are key parameters for plankton communities (Finkel et al., 2007). Epibenthic foraminiferal carbon isotope (δ13C) records (Zachos et al., 2001, 2008) are linked to ocean circulation and changes in global carbon cycling (e.g., massive methane releases, shifts in organic carbon and carbonate burial rates). Cenozoic sea-level variations (Miller et al., 2005) are attributed to temperature and ice volume changes and have affected carbon cycling by transforming continental shelf areas from sites of carbonate accumulation to sites of carbonate exposure and weathering (Merico et al., 2008). Finally, geological records (Pearson & Palmer, 2000; Pagani et al., 2005; Lowenstein & Demicco, 2006; Royer, 2006) and modeling studies (Berner & Kothavala, 2001; Li et al., 2009) agree on a long-term Cenozoic decline in pCO2, generally attributed to a global decrease in volcanic CO2 degassing, and/or tectonically enhanced weathering of silicate rocks accompanied by increased carbon burial (Raymo & Ruddiman, 1992; but see Willenbring & Von Blanckenburg, 2010).

Here we used a record of Cenozoic atmospheric pCO2 derived from the carbon isotopic composition of alkenones (long-chained organic compounds) (Pagani et al., 2005, 2011), a method that has been shown to agree with boron isotope data when studied in the same deep-sea sediment samples (Seki et al., 2010). In the modern ocean, production of alkenones is restricted to a few species of marine haptophyte algae within the order Isochrysidales, including the coccolithophore Emiliania huxleyi. Ancient alkenone-producing coccolithophores belong to the family Noelaerhabdaceae Jerkovic, 1970 (including the reticulofenestrids – ancestors of E. huxleyi) (Henderiks & Pagani, 2008). Other prominent Cenozoic coccolithophore families, such as the Coccolithaceae Poche, 1913, diverged from the Isochrysidales long before the first sedimentary evidence of alkenones (Farrimond et al., 1986; Brassell et al., 2004; Medlin et al., 2008) and have no known alkenone-producing representatives today. We merged the alkenone data with boron isotope estimates from foraminiferal calcite available over the same period (Pearson & Palmer, 2000; Pearson et al., 2009).

SCOR – Summed Common Species Occurrence Rate

We used coccolithophore species occurrence data as documented in the Neptune database of Cenozoic microfossils from Deep Sea Drilling Program (DSDP) and Ocean Drilling Project (ODP) sites (Lazarus, 1994, 2011; Spencer-Cervato, 1999). Each datum records the presence of an identified species in a unique sample from a deep-sea core associated with current day spatial coordinates of the site and an absolute geological age estimated by core-specific age models (Cande & Kent, 1995; Spencer-Cervato, 1999). To minimize the influence of time-varying boundary conditions at continental margins (e.g., river plumes, currents, and sea level) and equatorial upwelling, occurrences from 50 to 5 Ma were derived from sites located within the boundaries of major oceanic gyre systems (Fig. S1), using 1/3 million year (Myr) time bins, the maximum resolution afforded by Neptune for the Paleogene period (Spencer-Cervato, 1999). We used only species that were fully resolved taxonomically and present in at least 10 unique samples (irrespective of their assigned ‘abundance’ category in the Neptune database), leaving a total of 34 320 occurrences of 166 coccolithophore species.

Let Yj be the number of individuals of a given species in a given timebin, j. Yj has to be at least one for us to have any possibility of sampling that species. The probability pj of detecting the species in timebin j is hence 1 minus the probability that Yj is exactly zero:

display math

We assume that individuals of any given species i are randomly distributed in each timebin globally. A large number of individuals is possible, but observing a specific number of individuals in a given timebin can be considered a rare event and modeled as a Poisson-distributed variable. We note that the probability of finding one individual of a species is not independent of that of finding another individual of the same species, which violates an assumption of the Poisson. However, relaxing this assumption by using geometric or other distributions has little effect on the results (simulations not shown). Under the Poisson model, the probability of not finding an individual of the species i is inline image, where λij (the rate parameter) is the expected number of individuals of species i in timebin j, such that

display math


display math

In practice, pij is estimated as inline image, where the numerator is the number of samples in which species i is present (irrespective of its assigned ‘abundance’ category in Neptune) at time j and the denominator is the number of available sites in timebin j. Available sites are those that contain at least one species in the set of common species included in the analysis. For instance, if there were 100 available sites in the age bin 50–49.7 Ma in Neptune, and species i was present (and detected) in 5 of them, then λ for species i would be −ln(1−0.05).

We then define the Summed Common species Occurrence Rate (SCOR) as the total density of a given set of mj species in a particular timebin:

display math

and estimate the variance of this quantity by the delta method:

display math

SCOR is intended to reflect relative changes in the extent to which a set of species is globally common and widespread. As p approaches 1, the rate of increase in λ grows rapidly, such that very common/widespread species have a much greater influence on SCOR than less common species. If a species is found in every available sample in a time bin, its λ for that time bin will go to infinity. We thus omitted 35 time bins in which one, or at a maximum three, species (of Coccolithus and Chiasmolithus, as well as Cyclicargolithus floridanus) are recorded at all sites, leaving 101 time bins in the SCOR record. The number of sampled sites (holes), n, also changes with time, the effect of which is discussed below.

At this point, it is necessary to flag a few key issues: (1) Total species richness of a given taxonomic group is driven by rare species (most species are rare; McGill et al., 2007), whereas its total abundance is driven by common species. (2) Most fossil species have a ‘hat’-shaped temporal occurrence trajectory, being rare in the early and late stages of their life (Foote, 2007; Liow & Stenseth, 2007; Liow et al., 2010). (3) The SCOR is decoupled from species richness and relative abundance. As an example, consider a world composed of only two species, where each species is present in 99.99% of all samples. Richness is thus extremely low, but the SCOR is high (−ln[1−0.9999]*2 = 18.42). Conversely, if the world is composed of 100 species, each present in only 5% of all samples, then richness is high, whereas the SCOR is low (−ln[1−0.05]*100 = 5.13). Relative abundance is equal in both cases. We illustrate the relationships between SCOR, species richness, and sampling using a simulated database (named Uranus) in the Supporting information. (4) Absolute abundance, population sizes, or global fluxes cannot be measured directly from species occurrence data in Neptune. SCOR is based on the assumption that the more globally abundant and widespread a species is, the more likely it is to occur in a greater number of samples, all else being equal. If a species becomes globally more abundant and widespread in a time interval, then its λ will increase, and the SCOR will follow suit. Even if all species became exactly equally more common in absolute terms, with no change in relative abundance, their individual λ values will be higher and SCOR will capture the proportional change in absolute abundance. (5) In principle, the occurrence distribution of common species could reflect varying deep-sea carbonate preservation and/or artifacts of ODP sampling protocol. However, there is little evidence to suggest that the SCOR estimates are fundamentally biased by preservation or sampling (see 'Results'). We return to these issues below.

Coccolithophore cell size

Cenozoic size records were compiled for the reticulofenestrids, the ancestors of today's most prolific blooming coccolithophores (Henderiks & Pagani, 2008). Size measurements were made on fossil coccoliths, the calcite platelets produced by coccolithophores (Fig. 1), which provide a robust estimate of cell size variability and cellular calcite quota in the geological past (Henderiks, 2008). The maximum diameter (length) of individual coccoliths was measured with polarized light microscopy combined with a digital image analysis set-up, using four replicate slides for each sediment sample (Henderiks & Törner, 2006). We used deep-sea sediment samples that were selected from a suite of DSDP/ODP drilling sites covering tropical to temperate regions and all major ocean basins (Atlantic, Indian and Pacific oceans; Fig. S1). Age models were based on DSDP/ODP biostratigraphy calibrated to the geomagnetic polarity time scale (Cande & Kent, 1995). The size variability of the reticulofenestrid lineage over the time interval 45–5 Ma was derived from a total of 24 534 coccoliths, which, at a maximum resolution of 10 Kyr, and omitting samples with fewer than 20 specimens, yielded 161 sampled time bins with a median sample size of 174 specimens.

Figure 1.

Scanning electron microscopy image of two modern coccolithophores (a) Emiliania huxleyi and (b) Coccolithus braarudii, illustrating large size differences between species. The size of the coccoliths (individual calcite platelets) is proportional to the diameter of the cell they surround.

Information transfer analysis

A causal interaction between a driving variable A and a response variable B can be considered an inherently directional relationship (A→B). Statistical approaches to detecting the directionality of coupling commonly rely on a computational, or predictive, definition of causality by which knowledge of changes in A should improve our ability to predict changes in B (Granger, 1969). Information theory provides a general framework for implementing computational causality detection in time series analysis (Hlaváčková-Schindler et al., 2007), by quantifying information flow between variables. These techniques are applied across disciplines, including physiology (Pahle et al., 2008; Staniek & Lehnertz, 2008; Schippers et al., 2010), econometrics (Marschinski & Kantz, 2002), and earth sciences (Materassi et al., 2007; Hannisdal & Peters, 2010, 2011).

Here we used a non-parametric measure of directional information transfer based on the construct of transfer entropy (Schreiber, 2000; Verdes, 2005) to detect the relative strength and directionality of interactions among SCOR, size, and paleoenvironmental records. Unlike measures of correlation, information transfer is non-symmetric (A→B ≠ B→A), and is able to detect directional information flow resulting from a drive-response relationship, even in short and unevenly sampled geological records (Hannisdal, 2011a,b). We implemented the method in a two-step analysis: first, we estimated the relative influence each potential forcing factor has on a response variable, while taking into account mutual correlations among the factors (Verdes, 2005), with significance levels established by randomly shuffling the forcer time series (Fig. S2). This analysis will be referred to as conditional information transfer.

Second, we tested whether the information transfer between two variables is significantly directional compared to amplitude-adjusted fast Fourier transform surrogate time series (Schreiber & Schmitz, 2000; Vejmelka & Paluš, 2008). The surrogates are replicates of the original data that preserve the amplitudes and frequencies, but where the phases of the frequency components have been randomized, thus destroying any causal linkage. Crucially, by having the same distribution of states and the same frequency power spectrum as the original data, the surrogates effectively prevent false positive results that could arise from frequency bias (Paluš & Vejmelka, 2007; Vejmelka & Paluš, 2008) and have the potential to distinguish between correlation and causation (Fig. S3). This analysis will be referred to as directional information transfer (Hannisdal, 2011a). Paleoenvironmental time series were resampled to match the sampled time bins of the coccolithophore SCOR (N = 101) or size (N = 161) records. Before analysis, all records were detrended, power transformed (Box-Cox) to stabilize the variance, and normalized to mean zero and unit standard deviation. The information transfer analysis can be sensitive to differences in non-stationarity, and to minimize this possible bias the time series were detrended by subtracting the best-fit fifth-order polynomial, rendering the time series stationary for lags up to N/2 according to the KPSS test (Kwiatkowski et al., 1992).


Coccolithophore SCOR and size covary with pCO2

Two major placolith-bearing families, the Noelaerhabdaceae (including the reticulofenestrids) and Coccolithaceae (including the genera Coccolithus and Chiasmolithus) (Young & Bown, 1997) dominate the coccolithophore SCOR. Our estimates show that the SCOR of common coccolithophore species peaked in the middle Eocene and then declined irregularly through the Oligocene, followed by relatively stable, low SCOR in the Miocene (Fig. 2a). Following a peak in the late Eocene, maximum cell size in the reticulofenestrid lineage shows a marked, but irregular decrease through the Oligocene reaching a minimum in the early Miocene (Fig. 2b). The Oligocene size trend is attributed to the successive extinction of large morphospecies within this lineage, with a minor contribution from intraspecific size changes (Henderiks & Pagani, 2008). On the other hand, members of the Coccolithaceae, which maintained larger average size throughout the investigated time interval (Henderiks & Rickaby, 2007), are primarily responsible for the marked decline in SCOR (Fig. 3). We note that the SCOR patterns do not covary with proposed changes in species richness (Fig. S4; Bown et al., 2004). Taxonomic differences notwithstanding, the net result is a substantial global decrease in both size and SCOR of coccolithophores as a whole. This pattern has several features in common with reconstructions of Cenozoic changes in atmospheric pCO2 (Fig. 2c; Pearson & Palmer, 2000; Pagani et al., 2005, 2011; Pearson et al., 2009).

Figure 2.

Time series of coccolithophore Summed Common Species Occurrence Rate (SCOR) and size compared with Cenozoic environmental records. (a) SCOR estimated from common coccolithophore species occurrences in the major oceanic gyres (Fig. S1) with a two standard error envelope (2 SE). (b) Reticulofenestrid coccolith size (length) measurements with the 95th percentile highlighted (P95 size). (c) Estimates of atmospheric pCO2 based on the δ13C of alkenones (gray; Pagani et al., 2005, 2011) and the δ11B of foraminifera (open circles; Pearson & Palmer, 2000; Pearson et al., 2009). Solid line represents bin-averaged values for the time bins (= 101) of the SCOR record. In the case of gaps in the pCO2 data, time bins are expanded by adding the two adjacent bins until an average can be calculated. Note log scale. (d, e) Deep-sea carbonate δ18O and δ13C (Zachos et al., 2001, 2008). Solid lines are 6-point LOWESS filtered curves resampled at the SCOR time bin midpoints. (f) Estimated global sea-level changes (Miller et al., 2005), directly resampled (solid line) at the SCOR time bin midpoints.

Figure 3.

Comparing the Summed Common Species Occurrence Rate (SCOR) of different groups of calcifying plankton from the Neptune database. (a) Global SCOR estimates show a different pattern in the coccolithophores (black) compared to other calcareous nannofossils (nannoliths, including e.g., Sphenolithus and Discoaster; green) and to planktonic foraminifera (red) from the same sites (Fig. S1). (b) Coccolithophore SCOR (black) is dominated by two of the most prominent families, the Coccolithaceae (blue; not known to produce alkenones) and the Noelaerhabdaceae (red; including ancient alkenone producers, the reticulofenestrids). Non-alkenone producing coccolithophores are primarily responsible for determining the global SCOR pattern.

Could the observed trends in coccolithophore SCOR and size be artifacts of carbonate sediment preservation? Deep-sea dissolution is crucial in maintaining the steady state between weathering supply and burial (removal) of carbonates, and the depth at which most carbonate is dissolved can be traced as a ‘snowline’ on the seafloor, known as the carbonate compensation depth (CCD). If the burial of carbonate exceeds supply, more carbonate gets dissolved and the CCD will shoal until a new steady state is reached (Zeebe & Westbroek, 2003). However, a major preservational bias is unlikely, given the following observations: (1) If deep-sea carbonate preservation were the overriding determinant of the coccolithophore SCOR, then we would expect other calcareous microfossils from the same carbonate samples to show very similar SCOR patterns. On the contrary, different coccolithophore families, other calcareous nannofossils, and planktonic foraminifera in the Neptune database all show divergent overall trends as well as different short-term variability in their SCOR patterns (Fig. 3). Note that the absolute SCOR values are not directly comparable among groups, and should not be interpreted as a measure of the relative abundance of different groups. All time series are normalized prior to time series analysis, which compares the relative changes through time in terms of transition probability distributions. (2) A rapid deepening of the CCD is recorded at the Eocene/Oligocene boundary (Fig. S5; Van Andel, 1975; Coxall et al., 2005), which has been attributed to glacioeustatic sea-level fall causing exposure and dissolution of shallow-marine carbonates (Merico et al., 2008). If dissolution had driven the overall decline in SCOR, then we would expect an opposite shoaling CCD trend over the Cenozoic. (3) Conversely, if the deepening of the CCD had induced an apparent size decrease by enhanced preservation of small coccoliths, then we would expect a trend in the minimum size, but the observed size decrease is in the maximum only (Fig. 2b). Less dissolution cannot explain the selective loss of large forms. (4) The transfer of coccoliths from surface waters to deep-sea sediments imposes a harsh preservational filter, but for this filter to distort our results it would have to operate differently at different times. We find no clear systematic changes in preservation indicators, including the relationship between Neptune site water depth, sample burial depth, and the number of species (Fig. S6a–c); the semi-quantitative measure of average state of preservation (Fig. S6d); the range of paleodepths sampled through time (Fig. S6e); or the proportion of dissolution-susceptible vs. resistant taxa as indicated by the ratio of murolith to placolith morphotypes (Fig. S6f). (5) Whereas the number of species is strongly positively correlated with the number of holes sampled, the coccolithophore SCOR is not significantly correlated with the number of holes and shows weak negative correlation with the number of species (Fig. S7). (6) The SCOR is theoretically decoupled from species richness, because only the most common species are relevant. The coccolithophore SCOR pattern can be obtained with fewer than 50 of the most commonly occurring species included in the set, and the addition of less common species has little effect (Fig. S8), making SCOR more robust to preservation, dilution, and/or sampling than sampling-standardized species richness (Fig. S9).

Today, marine calcifiers overproduce calcite by up to four times more than what is available through riverine supply from weathering (Broecker & Peng, 1982). However, surface waters remain supersaturated with respect to calcite, and this has probably been the case for most of the past 100 million years (Tyrrell & Zeebe, 2004). The answer to this apparent contradiction is that up to 80% of carbonate is recycled in the upper 500–1000 m of the ocean, most likely by respiratory dissolution (Archer, 1996; Milliman et al., 1999). In the Eocene, coccolithophores were overall more common and widespread, and had overall larger size. In principle, this could imply higher coccolithophore carbonate productivity, which could have been sustained by similar efficient shallow carbonate recycling; by increased supply of building blocks from weathering (Li et al., 2009); and/or by the shallower Eocene CCD.

Relative information transfer from environmental records

Atmospheric CO2 plays a key role in the coupled geo-biosphere system, hence the observed covariation between Cenozoic pCO2 and coccolithophore SCOR is likely to involve common interactions with changes in temperature and other climatic factors (Fig. 2). To assess the importance of such shared responses and possible confounding factors, we quantified the conditional information transfer between the SCOR and pCO2 records while excluding shared information with the Cenozoic δ18O, δ13C, and sea-level records (Fig. 4).

Figure 4.

Information transfer analysis quantifying the relative strength of statistical dependencies between coccolithophore Summed Common Species Occurrence Rate (SCOR) and size and Cenozoic environmental records. (a) Conditional information transfer from pCO2 to SCOR (X) excluding shared information with δ18O (pCO2 : X, δ18O) and from δ18O to SCOR given pCO218O : X, pCO2). Scale values along abscissa correspond to the magnitude of changes in the normalized SCOR record in units of standard deviation. Solid and dashed gray lines are 95th percentiles of 5000 replicate calculations with randomly shuffled pCO2 and δ18O records, respectively, and information transfer values exceeding these are considered significant. (b, c) As in (a), but comparing pCO2 to δ13C and sea level (SL), respectively. (d–f) As in (a–c), but with size as the target (X) variable.

Information transfer varies as a function of the scale of the changes in the target time series, and the total area under the resulting curve is used as a basis for assessing statistical significance. For example, the conditional information transfer from pCO2 to coccolithophore SCOR is significant despite mutual correlations with δ18O, whereas the δ18O record does not provide any significant information on changes in SCOR beyond that contained in the pCO2 record (Fig. 4a). Overall, pCO2 shows a dominant statistical dependence with coccolithophore SCOR relative to the other environmental factors (Fig. 4a–c). Mutual responses to changes in temperature and other climatic factors thus seem less likely to explain the relationship between CO2 and coccolithophores on million-year time scales. Conditional information transfer analyses of the reticulofenestrid size data give very similar results (Fig. 4d–f), with one exception: δ13C also shows a significant statistical dependence with the size record, albeit less pronounced than pCO2 (Fig. 4e). In contrast, the reconstructed SCOR curves for nannoliths and planktonic foraminifera (Fig. 3a) did not yield significant information transfer with any of the Cenozoic paleoenvironmental records. Among all the variables (Fig. 2), the strongest overall relationship is found between the δ18O and δ13C records, but interrelationships among the paleoenvironmental variables are not pursued further here.

Coccolithophores and CO2 – testing the causal hypothesis

A coupling between atmospheric pCO2 and coccolithophore size and SCOR may be a two-way interaction on time scales shorter than the response time of the ocean to an imbalance in the calcium carbonate budget (~10 000 years), through the biological pump and carbonate compensation feedbacks (Rost & Riebesell, 2004; Sarmiento & Gruber, 2006). On longer time scales, however, volcanic CO2 degassing and silicate weathering are generally considered to be the dominant controls on atmospheric CO2 (Berner & Kothavala, 2001; Zeebe & Caldeira, 2008). Thus, any CO2 response to a perturbation of the planktonic ecosystem and the subsequent compensation would generally not be resolved in our data (1/3 Myr and 10 Kyr maximum resolution for the SCOR and size records, respectively). If the observed relationship between coccolithophores and pCO2 involved a causal linkage on million-year time scales, then we would expect any directionality of coupling to emerge from a direct or indirect phytoplankton response to pCO2 forcing.

Here we tested this hypothesis by comparing the non-symmetric measure of directional information transfer between coccolithophore and paleoenvironmental records to a null distribution of values obtained using a large number of surrogates (Schreiber & Schmitz, 2000; Vejmelka & Paluš, 2008; Hannisdal, 2011a). If two environmental records are equally sensitive to a shared underlying forcing mechanism, then information transfer between the two records should be comparable in both directions. If, however, one record is more sensitive to a shared forcing, then it should be a better predictor of the other. Significance is determined by the scale-integrated information transfer, such that the total area under the information transfer curve must exceed that of the null distribution, regardless of the shape of the curve (Fig. S3). Finally, if information transfer in one or both directions is significant, then the difference in the area under the curves is used to test for significant directionality (asymmetry) of information flow.

We found significant information transfer from the pCO2 record to the coccolithophore SCOR but not vice versa, with a directionality exceeding that of the surrogates (Fig. 5a and b). The size record gave similar results (Fig. 5d and e). A sensitivity analysis was used to assess the effect of temporal sampling resolution (number and spacing of time bins) and the effect of uncertainty (noise) on the values within each time bin. We iteratively subsampled the records at randomly spaced time bins (500 iterations), varying the number of sampled time bins from 10 to 100 for the SCOR record, and from 16 to 160 for the size record. Within each time bin, we randomly drew values from Gaussian distributions defined by the reported midpoints and estimated standard error envelopes of each record (Fig. 2a–c; for the size record, we used the estimated standard error of the mean around the 95th percentile, and for pCO2, we used the ranges reported by Pagani et al., 2005, 2011). Then we calculated information transfer and determined the significance of the directional asymmetry. From this we obtained a frequency of detection (proportion of results that are significant), representing the expected statistical power of the test as a function of the number of time bins sampled, when noise was added to each proxy record. To evaluate how the statistical power is affected by the noise, we repeated the analysis with noise levels reduced by a factor of 10. Even if the statistical power necessarily declines with reduced sampling and added noise (Fig. 5c and f), the sensitivity analysis supports the conclusion of a directional information transfer from the pCO2 record to the coccolithophore SCOR and size records. Thus, the inferred relationships among the records reflect properties of the signals, not noise or temporal sampling.

Figure 5.

Testing for significant directionality of information transfer between pCO2 and coccolithophore Summed Common Species Occurrence Rate (SCOR) and size. (a) Information transfer between SCOR and pCO2 records quantified in both directions, without conditioning on other variables. Significance levels (gray) are 95th percentiles of 5000 calculations using amplitude-adjusted fast Fourier transform surrogate time series. Total information transfer (area under the curve) is significant only in one direction (pCO2 → SCOR). (b) The directionality of coupling, measured as the difference in area under the curves in each direction (solid line) is significantly greater than for the 5000 surrogates (histogram). (c) Statistical power (frequency of significant directionality) as a function of the number of randomly spaced time bins sampled. In each bin, data were randomly sampled within a noise envelope of two standard errors (dashed line) or an order of magnitude less (solid line). Stippled line represents average index of possible bias due to differences in non-stationarity between the two time series (∆ KPSS). (d–f) As in (a–c), but with size as the target variable.

Finally, to evaluate whether or not differences in non-stationarity could be a contributing factor to the directional asymmetry, we calculated the KPSS test score (1 if the null hypothesis of stationarity is rejected, 0 if not) over a range of lags from 0 to N/2, where N is the number of time bins. A bias index, ∆ KPSS, was defined as the sum of the absolute values of the difference between the two test score vectors, divided by N/2. Thus, a maximum ∆ KPSS value of 1 means that one time series is stationary at all lags, whereas the other time series is non-stationary at all lags, representing a maximum possible bias. Conversely, a ∆ KPSS value of 0 means that both series are either fully stationary, or non-stationary at the exact same lags, representing a minimum possible bias. Both sensitivity analyses yielded low ∆ KPSS values that converge on 0 with increasing temporal resolution (Fig. 5c and f), suggesting that the asymmetry of the directional information transfer is unlikely to be biased by differences in non-stationarity. These findings suggest that the pCO2 record captures a signal of one or more processes that have also influenced global changes in coccolithophore SCOR and reticulofenestrid cell size.


Recent studies proposing a link between climate and plankton evolution over the Cenozoic have targeted apparent trends in species richness, using correlations between patterns of richness or turnover and paleoenvironmental records as evidence for biotic response to abiotic forcing (Falkowski et al., 2004; Bown, 2005; Rabosky & Sorhannus, 2009; Cermeño, 2012). However, species richness is driven by rare species, and richness estimates from the deep-sea nannofossil record suffer from incomplete data, reworking, age model errors and taxonomic errors (Lazarus, 2011; Lloyd et al., 2012). Moreover, it is unclear what theoretical relationship species richness might have with productivity and carbon cycling – two important components of the plankton-climate coupling. Here we focus instead on the commonly occurring species, using the SCOR as a robust measure of the extent to which coccolithophores were globally common and widespread, thus offering a potential link to carbonate production and carbon cycling. Our results suggest that the SCOR holds some promise for reconstructing long-term ecological and evolutionary patterns in the deep-sea microfossil record.

Global changes in the size structure of planktonic assemblages may relate more directly to net production and the oceanic carbon pump than changes in species richness. Macroevolutionary trends in plankton body size have been attributed to long-term changes in ocean temperature, stratification, and nutrient supply (Schmidt et al., 2004; Finkel et al., 2007). Although these studies provide an explicit mechanistic hypothesis, their reported correlations between trended Cenozoic records do not constitute a proper test. Furthermore, the attribution of a universal abiotic driver of plankton size (Finkel et al., 2007) is not supported by observed size patterns in groups such as coccolithophores and radiolaria (Henderiks & Pagani, 2008; Lazarus et al., 2009). Our results suggest that the interaction between climate change and planktonic communities in deep time is taxon-dependent.

How could long-term changes in pCO2 directly or indirectly influence coccolithophores? Arguably, increased CO2 availability during the Eocene could have increased rates of photosynthesis, assuming ancient coccolithophores – like their modern descendants – did not possess an efficient CO2-concentrating mechanism (Rost et al., 2003), and given that the CO2/carboxylation specificity of Rubisco increases at higher CO2 levels (Giordano et al., 2005). In addition, coccolithophores may fix more carbon per unit of nutrient under high CO2 conditions (Riebesell et al., 2007), although it is not known whether such shifts in the C : N : P stoichiometry would be stable over the time scales we consider here. Since the early Oligocene, declining CO2 availability may thus have contributed to the extinction of large-celled coccolithophores (Henderiks & Pagani, 2008) and acted to reduce the SCOR of the Coccolithaceae, which previously thrived in the Eocene greenhouse world (Fig. 3b).

Our results depend in large measure on the alkenone-based pCO2 reconstruction, which builds on the stable carbon-isotope composition of alkenones and the total carbon-isotope fractionation that occurred during algal growth (εp37:2), under the assumption of diffusive uptake of CO2 (Pagani et al., 2005). In addition to CO2 concentrations, this method is sensitive to physiological factors, such as specific growth rate and cell geometry (Popp et al., 1998; Pagani et al., 2005; Henderiks & Pagani, 2008). The sensitivity to CO2 concentration is reduced under nutrient-replete conditions, and εp37:2 becomes lower (Riebesell et al., 2000). However, if algal growth rate were a dominant control on the carbon fractionation of alkenones, then Eocene alkenone producers must have had substantially lower growth rates in eutrophic waters than modern E. huxleyi in oligotrophic settings, which is highly unlikely. The changes in cell size of ancient alkenone producers, on the other hand, could imply even higher pCO2 just prior to early Oligocene glaciation ~34 Ma (Henderiks & Pagani, 2008).

If we relax the assumption of diffusive CO2 uptake, the observed decline in photosynthetic fractionation reflected in alkenone δ13C could be the result of changes in two additional physiological factors: (i) an increase in the ratio of bicarbonate to CO2 uptake in photosynthesis, and (ii) a reduction of cell leakiness, or CO2 efflux/influx ratio (Burkhardt et al., 1999). The latter is unlikely to have caused the declining trend, because the observed size decrease in the (alkenone-producing) reticulofenestrids (Fig. 2b) would imply an increase in the efflux/influx ratio. Conversely, a shift towards greater utilization of bicarbonate is consistent with a response to an externally imposed decrease in pCO2 causing an increase in the seawater carbonate ion concentration that is offset by a concomitant decrease in the Ca concentration, keeping the ocean CaCO3 saturation state relatively constant during the Cenozoic (Tyrrell & Zeebe, 2004). Furthermore, pCO2 could influence calcifying phytoplankton productivity indirectly via its control on the rate of terrestrial rock weathering on this time scale (Zeebe & Caldeira, 2008; Li et al., 2009), whereby a lowering of atmospheric pCO2 would reduce the weathering flux of dissolved constituents necessary for growth and calcification. Assessing the relative importance of direct and indirect mechanisms will require a better understanding of coccolithophore physiology, as well as improved Cenozoic records of ocean chemistry and nutrient fluxes.

More work is needed to understand how the coccolithophore-CO2 relationship behaves across different time scales. Although mechanisms cannot be extrapolated, our findings are consistent with inferred coccolith mass increase with rising atmospheric pCO2 over the past two centuries (Halloran et al., 2008; Iglesias-Rodriguez et al., 2008), and theoretical life-history modeling has been shown to predict an increase in coccolithophore size and calcification as an adaptive evolutionary response to ocean acidification (Irie et al., 2010). Low rates of sedimentation in the deep sea set a limit to the temporal resolution at which truly global signals can be resolved. On shorter time scales, any global link between plankton and atmospheric CO2 may be effectively concealed by more regional signals of oceanographic setting and biogeography.

In summary, we find a striking relationship between macroevolutionary changes in the coccolithophores and estimated changes in atmospheric pCO2 over a period of 50 million years, spanning one of Earth's major climatic (greenhouse-icehouse) transitions. Coccolithophores were globally more common and widespread, larger, and more heavily calcified in the pre-34 Ma greenhouse world, and declined in both ecological prominence and body size along with pCO2 during the Oligocene (34–23 Ma). Our results suggest that atmospheric pCO2 exerted an important long-term control on coccolithophores, either directly through its availability for photosynthesis or indirectly via weathering supply of resources for growth and calcification.


This research used samples provided by the Ocean Drilling Program (ODP). The ODP (now IODP) is sponsored by the US National Science Foundation and participating countries under management of the Joint Oceanographic Institutions (JOI), Inc. Comments from our anonymous reviewers helped improve the paper. JH is grateful for support from the Royal Swedish Academy of Sciences through a grant financed by the Knut and Alice Wallenberg foundation (KAW 2009.0287) and from the Norwegian Resesarch Council (project 197823/V40).