Southern Ocean phytoplankton turnover in response to stepwise Antarctic cooling over the past 15 million years

Significance

Based on data of unprecedented resolution, we show that phytoplankton (diatoms) in the Southern Ocean have experienced five major pulses of species extinction and origination over the past 15 My that were linked to large cooling transitions in southern high latitudes. Our findings suggest that phytoplankton communities around Antarctica have been robust to “baseline” glacial–interglacial climate variability but were sensitive to large-scale changes in mean climate state driven by a combination of long-period variations in orbital forcing and atmospheric carbon dioxide perturbations.

Abstract

It is not clear how Southern Ocean phytoplankton communities, which form the base of the marine food web and are a crucial element of the carbon cycle, respond to major environmental disturbance. Here, we use a new model ensemble reconstruction of diatom speciation and extinction rates to examine phytoplankton response to climate change in the southern high latitudes over the past 15 My. We identify five major episodes of species turnover (origination rate plus extinction rate) that were coincident with times of cooling in southern high-latitude climate, Antarctic ice sheet growth across the continental shelves, and associated seasonal sea-ice expansion across the Southern Ocean. We infer that past plankton turnover occurred when a warmer-than-present climate was terminated by a major period of glaciation that resulted in loss of open-ocean habitat south of the polar front, driving non-ice adapted diatoms to regional or global extinction. These findings suggest, therefore, that Southern Ocean phytoplankton communities tolerate “baseline” variability on glacial–interglacial timescales but are sensitive to large-scale changes in mean climate state driven by a combination of long-period variations in orbital forcing and atmospheric carbon dioxide perturbations.

In the face of warming oceans and changing seawater chemistry and circulation, there is growing evidence of biogeographic, community, and adaptive changes in the living marine microflora (1–3). Predicting how environmental change will influence phytoplankton communities, which account for ∼50% of global primary productivity, is hampered by the large spatial scale of the forcings on the biotic system and the long response time (2). Here, we use the history of postmid-Miocene (post-15 Ma) diatoms preserved in geological archives from the Southern Ocean and Antarctic margin to reveal long-term patterns and rates of phytoplankton turnover (origination rate plus extinction rate) and to constrain interpretations of macroevolutionary processes. These floras are of critical interest today given the expected sensitivity of high-latitude ecosystems to polar amplification of climate change. Globally, diatoms are the dominant living phytoplankton group, accounting for ∼20% of primary productivity, and their macroevolutionary history is linked to changes in climate (4). Diatom species are highly endemic south of the Antarctic Polar Front, which today forms an oceanographic and biogeographic boundary (5). Within this region, there are currently two distinct biomes: a specialized flora occupying the sea-ice zone (6) and a high-nutrient low-chlorophyll flora occupying the open ocean (7).

The climate history of the Antarctic and Southern Ocean over the past 15 My is one of stepwise cooling punctuated by transient warm periods. Highly variable ice sheets that characterized the mid-Miocene Climate Optimum (8, 9) subsequently expanded during the cooling associated with the Miocene Climate Transition ∼14 Ma (10, 11), at which time a large terrestrial ice sheet became a relatively permanent feature on the East Antarctic continent. Marine-based ice, grounded on bedrock below sea level in West and East Antarctica, continued to expand and contract through variable climates of the late Miocene and into the Pliocene (14–3 Ma) (12, 13), driving global sea-level fluctuations of between 20- to 30-m amplitude and up to 20 m above the present-day level (14). These marine ice sheets then stabilized as global climate cooled again in the late Pliocene and early Pleistocene (3–2.5 Ma) and perennial sea ice became a “permanent” feature in the Southern Ocean (15).

The response of Southern Ocean diatoms to these major changes has been, until now, poorly resolved. To realize the potential of their rich microfossil record, it is first necessary to mitigate pervasive biases that are inherent to paleontological records (Supporting Information). To achieve this, we used a quantitative biostratigraphic method, constrained optimization [CONOP (16)] (Materials and Methods and Supporting Information), applied to a regional database of fossil species occurrences in the Southern Ocean and Antarctic margin that was drawn from 34 drill cores (17) (Fig. 1 and Table S1). Data were restricted to regions south of the present-day Antarctic Polar Front, which represents an effective biogeographic barrier to the dispersal of modern assemblages. Because of uncertainties regarding the position of this front in the past (18–20), we used alternative CONOP models based on a conservative, more southerly dataset of 27 drill cores that are situated well south of the modern front, and a less conservative dataset that included an additional seven drill cores from regions immediately south of the front.

Fig. 1.

Map of Antarctica and the Southern Ocean, showing location of sites studied here. See Table S1 for details.

Table S1.

Drill cores used in this study and sources of data

Drill core	Latitude	Longitude	Water depth, m below seafloor	Location	Oceanography	Refs.	Age info*	Sample interval, m	Minimum sampled age, Ma	Maximum sampled age, Ma
DVDP 10	77° 34.72′ S	163° 30.7′ E	2.8 above sea level	McMurdo Sound	Shelf	58		117.76	0.00	5.40
DVDP 11	77° 35.4′ S	163° 24.67′ E	80 above sea level	McMurdo Sound	Shelf	58		290.70	0.00	5.36
ANDRILL 2A	77° 45.49′ S	165° 16.61′E	383.57	S McMurdo Sound	Shelf	59, 60	A	722.39	11.08	19.49
ANDRILL 1B	77° 53.36′ S	167° 5.64′ E	950	McMurdo Ice Shelf	Shelf	61–63	P A	589.70	0.07	6.39
CIROS 2	77° 41′ S	163° 32′ E	211	McMurdo Sound	Shelf	58	A	166.35	0.00	4.62
ODP 693	70° 49.89′ S	14° 34.41′ W	2,359	Weddell Sea	∼Summer sea-ice limit	64		275.72	0.00	23.58
ODP 1096	67° 34.01′ S	76° 57.81′ W	3,152.5	Antarctic Peninsula	∼Summer sea-ice limit	65		573.40	0.00	5.45
ODP 1101A	64° 22.33′ S	70° 15.67′ W	3,279.7	Antarctic Peninsula	∼Summer sea-ice limit	65	P	196.90	0.13	4.13
IODP U1359	64° 54.24′ S	143° 57.63′ E	3,010	Wilkes Land Margin	Winter sea-ice zone	Unpublished	P	612.83	0.26	15.52
IODP U1361A	64° 24.57′ S	143° 53.19′ E	3,466	Wilkes Land Margin	Winter sea-ice zone	Unpublished		367.27	2.08	13.65
IODP U1356A	63° 18.61′ S	135° 59.94′ E	4,003	Wilkes Land Abyssal Plain	Winter sea-ice zone	Unpublished		386.95	4.82	16.77
DSDP 274	68° 59.81′ S	173° 25.64′ E	3,326	NE of Cape Adare	Winter sea-ice zone	66		389.16	0.65	28.28
ODP 1165B	64° 22.8′ S	67° 13.1′ E	3,537.6	Prydz Bay	Winter sea-ice zone	67		49.69	0.26	5.32
ODP 695A	62° 23.48′ S	43° 27.1′ W	1,300	Scotia Ridge, N Weddell Sea	Winter sea-ice zone	64, 68–70	P	337.62	0.26	9.33
ODP 696	61° 50.95′ S	42° 55.98′ W	650	Scotia Ridge, N Weddell Sea	Winter sea-ice zone	64		509.65	3.83	15.87
ODP 690B	65° 9.63′ S	1° 12.3′ E	2,914	Maud Rise	Winter sea-ice zone	64, 68, 69	P	57.80	1.73	28.40
ODP 689B	64° 31.01′ S	3° 6′ E	2,080	Maud Rise	Winter sea-ice zone	64, 71	P	76.82	3.02	28.44
ODP 744	61° 34.66′ S	80° 35.46′ E	2,306.4	S Kerguelen Plateau	Winter sea-ice zone	72, 73	P	137.62	0.65	28.25
ODP 745B	59° 35.71′ S	85° 51.6′ E	4,082.5	Kerguelen Plateau	Winter sea-ice zone	72–74	P	210.63	0.00	8.52
ODP 746A	59° 34.12′ S	85° 52.09′ E	4,059.5	Kerguelen Plateau	Winter sea-ice zone	72–74		115.47	4.70	12.13
DSDP 269	61° 40.57′ S	140° 4.21′ E	4,285	SE Indian Ocean	∼Winter sea-ice limit	66		854.70	0.13	18.38
DSDP 267	59° 15.74′ S	104° 29.30′ E	4,564	SE Indian Ridge, Antarctic	∼Winter sea-ice limit	66		303.29	1.75	28.28
ODP 748B	58° 26.45′ S	78° 58.89′ E	1,290.9	Kerguelen Plateau	∼Winter sea-ice limit	75–77	P	115.18	0.00	29.00
ODP 751A	57° 43.56′ S	79° 48.89′ E	1,633.8	Kerguelen Plateau	∼Winter sea-ice limit	75, 76, 78		164.73	1.66	18.12
ODP 747A	54° 48.68′ S	76° 47.64′ E	1,695	Kerguelen Plateau	North of winter sea-ice limit	75, 76, 78	P	157.53	0.33	28.45
ODP 1138A	53° 33.1′ S	75° 58.5′ E	1,141.4	S Kerguelen Plateau	North of winter sea-ice limit	79	A	268.71	0.00	23.98
DSDP 266	56° 24.13′ S	110° 6.7′ E	4,173	SE Indian Ridge, Antarctic	North of winter sea-ice limit	66		366.53	0.46	15.70
DSDP 265	53° 32.45′ S	109° 56.74′ E	3,582	SE Indian Ridge, Antarctic	South of polar front	66		349.31	0.72	12.17
ODP 1094	53° 10.82′ S	5° 7.81′ E	2,807.3	S Atlantic Ocean	South of polar front	71, 80, 81	P	168.54	0.00	1.99
ODP 737A	50° 13.67′ S	73° 1.97′ E	564	N Kerguelen Plateau	South of polar front	72–74	P	296.80	2.75	13.12
ODP 736	49° 24.12′ S	71° 39.61′ E	629	N Kerguelen Plateau	∼Polar front	72, 73		359.80	0.13	4.88
ODP 1093	49° 58.59′ S	5° 51.93′ E	3,626.2	S Atlantic Ocean	∼Polar front	71, 80, 81	P	584.53	0.00	3.95
ODP 701C	51° 59.09′ S	23° 12.7′ W	4,636.7	S Atlantic Ocean	South of polar front	82, 83		194.64	0.26	9.23
ODP 699A	51° 32.54′ S	30° 40.62′ W	3,705.5	S Atlantic Ocean	∼Polar front	83		62.12	0.13	8.50

The lowest seven rows relate just to the 34-drill core dataset.

^*Age information: A, radiometric date on tephra; P, paleomagnetic polarity reversal data.

From vetted observations of local highest and lowest fossil occurrences in each drill core, CONOP produced model composite histories that were used to identify the order and timing of origination/immigration and regional extinction events for the common, abundant, and robust diatom species (Supporting Information). Conservatively, using three CONOP composites, we used model ensemble and bootstrapping procedures to capture uncertainties related to biogeographic constraints and varying assumptions about the nature of the fossil record, assumptions that are reflected in different CONOP compositing approaches (Supporting Information). Each model composite sequence of events produced by CONOP was dated independently to yield an average temporal resolution of ∼63,000 y (63 ky) over the study interval. This approach provides a “continuous” record of individual diatom appearance and disappearance events that is free from the biases and artificial discontinuities imposed by binning at the time resolution of zones and stages. The CONOP composites used here are given in Datasets S1–S3.

We do not discriminate true global speciation and extinction of endemic species from ecologically controlled immigration and local extinction of species that are also found north of the Antarctic Polar Front, a task that would require high-resolution, global biogeographic and biostratigraphic data (Materials and Methods, and see refs. 4 and 5). For the purposes of the present study, however, we are interested in the timing and patterns of species turnover in the Southern Ocean region and, although of considerable interest, it is not essential to discriminate turnover driven by evolution/extinction from turnover driven by biogeographic changes.

Results

Using the ensemble of age-calibrated composite histories of diatom first and last appearances, we calculated lineage-million-year origination, extinction, and turnover (origination + extinction) rates (21) for a smoothed 200-ky moving window at 50-ky time increments since 15 Ma (Fig. 2 and Fig. S1 A–C). For comparison, the average duration of a species in our dataset is ∼7 My. This new macroevolutionary history of diatoms is an order of magnitude more finely resolved than previously available records (cf. refs. 4 and 5) and reveals hitherto unknown patterns and pulses of phytoplankton species turnover, separated by intervals of moderate to low turnover. Results are apparently unbiased by effects of drill core coverage, sampling density, or the presence of widespread hiatuses in sedimentation (Supporting Information and Figs. S2 and S3).

Fig. 2.

Turnover pulses of diatoms in the Southern Ocean and Antarctic margin over the past 15 million years (Ma) compared with key paleoenvironmental proxies. (A) Species lineage-million-year (lmy) turnover rate. The bold line is the model ensemble and bootstrapped median, the dark region is the model ensemble (nonbootstrapped) ±1 SD uncertainty bound, and the pale region is the bootstrapped uncertainty bound (Supporting Information). Pink bars identify turnover pulses discussed here. (B) Benthic δ¹⁸O curve from ref. 42; raw data shown in gray, loess smooth in black. (C) Benthic δ¹⁸O curve from ref. 28; raw data shown in gray, loess smooth in black. (D) Opal accumulation rates (MAR_opal) at ODP Sites 1095 (black) and 1096 (blue), from ref. 43; this measure is taken as a proxy for sea-ice extent on the Antarctic margin such that declining opal accumulation indicates an expansion in sea ice. (E) Estimates of atmospheric pCO₂ based on various proxies, from ref. 44 (1, alkenone; 2, boron), ref. 45 (3, alkenone), and ref. 46 (4, boron); raw data in pink, loess smooth in red. (F) Benthic δ¹⁸O curve for the time interval 15–13 Ma, from ref. 11; raw data shown in gray, loess smooth in black. (G) Benthic δ¹³C curve from ref. 11; raw data shown in gray, loess smooth in black. (H) Estimates of atmospheric pCO₂ based on various proxies, from ref. 47 (1, boron), ref. 45 (2, alkenone), and ref. 48 (3, B/Ca and 4, alkenone).

Fig. S1.

(A) Species lineage-million-year (lmy) turnover rate (cf. Fig. 2A), which is calculated as the sum of extinction (B) and origination rates (C). (D) Proportion of endemic species in our data. Uncertainty intervals (colored regions) explained in Fig. 2A.

Fig. S2.

(A and B) Median extinction and origination rate curves from Fig. S1. (C) Number of drill cores spanning each part of the CONOP composite (34 drill core dataset). (D) Event support metric for all events in our data, this being the number of local observations for each event in each composite (ignoring observations at the top or bottom of a core). The mean event support is 8.6 local observations of each event. Also shown is a locally weighted polynomial regression.

Fig. S3.

Summed proportions of local FADs that are contracted upward and LADs that are contracted downward for intervals of high and low floral turnover. For the intervals of time indicated by the horizontal gray bars, these proportions were calculated from just those drill cores that are inferred to lack major hiatuses or condensed sections within each interval. Data were summed across all three composites. Vertical error bars indicate ± 1 binomial SE. See Potential Biases in the Origination and Extinction Rates for further explanation.

Here, we focus on five conspicuous peaks of turnover with rates greater than 0.5 (labeled turnover pulses A–E in Fig. 2A) and explore their relationship to known paleoclimatic events. The ages of these pulses are approximately: 14.65–14.45 Ma (A), 13.75–13.55 Ma (B), 4.90–4.40 Ma (C), 3.55–3.40 Ma (D), and 3.00–1.95 Ma (E).

Fig. 2 reveals that all of the turnover pulses coincide with positive excursions or maxima in benthic marine δ¹⁸O values that mark periods of global cooling and/or ice volume increase. Likewise, all except D occurred during episodes of atmospheric CO₂ decline inferred from geological proxy records, although we caution that the CO₂ geological proxy data are presently limited in temporal resolution and are subject to large assumptions and uncertainties (22). Turnover pulses A and B also coincide with two major middle Miocene positive excursions in δ¹³C, and pulses C to E align with periods of declining to low biosiliceous productivity on the Antarctic margin, represented by opal accumulation rate (MAR_opal). Taken together, these results suggest that the diatom turnover pulses identified here were related in some way to major global climatic and oceanographic changes, coupled to perturbations in the carbon cycle, which occurred during the middle Miocene to earliest Pleistocene.

In more detail, pulses A and B coincide with two major cooling steps within the Miocene Climate Transition, which marks progressive change from eccentricity-paced glacial–interglacial cycles to obliquity-paced cycles (11). During these cooling steps, bottom water temperatures decreased by around 3 °C and sea-surface temperatures decreased by up to about 7 °C (10). At the same time, sea level dropped by between 30 and 60 m (23), consistent with expansion of a terrestrial Antarctic Ice Sheet across the continental shelf and into marine basins in East and West Antarctica. The presence of grounded ice on the continental shelf is indicated in the ANDRILL-1B core and by regional seismic records (9) and supported by ice-sheet-climate models (8). Model simulations also indicate that sea-ice expansion is likely to have occurred in concert with advance of grounded ice sheets onto the shelf (24). Although definitive evidence of sea ice during turnover pulses A and B is lacking, changes in diatom and foraminifera assemblages in the ANDRILL-2A core have been used to suggest that sea ice appeared after 16 Ma and became more persistent after about 15 Ma (9, 25). (We note that certain key diatom taxa used to infer the presence of sea ice in ref. 9, notably Fragilariopsis truncata and Synedropsis cheethamii, have not been consistently identified in older floral lists from the Southern Ocean and failed to meet the threshold for inclusion in the present CONOP analyses.) In addition, probable sea-ice associated Miocene diatoms have been recorded from the Ross Sea region (26, 27).

Turnover pulses C and D likewise are associated with cooling episodes that culminated in deep-sea temperatures significantly colder than today (14). Evidence for the advance of grounded ice in the ANDRILL-1B core indicates that marine-based ice sheets in West Antarctica were more expansive than today, at least during the onset of pulse C (12). Whereas the onset of pulse D seems to predate significant cooling in Fig. 2C, in fact peak turnover at ∼3.4 Ma lies within the sustained, stepwise cooling trend that began at Marine Isotope Stage (MIS) MG8 at ∼3.55 Ma and persisted until MIS M2 at ∼3.3 Ma (28). The initiation of this cooling trend is coincident within uncertainty (gray region in Fig. 2A) with the onset of pulse D. Furthermore, pulse D is preceded by MIS Gi4 and Gi2, between 3.7 and 3.6 Ma, which are characterized by δ¹⁸O excursions with values that are similar to the Holocene, reflect major cooling and ice sheet advance, and may also have forced the onset of turnover. Importantly, the end of pulse D coincides with the final transition from subpolar to polar conditions on the Antarctic margin and termination of early- to mid-Pliocene warm conditions (15, 29).

Finally, the onset of turnover pulse E is associated with large-scale δ¹⁸O excursions (e.g., 3.1–3.0 Ma and 2.7–2.4 Ma) and minima in deep-sea temperatures (14), a trend toward increased Antarctic summer sea-ice extent and duration (15), and stabilization of marine margins of the East Antarctic Ice Sheet by ∼2.5 Ma (30).

Each of the cooling episodes described above was preceded by a warmer climate and, except in the case of turnover pulse E, followed by times of relative global warmth and high oceanic productivity, bottom-water temperature, and sea level (12, 14, 15, 23). Each successively younger “warm” interval, however, occurred on a cooler background climate state, perhaps controlled by variations in atmospheric CO₂.

Discussion

Results from this study suggest past, region-wide sensitivity of the diatom flora to large changes in ice extent in the pre-Pleistocene world. We propose the following model to explain this sensitivity (Fig. 3). High-latitude cooling and associated changes in atmospheric circulation, driven by global-scale processes, resulted in increased production of cold Antarctic surface waters, enhanced stratification in the Southern Ocean (31), and deepening of the pycnocline (32). These changes in ocean structure and dynamics at the Antarctic margin inhibited upwelling of warm deep water onto the continental shelf and promoted growth of perennial sea ice and expansion of winter sea ice across the Southern Ocean. We suggest that these episodes of major sea-ice expansion increased surface albedo, reduced Southern Ocean ventilation, and enhanced CO₂ sequestration to the deep ocean (33), which established a positive feedback on global climate that drove major expansion of Antarctica’s ice sheets across the continental shelves. We infer that, before each of the turnover pulses discussed here, there were year-round sea-ice-free conditions in the open Southern Ocean, with well-mixed, nutrient-rich surface waters, the “Permanent Open Ocean Zone” of ref. 34 (Fig. 3A). At such times, new species were evolving in stratified coastal waters that were sea-ice-covered in winter. During significant cooling events, when winter sea ice expanded beyond the continental margin and toward the polar front, and surface waters became stratified, the open ocean environment was greatly reduced (Fig. 3B). Given that ocean frontal boundaries can represent significant biogeographic barriers (5) and that phytoplankton diversity is governed by species-area effects (35), it seems likely that loss of open-ocean habitat during these cooling events drove extinction pulses in non-ice-adapted, Southern Ocean diatoms. Throughout the mid- to late Miocene and early Pliocene, large-magnitude cooling events seem to have been paced by long-period orbital cycles and as a result were transient in nature, and thus the forcing of species evolution and community adjustment by each new climatic regime was short-lived. However, long-term cooling, especially after 5 Ma, resulted in the incremental and stepwise evolution of endemic, ice-adapted diatoms that are a dominant component of the flora today (Supporting Information and Fig. S1D). Based on evidence available to date, the model proposed above is supported most strongly for turnover pulses C to E. We argue that the model would still have operated during times of lesser sea-ice extent—perhaps pulses A and B—when the entire biogeographic system may have been more sensitive to relatively brief but significant episodes of cryospheric expansion and/or was located closer to the Antarctic margin than today.

Fig. 3.

Schematic environmental reconstructions for the Antarctic continental shelf and Southern Ocean during intervals of (A) warmth and ice minimum and (B) peak cold and maximum ice extent; exact limits of Pliocene and Miocene sea ice unknown. Relatively rapid transition between these two end-member environmental states drives extinction/speciation of warm/cold-adapted phytoplankton (diatoms), causing major species turnover. Two schematic, representative diatom taxa are illustrated. Flow of Antarctic Circumpolar Current and westerly winds is out of the page. Flow of polar easterlies is into the page. Latitudinal position of peak flow is indicated by circles and relative strength of flow is indicated by bold (strongest) to dashed (weakest) lines. Ventilation of CO₂-enriched deep water is indicated by wavy arrows (thin line indicates reduced ventilation due to stratified surface water and sea-ice cover). AABW, Antarctic bottom water; ASW, Antarctic surface water; PF, polar front; UDW, upper deep water; WDW, warm deep water (red, relatively warmer; green, relatively cooler).

We cannot identify here specific ecological or physiological mechanisms that drove species turnover—whether presence of seasonal ice per se, or stratification, water temperature, nutrient supply, duration of growing season in open water, and so on—but all of these are strongly regulated by sea-ice extent in the Southern Ocean, and threshold effects in response to cryospheric expansion seem to have been the predominant factors in driving phytoplankton turnover in the Southern Ocean over the past 15 Ma. Also, as noted elsewhere, we cannot yet discriminate whether turnover was driven by regional or global extinction, and in situ evolution or immigration, or the precise relative timings of these components. We note, however, that our findings may be consistent with recent models of evolution in response to fluctuating but trending global temperature change, which reproduce patterns of bounded trait evolution (stasis) on short time frames with pulses of accelerated evolutionary divergence at time spans of longer than a million years (36).

The turnover pulses we document apparently were consequences of cooling, but our findings suggest that Southern Ocean phytoplankton communities are sensitive to large-scale changes in mean climate state driven by a combination of long-period variations in orbital forcing and atmospheric carbon dioxide perturbations. The potential for irreversible change in these phytoplankton communities in response to future, geologically abrupt warming in southern high latitudes—a consequence of polar amplification of global warming—remains an open question. This question will be addressed by a growing and diverse body of evidence that spans geological to ecological timescales, such as data on past, regional-, and planetary-scale ecological state changes (37), ecological niche models that are informed by both theory and empirical observations (3, 38), observations of adaptive evolution in phytoplankton (39, 40), and physiological experiments (41).

Materials and Methods

Our dataset was constructed from diatom occurrences recorded in 34 drill cores from the Antarctic margin and Southern Ocean south of the present-day Antarctic Polar Front (Fig. 1 and Table S1). All data were vetted before analysis, to eliminate records that may have been reworked, and species with fewer than three occurrences, with highly inconsistent or ambiguous stratigraphic ranges, or ranges that extend through the entire interval of interest. The final dataset contains 2,396 lowest and highest occurrences for 139 species that existed during the past 15 My.

To mitigate the effects of biases in the fossil record, we used the quantitative biostratigraphic method of constrained optimization, as implemented in the computer program CONOP9 (16), to generate best-fit, composite sequences or timelines of first and last appearances of species. These composites were age-calibrated using paleomagnetic polarity reversal data and one ⁴⁰Ar/³⁹Ar age on a volcanic tephra (Table S2). We calculated species-level, lineage-million-year origination and extinction rates (21), which are insensitive to the effects of overall diversity, and from these derived turnover as origination + extinction. To incorporate uncertainties relating to different assumptions about the nature of biases in the fossil record and the position of the polar front through time, our final analyses were based on composites from three different CONOP models (Table S2 and Datasets S1–S3) that were integrated using a model ensemble procedure; at the same time, bootstrapping was used to capture uncertainties inherent in the selection of species (details are given in Supporting Information).

Table S2.

Key features, diagnostics, and CONOP runtime parameters for the three composite sequences used here

Model	Relaxed hybrid (Dataset S1)	Strict hybrid (Dataset S2)	Strict hybrid (Dataset S3)
Properties of the data
No. of drill cores	34	34	27
No. of calibration ages	55	55	55
Total no. of species	151	151	139
No. of species ≤15 Ma	137	139	120
No. of FAD/LAD events ≤15 Ma	237	239	208
Total no. of FAD/LAD observations	2,836	2,836	2,344
Total no. of FAD/LAD observations ≤15 Ma	2,380	2,396	1,894
Properties of the solution
Total penalty, levels	8,215	8,962	6,256
Average penalty levels per section	242	264	232
Average penalty levels per FAD/LAD observation	2.9	3.2	2.7
Proportion of FAD/LAD observations ≤15 Ma that remain unmoved in CONOP solution, %	52	51	51
Proportion of FAD/LAD observations ≤15 Ma that remain unmoved in CONOP solution if core-end samples ignored, %	41	40	41
Average level spacing ≤15 Ma, My	0.061	0.060	0.068
Maximum level spacing ≤15 Ma, My	0.490	0.540	0.490
Minimum level spacing ≤15 Ma, My	0.010	0.010	0.010
Key runtime parameters
Coexistence constraint	Off	sensu stricto	sensu stricto
“Let contract”	On	sensu stricto	sensu stricto
Weight for contracting FAD (moving up)	5	12	12
Weight for contracting LAD (moving down)	2	4	4

About half (46%) of the species included here are apparently endemic to the Antarctic margin and Southern Ocean (Fig. S1D). For these taxa, our modeled times of first and last appearance represent the best available approximations to the true times of speciation and extinction. In contrast, the remaining species are known also from records north of the Antarctic Polar Front and are cosmopolitan, or restricted to southern polar to temperate or subtropical water masses, or bipolar in distribution. Their modeled times of first and last appearance may correspond, therefore, either to immigration and local extinction (e.g., ref. 5) or to true global evolution and extinction. Without recourse to high-resolution biogeographic and biostratigraphic data from north of the Antarctic Polar Front, we cannot discriminate between these possibilities but, as noted above, here we are focused on the timing and patterns of species turnover in the Southern Ocean region and its relationship to environmental drivers, and not on the wider biogeographic relationships of taxa.

Data

Our data were drawn from 34 drill cores obtained over 35 y of scientific drilling (Fig. 1 and Table S1). Our full dataset extends back as far as the Oligocene, although interpretations presented here and in the main text are restricted to the time interval 15 Ma to the present. All data were vetted thoroughly before analysis, as explained in ref. 17. We stress that during data selection, to be conservative, any records that may have been reworked from older into younger strata were eliminated, although our analytical methods also are designed to cater for this possibility.

Our analyses operate at the species level and subspecies and varieties have been treated as species-rank taxa. We included informally identified taxa or those in open nomenclature only in cases where the taxonomic concepts are well-documented and/or used consistently between multiple authors. All “immortal” taxa that range through the entire Neogene were excluded because they are not relevant to the present study of taxonomic turnover. We also excluded sparsely sampled species that have been recorded in fewer than three drill cores because their inferred ranges are likely to be unreliable. Finally, we eliminated from analysis 10 species with excessively inconsistent, ambiguous, or poorly constrained stratigraphic ranges (17). It must be remembered that the set of fossil taxa (even if every specimen observed were identifiable, recorded, and included in our models) will not represent the full set of living taxa that populated a region at a given time: Many of the smaller and more lightly silicified diatom species in any genus are not preserved in the fossil record, a bias that relates in particular to sea-ice taxa. Our analyses, therefore, pertain to the “nonimmortal,” common, abundant, and robust diatom species, which we infer will represent a meaningful proxy for the southern polar diatom flora in general.

Following filtering, our 27-drill-core dataset contains 139 species and 2,344 observed lowest or highest occurrences within the last ∼28 Ma, and the larger dataset contains 151 species and 2,836 range-end occurrences over the same period. For the interval of focus here, the past 15 Ma, our CONOP models include 120–139 species and are based on 1,894–2,396 observed occurrences (Table S2).

CONOP Analyses

In the following discussion, “FAD” denotes “first appearance datum” and refers to the first or lowest occurrence of a particular fossil taxon in time and stratigraphic depth. Likewise, “LAD” denotes “last appearance datum,” the last or stratigraphically highest occurrence of the taxon in question. “Local” observations of FADs and LADs are the levels of these events as recorded in a given drill core or stratigraphic section. In this simplified nomenclature, we explicitly equate stratigraphic position and geological age.

General discussions of the constrained optimization approach to quantitative biostratigraphy, as implemented in the computer program CONOP9 (16), can be found in refs. 49–51, and a recent short review is given in ref. 52. Many details of the methods as applied in the present study have been described in refs. 17 and 53 and are not repeated here.

In essence, however, because of the nature of geological sampling, local observations of the ordering of biostratigraphic events are always imperfect and, to varying degrees, contradictory. These problems arise because of incomplete and biased sampling (e.g., ref. 4), diachroneity in species distribution, geological reworking of fossils out of their strata of deposition (“stratigraphic leakage”), drilling disturbance, and inconsistent taxonomic practice. To overcome these problems and using many sets of local observations of stratigraphic superposition in sections and cores, CONOP generates a parsimonious, best-fit, composite sequence or timeline of events, where the events may include biostratigraphic, physical, or geochemical datums (52). During successive iterations and starting with an initially random sequence, CONOP estimates the quality of fit of a given composite permutation to the local observations using some objective penalty function. The choice of the objective function and penalty weighting regimes are guided by the user’s underlying assumptions about the nature of the stratigraphic record; different selections yield different CONOP models. Likewise, different event types can be governed by specific rules that reflect particular combinations of uncertainty. For example, the true stratigraphic height (as opposed to the sampled and observed height) of a biostratigraphic datum will be imprecisely located in any stratigraphic section and imprecisely dated, whereas a volcanic ash bed is likely to be precisely located and may be precisely dated. The “constraint” of constrained optimization refers to the requirements that the FAD must be placed before the LAD and the composite should honor all observed coexistences (overlapping ranges) of taxa, although this latter constraint can be relaxed.

Here, and building on the analyses of ref. 53, we used the robust and well-tested CONOP “hybrid range” model and a penalty function based on event levels. The hybrid range model permits but penalizes range contraction during the compositing process, to reflect the possibility that observed LADs may have been reworked upwards and/or observed FADs may have been bioturbated or caved downward. As an extension to the approach described in (53), we used both “strict” and “relaxed” hybrid models. The strict hybrid model enforces the coexistence constraint (discussed above) and prevents FADs or LADs from contracting entirely through their observed range. The relaxed hybrid model, in contrast, does not enforce the coexistence constraint and allows the composite range to move out of its observed local range, an extreme adjustment that implies large-scale displacement of all observed data and is unlikely to occur. The two variants of the hybrid range model also use different weighting schemes: The strict model penalizes upward adjustment of FADs 12:1 and downward adjustment of LADs 4:1, whereas for the relaxed model these ratios are 5:1 and 2:1 (Table S2). These weighting schemes span geologically plausible values that are consistent with our expectations regarding likelihood of reworking and they effectively discount the evidence of ∼10–15% of stratigraphically outlying event occurrences.

Each composite was age-calibrated using 117 local, highly vetted observations of 54 paleomagnetic polarity reversal datums from selected cores (Table S2), of which 45 events lie within the 15-My time interval of focus here. These reversal events were dated using the global geomagnetic polarity timescale (54). Polarity reversal observations were not used if they coincide with breaks in stratigraphy or core recovery, are based on poor or ambiguous remnant magnetization data, or rely entirely on diatom data for correlation to the global geomagnetic polarity timescale (to avoid circularity). In addition to the paleomagnetic data, one dated tephra bed, ash A2 in core ODP 1138A, was also used; other dated tephra beds were rejected because of their large error bars. All these dated events were included within the CONOP compositing process, and then age-calibration of a given composite was achieved via linear interpolation between the dated tie-points (17). On average, the spacing of constraining age tie-points in the composites is ∼0.31 My for the interval 0–15 Ma.

In our final age-calibrated models, the average spacing between diatom FAD/LAD event levels in the age-calibrated composites over the 0–15 Ma interval of interest, and ignoring the clustering of LADs of living taxa at 0 Ma, ranges from 0.060 to 0.068 My (Table S2).

Although we have tested many CONOP models, involving different selections of drill cores and different approaches to compositing (e.g., ref. 53), our final results are based on just three models: the strict hybrid range model for the 27-well-core dataset and both strict and relaxed hybrid range models for the 34-well-core dataset (see main text for discussion of well core selection). These models were selected because they satisfy biogeographic requirements and they span well-tested, robust, and geologically reasonable approaches to compositing using CONOP. In fact, however, the dominant patterns reported in the main text, of pulsed turnover in the middle Miocene and Pliocene and generally low rates of turnover in the late Miocene, are consistent across alternative models examined individually (results not presented here). In total, we estimate that several billion candidate composite solutions were tested and ranked over the course of the selected CONOP runs.

Note that, unlike the convention used in previous studies of this sort, during subsequent calculation of origination and extinction rates, we assumed that the FAD of a taxon lies at the sample level below its placement in the CONOP composite and, likewise, the LAD lies at the level above its placement. This very minor adjustment reflects the fact that first and last occurrences of the composite represent positive evidence for the presence of a given taxon, and its FAD and LAD must lie beyond these points. This adjustment was not applied to events at the limits of the time series, which would imply extrapolation of the composite. Also, we did not place events midway between their recorded range limit and the level beyond because this would have required creation of synthetic event levels and implied unrealistic temporal resolution.

Generating Model Ensemble Rates

Rather than compare patterns of origination and extinction based on individual CONOP composites, and unlike previous studies, we have taken a model ensemble approach based on resampling, an approach that captures overall uncertainties related to variations between the CONOP models. At the same time, we incorporated uncertainties related to both age calibrations and the selection of species within the composites. Our model ensemble procedure used the following sequence of steps and, along with other analyses described below, was undertaken in the R language and environment for statistical computing (55).

• i)For a given iteration, one of the model composites was chosen at random.
• ii)Event ages were sampled randomly from levels lying within their calibrated age spans, where these spans reflect uncertainties in the ages of events used to calibrate the composite. During this process, FAD-before-LAD constraints for a given species were enforced. In reality, because most events have a unique age, this step was only relevant in a handful of cases.
• iii)For the resulting timeline of FAD and LAD ages, we imposed a moving-window time bin for the calculation of origination and extinction rates. For the results presented here, this moving window comprised five composite levels, centered on the third level, with an average window duration of ∼0.3 My. Although we experimented with moving windows based on duration in millions of years, we prefer to operate on composite levels because this implicitly factors in the effect of uneven spacing of data.
• iv)Within each moving-window time bin, we calculated lineage-million-year origination and extinction rates (21), these being the number of FADs and LADs, respectively, divided by the sum of species lineage segments lying within the window. These rate metrics are insensitive to the effects of varying overall diversity and are scaled to the number of lineages at risk and the time they are at risk (56). From these two rates, we calculated turnover rate as origination + extinction (Fig. S1).
• v)Steps 1–4 were repeated 1,000 times.
• vi)To impose a uniform timescale across results from the different model composites, we used a centered 0.2-My fixed-duration moving window, moving by 0.05-My increments, and calculated the ±1 SD values (i.e., the ∼16th and 84th percentiles) for the results of step v. These are the model ensemble (nonbootstrapped) uncertainty bounds plotted in Fig. 2.
• vii)Finally, steps i–vi were repeated with the addition of bootstrapping of species between steps i and ii; in other words, for a given, randomly selected composite with n species, n species were resampled with replacement. This operation provided a means of modeling uncertainties that are inherent within the chosen cohort of species. From these results, we calculated the median origination and extinction rates and the bootstrapped ±1 SD uncertainty bounds plotted in Fig. 2 (these latter uncertainty bounds are in fact the model ensemble + bootstrapped uncertainties).

Taken together, the incorporation of uncertainties from different models, from the age-calibrated uncertainties, and from the selection of species, is highly conservative. We believe that the resulting, highly resolved, timeline of diatom first and last appearances in the Southern Ocean is extremely robust. The two sets of uncertainty bounds shown in Fig. 2 represent more and less conservative interpretations and the medians represent our best estimates of the “true” rate time series.

Potential Biases in the Origination and Extinction Rates

It is important to consider whether some of the structure in the origination and extinction rate curves may be influenced by sampling biases related to the density and distribution of input data, even though the CONOP method and our model ensemble procedure are designed to overcome major biases. One obvious potential bias may result from edge effects related to the limits of drill cores: Perhaps artificial biostratigraphic range truncations clustered at the lower or upper limit of cores could result in spurious peaks of apparent origination or extinction? Because of this problem, CONOP ignores FADs that coincide with the base of a section and LADs that coincide with the top of a section when determining the composite range of a taxon; thus, the impact of edge effects should be minimized. In any case, to test this, we calculated the number of drill cores in our dataset that span each interval of time, using our CONOP-derived, revised age models for each core (Table S1), to identify significant steps in core coverage (Fig. S2C). Visually, there is no obvious correlation between variations in core coverage and extinction or origination rate (compare Fig. S2 A–C) and Spearman’s rank-order correlation coefficients for detrended (first-differenced) time series are low and nonsignificant: number of sections vs. extinction, ρ = 0.093, P = 0.109; number of sections vs. origination, ρ = −0.106, P = 0.068. Thus, there is no evidence to suggest that edge effects related to the limits of drill cores have caused spurious peaks in extinction or origination rates.

To test this in more detail, we examined whether the density of sampling of events may have influenced our measured taxic rates. Fig. S2D shows the distribution of FADs and LADs in the composite against an event-support metric, which is simply the number of local observations of a given event in each of the three composites. Overall, on average, each event in each composite has 8.6 local observations, excluding those FADs and LADs that occur at the bottom and top of a core, respectively (explained above). A locally weighted polynomial regression (loess) smooth has been fitted to the event-support data and interpolated values extracted to match our time series of origination and extinction rates to calculate correlations. Comparison of Fig. S2 A and B with Fig. S2D suggests that there is no apparent relationship between the event-support curve and our taxic rates. Spearman’s rank-order correlation coefficients for the detrended time series are as follows: event support vs. extinction, ρ = 0.006, P = 0.917; event support vs. origination, ρ = −0.082, P = 0.161. From these results we infer that major patterns in our time series of origination and extinction rates are not being driven to any significant extent by uneven sampling in the input data. [Note that apparent discrepancies between the number of drill cores and event support in the youngest ∼0. 4 Ma (Fig. S2 C and D) reflect the presence of unconformities in the youngest parts of some drill cores, which reduces their effective age span even though some biostratigraphic events are recorded within these condensed intervals.]

Because our approach allows for range contraction, the possibility remains that region-wide hiatuses or variations in sedimentation rate could potentially cause spurious volatility in apparent turnover rates if the condensed sections outnumber the complete sections (57). Specifically, a dominance of condensed sections may cause FADs in the complete drill cores to contract upward to the top of the condensed interval, creating a spurious spike in FADs at that level. Likewise, LADs may contract downward to the base of the condensed interval, creating a spike of LADs. We consider this an unlikely scenario given the wide latitudinal and geographic spread of our drill cores and their range in depositional settings, and consequently the low probability that any condensed interval will be strongly expressed across the majority of sections. Furthermore, our weighting schemes penalize range contractions, meaning that condensed sections would have to greatly outnumber complete sections before they would drive the composite solution. To test for the presence of artifacts related to regional stratigraphic condensation, however, we compiled statistics on the proportion of local observations of FADs and LADs that are contracted into the composite. We compared these statistics for just the drill cores that are inferred to be essentially complete through intervals of high and low floral turnover, where estimates of completeness were based on the original age models and our CONOP-derived, revised age models. We find no evidence to suggest that range contractions in the most complete drill cores are more prevalent during times of high turnover than during times of low turnover, suggesting that widespread condensation has not resulted in spurious peaks in origination and extinction rates. Whereas there are significant variations in proportions of local event contractions, these proportions are generally similar in adjacent intervals of low and high turnover (Fig. S3). Overall, considering just the data plotted in Fig. S3, on average there are 10.3% range contractions across intervals of high turnover and 8.6% contractions across intervals of low turnover; these two proportions are not statistically significantly different (Fisher’s exact test, P = 0.427). Thus, there is no evidence to suggest that our turnover pulses are artifacts of stratigraphic incompleteness.