Philippa Kaur

Epithemia isolation and cultureThe Epithemia cells were isolated from 0.5 L of seawater collected from depths of 25, 75, and 100 m in the North Pacific Subtropical Gyre (22°45′ N, 158°00′ W). Seawater was collected during the near-monthly Hawaii Ocean Time-series (HOT) expeditions to the long-term monitoring site Station ALOHA (water depth ca. 4800 m) in October 2014 (HOT cruise #266) and February–July 2019 (HOT cruises #310–313). Serial dilution (unialgal strains UHM3202, UHM3203, UHM3204) or micropipette isolation of single cells (clonal strains UHM3200, UHM3201, UHM3210, UHM3211) were used to establish the Epithemia cultures, which were grown in a seawater-based, low-nitrogen medium. Filtered (0.2 µm) and autoclaved, undiluted Station ALOHA seawater was amended with 2 μM EDTA, 50 nM ferric ammonium citrate, 7.5 μM phosphoric acid, trace metals (100 nM MnSO4, 10 nM ZnCl2, 10 nM Na2MoO4, 1 nM CoCl2, 1 nM NiCl2, 1 nM Na2SeO3), vitamins (50 μg/L inositol, 10 μg/L calcium pantothenate, 10 μg/L thiamin, 5 μg/L pyridoxine HCl, 5 μg/L nicotinic acid, 0.5 μg/L para-aminobenzoic acid, 0.1 μg/L folic acid, 0.05 μg/L biotin, 0.05 μg/L vitamin B12), and 106 μM Na2SiO3. Although not tested here, simpler formulations of diazotroph media such as PMP40 or RMP41 may also be suitable for growing Epithemia, when made with 100% seawater and adding Na2SiO3. The cultures were subsequently incubated at 24 °C on a 12:12 h light:dark cycle with 50–100 μmol quanta m−2 s−1 using cool white fluorescent bulbs. All E. pelagica and E. catenata symbioses were stable under these medium and incubation conditions. E. pelagica was successfully isolated from at least one of the three depths that were targeted during each sampling occasion.Morphological observationsEpithemia living and fixed cells were imaged by light and epifluorescence microscopy using a Nikon Eclipse 90i microscope at 40×–60× magnification. Diatom cell sizes were determined using >60 live, exponentially growing cells, imaged in either valve view (E. pelagica) or girdle view (E. catenata). Endosymbiont (spheroid body) cell sizes were averaged from DNA-stained cells for E. pelagica UHM3200 (n = 78) and E. catenata UHM3210 (n = 91), imaged by epifluorescence microscopy after preparing samples as follows: Epithemia cells were fixed in 4% glutaraldehyde for 30 min, pelleted at 1000 × g for 1 min, the supernatant was exchanged with 0.5% Triton X-100 (in autoclaved filtered seawater), samples were incubated for 10 min with gentle agitation, cells were then pelleted at 4000 × g for 1 min, supernatant was exchanged with autoclaved filtered seawater and fixed in 4% glutaraldehyde, and samples were stained with 1× final concentration of SYBR Gold nucleic acid stain (Invitrogen, cat. # S11494) for 2 h. For routine observations of endosymbionts (e.g., determining presence/absence and number per host cell), osmotic shock was used to disrupt the cell contents of diatom host cells and improve visualization of the endosymbionts. This was achieved by gently pelleting cells and exchanging the medium with either ultrapure water or 2–3 M NaCl solution, followed by immediate observation. While this is a simple technique for detecting and visualizing endosymbionts (Fig. 1c, f), it does not accurately represent the natural location of endosymbionts within the host cells, as seen when compared to fixed cell preparations for epifluorescence microscopy (Fig. 1n, o). To assess the presence of fluorescent photopigments in endosymbiont cells, live host cells were pelleted at 4000 × g for 5 min and crushed using a microcentrifuge tube pestle (SP Bel-Art, cat. # F19923-0000) to release the endosymbionts. The crushed pellet was resuspended in 75% glycerol containing live Synechococcus WH7803 cells (positive control for fluorescence), and samples were observed by epifluorescence microscopy using filter cubes appropriate for observing phycoerythrin (EX: 551/10, BS: 560, EM: 595/30) and chlorophyll (EX: 480/30, BS: 505, EM: 600LP).The loss of endosymbionts from Epithemia cultures (UHM3200 and UHM3210) was observed after propagating cells for four months in nitrogen-replete medium (K)18, where approximately 5–10% of the culture was transferred to fresh medium about every two weeks. Observations were only made at the end of the four-month period. Endosymbionts were not observed growing freely in these cultures, and the absence of endosymbionts within host cells was confirmed by the failure to observe spheroid bodies by light microscopy after osmotic shock of the diatoms, as well as a failure to amplify the endosymbiont SSU (16S rRNA) and nifH genes from cellular DNA extracts. PCR reactions were performed in parallel with DNA extracts from control cultures (grown in low-nitrogen medium), using the same template DNA amount (10 ng) and PCR conditions (see methods for Marker gene sequencing and phylogenetics).Ultrastructural observations by electron microscopy (EM) were conducted for E. pelagica UHM3200 and E. catenata UHM3210. EM preparations of diatoms typically involve the oxidative removal of organic matter to uncover the fine details of frustule ultrastructure. However, in the case of E. catenata, oxidatively cleaned cells lacked structural integrity, leading to collapsed frustules when dried and viewed by scanning EM (SEM). For this reason, both species were prepared for SEM with and without (Fig. 1a, d) the oxidative removal of organic matter, and cleaned E. catenata frustules were further analyzed by transmission EM (TEM). To remove organic matter, 100 mL of exponentially growing culture was pelleted by centrifugation at 1000 × g for 10 min and resuspended in 30% H2O2. Cells were boiled in H2O2 for 1–2 h, followed by rinsing cells six times in ultrapure water by sequential centrifugation at 1000 × g for 10 min and resuspension of cell pellets. Suspensions of the cleaned cells were dried on aluminum foil and mounted on aluminum stubs with double-sided copper tape. For some E. catenata SEM preparations, the cleaned frustules were dehydrated in an ethanol dilution series and exchanged into hexamethyldisilazane (HMDS) prior to drying on aluminum foil; this was to minimize the collapse of frustules resulting from drying. To prepare cells with organic matter intact, 25 mL of exponentially growing culture was mixed with an equal volume of fixative solution (5% glutaraldehyde, 0.2 M sodium cacodylate pH 7.2, 0.35 M sucrose, 10 mM CaCl2) and incubated overnight at 4 °C. Cells were gently filtered onto a 13 mm diameter 1.2 μm pore size polycarbonate membrane filter (Isopore, Millipore Sigma), washed with 0.1 M sodium cacodylate buffer (pH 7.4, 0.35 M sucrose), fixed with 1% osmium tetroxide in 0.1 M sodium cacodylate (pH 7.4), dehydrated in a graded ethanol series, and critical point dried. Filters were mounted on aluminum stubs with double-sided conductive carbon tape. All SEM stubs were sputter coated with Au/Pd, prior to observing on a Hitachi S-4800 field emission scanning electron microscope at the University of Hawai’i at Mānoa (UHM) Biological Electron Microscope Facility (BEMF). Cleaned E. catenata cells were prepared for TEM by drying a drop of sample on a formvar/carbon-coated grid and observing on a Hitachi HT7700 transmission electron microscope at UHM BEMF.Additional light microscopy of hydrogen-peroxide cleaned frustules was conducted for E. pelagica UHM3201 and E. catenata UHM3210. Samples were mounted in Naphrax (PhycoTech, Inc., cat. # P-Naphrax200) and observed at 100× using an Olympus BX41 Photomicroscope (Olympus America Inc., Center Valley, Pennsylvania) with differential interference contrast optics and an Olympus SC30 Digital Camera at California State University San Marcos.A key to the strains used in each micrograph is provided in Supplementary Table 2.Marker gene sequencing and phylogeneticsFor each Epithemia strain, 25–50 mL of culture was pelleted at 4000 × g for 10 min, and DNA was extracted from the pellet using the ZymoBIOMICS DNA Miniprep Kit (Zymo Research, cat. # D4300). Marker genes were amplified with the Expand High Fidelity PCR System (Roche, cat. # 4743733001), using conditions previously described for genes SSU encoding 18S rRNA (Euk328f/Euk329r)42, LSU encoding 28S rRNA (D1R/D2C)43, rbcL (rbcL66+/dp7−)44,45, psbC (psbC+/psbC−)44, and cob (Cob1f/Cob2r)21. For the endosymbionts, a partial sequence for the SSU (16S rRNA) gene was amplified using a primer set targeting unicellular cyanobacterial diazotrophs, CYA359F/Nitro821R46,47, and the nifH gene was amplified using new primers specific to the nifH of Cyanothece-like organisms, ESB-nifH-F (5′-TACGGAAAAGGCGGTATCGG-3′) and ESB-nifH-R (5′-CACCACCAAGRATACCGAAGTC-3′), with a 55 °C annealing temperature and 75 s extension time. All primers were synthesized by Integrated DNA Technologies (IDT). Amplified products were cloned and transformed into E. coli using the TOPO TA Cloning Kit for Sequencing (Invitrogen, cat. # K457501), and plated colonies were picked and grown in Circlegrow medium (MP Biomedicals, cat. # 113000132). Plasmids were extracted with the Zyppy Plasmid Miniprep kit (Zymo Research, cat. # D4019) and sequenced from the M13 vector primers using Sanger technology at GENEWIZ (South Plainfield, NJ). For the diatom SSU (18S rRNA) gene, sequencing reactions were also performed using the 502f and 1174r primers48.Phylogenetic trees (Fig. 2) were inferred using concatenated alignments for both diatom host genes (SSU encoding 18S rRNA, psbC, rbcL) and endosymbiont genes (SSU encoding 16S rRNA, nifH). For each gene, nucleotide sequences were aligned using MAFFT v7.45349 (L-INS-i method), and sites with gaps or missing data were removed. An appropriate nucleotide substitution model was selected for each gene alignment using jModelTest v2.1.1050. Bayesian majority consensus trees were inferred from the concatenated alignments using MrBayes v3.2.751 with two runs of 4–8 chains, until the average standard deviation of split frequencies dropped below 0.01. Maximum likelihood bootstrap values were generated for the Bayesian tree using RAxML v8.2.1252, implemented with 1000 iterations of rapid bootstrapping. To further analyze the phylogenetic position of the new Epithemia species in the broader context of Surirellales and Rhopalodiales diatoms, individual gene trees (SSU encoding 18S rRNA, LSU, rbcL, psbC, and cob; Supplementary Figs. 13–19) were constructed from sequences aligned using MAFFT (automatic detection method) and trimmed using trimAl v1.253 (gappyout method). rRNA gene phylogenies were also inferred using sequences aligned according to the global SILVA alignment for SSU and LSU genes using SINA54, which were either left untrimmed in the case of the LSU gene or trimmed to remove highly variable positions (SINA’s “012345” positional variability filter) and gappy positions (trimAL v1.2, gappyout method) in the case of the SSU gene. These trimming strategies were selected based on their ability to maximize the monophyly of the previously described Rhopalodiales clade and minimize the separation of known conspecific strains, such as the strains of E. pelagica described here. All gene phylogenies were inferred using the Bayesian methods described above. To investigate the level of support for constrained tree topologies placing E. catenata within or outside of the genus Epithemia and family Rhopalodiaceae, SH55 and AU56 statistical tests were performed in IQ-TREE 257 (implementing ModelFinder58) using all alignments from the individual gene trees (Supplementary Table 3).Given E. catenata’s unusual morphology, test trees were inferred with the inclusion of diatom sequences from orders Bacillariales (Nitzschia, Pseudo-nitzschia), Cymbellales (Didymosphenia), Naviculales (Amphiprora, Navicula, Pinnularia), and Thalassiophysales (Amphora, Halamphora, Thalassiophysa); however, E. catenata was consistently placed within Rhopalodiales, and these trees were not pursued further.An additional nifH phylogeny was constructed using all environmental sequences from NCBI’s non-redundant nucleotide (nt) database >300 bp and sharing >95% nucleotide sequence identity with EpSB and EcSB nifH sequences (Supplementary Fig. 23), including 51 environmental sequences from prior studies investigating marine diazotrophs34,59,60,61,62,63,64,65,66. Environmental nifH sequences were aligned to the previously generated nifH sequence alignment using MAFFT (automatic method detection and addfragments options), and the best-scoring maximum likelihood phylogeny was inferred using RAxML with 1000 iterations of rapid bootstrapping. NCBI accession numbers for all tree sequences are in the Source Data file.Analysis of Epithemia endosymbiont nifH sequences in environmental datasetsNucleotide sequences for EpSB and EcSB nifH were queried against NCBI’s non-redundant nucleotide (nt) database using webBLAST67 (megablast; https://blast.ncbi.nlm.nih.gov/) and SRA databases for nifH amplicon sequencing projects from the marine environment using the SRA Toolkit68 (dc-megablast, with database validation using vdb-validate; https://github.com/ncbi/sra-tools). Database hits with 98–100% nucleotide identity over an alignment of the entire subject sequence (BLAST alignment length = subject sequence length) were identified, and the associated sample’s latitude and longitude coordinates (where available) were mapped. Coordinates were also mapped for metagenome and metatranscriptome samples containing matches to unigene MATOU-v1_93255274 from the Marine Atlas of Tara Oceans Unigenes69, a unigene that shares 100% identity over the entire length of the EpSB UHM3202 nifH sequence and >99.4% identity with all other EpSB nifH sequences.The presence of EpSB and EcSB nifH sequences was examined in metagenomes prepared from sinking particles collected at 4000 m depth at Station ALOHA27,28. The sinking particles were collected during intervals of 12, 10, and 8 days during 2014, 2015, and 2016, respectively, using a McLane sediment trap equipped with a 21-sample bottle carousel. The presence of EpSB and EcSB nifH sequences in the metagenomes was assessed by blastn70, after first removing low quality bases from metagenomic reads using Trimmomatic v0.3971 (parameters: LEADING:20 TRAILING:20 MINLEN:100). For each sediment trap metagenome, the total number of reads matching EpSB or EcSB nifH nucleotide sequences with 100% identity were tallied and normalized to the total number of reads in the database. Only EpSB-matching reads were detected in this analysis.Quantitative PCRSpecific PCR primers were designed targeting a 102 bp region of E. pelagica’s LSU gene (Epel-LSU-F, 5′-GAAACCAGTGCAAGCCAAC-3′; Epel-LSU-R, 5′-AGGCCATTATCATCCCTTGTC-3′) and an 85 bp region EpSB’s nifH gene (EpSB-nifH-F, 5′-CACACTAAAGCACAAACTACC-3′; EpSB-nifH-R, 5′-CAAGTAGTACTTCGTCTAGCTC-3′) and were synthesized by IDT. Gene copy concentrations were quantified for Station ALOHA water samples (~2 L) collected by Niskin bottles at 5, 25, 45, 75, 100, 125, 150, and 175 m on January 16 and July 1 (except 5 m), 2014, during HOT cruises #259 and #264. Samples were filtered onto 25 mm diameter, 0.02 μm pore size aluminum oxide filters (Anotop; Whatman, cat. # WHA68092102) and stored at −80 °C until extracting DNA using the MasterPure Complete DNA and RNA Purification Kit (Epicentre, cat. # MC85200) according to Mueller et al.72. Briefly, a 3-mL syringe filled with 1 mL of tissue and cell lysis solution (MasterPure) containing 100 μg mL−1 proteinase K was attached to the outlet of the filter, and the filter inlet was sealed with a second 3-mL syringe. The lysis solution was pulled halfway through to saturate the filter membrane, and the entire assembly was incubated at 65 °C for 15 min while attached to a rotisserie in a hybridization oven rotating at ca. 16 rpm. The lysis buffer was then drawn fully into the inlet syringe, transferred to a microcentrifuge tube, and placed on ice. The remaining steps for protein precipitation and removal and nucleic acid precipitation were carried out following the manufacturer’s instructions. For each sample, DNA was resuspended in a final volume of 100 μL. Quantitative PCR (qPCR) was performed using the PowerTrack SYBR Green Master Mix system (Applied Biosystems, cat. # A46109) and run on an Eppendorf Mastercycler epgradient S realplex2 real-time PCR machine. Reactions (20 µL total volume) were prepared according to the manufacturer’s protocol, containing 500 nM of each primer. Sample reactions (four replicates) contained 2 μL of environmental DNA extract (24–76 ng DNA), while standards (three replicates) contained 2 μL of gBlocks Gene Fragments (IDT) that were prepared at 1, 2, 3, 4, 5, and 6 log gene copies/μL. The gBlocks Gene Fragments were 500 bp in length and encompassed the entire E. pelagica UHM3201 LSU sequence and positions 1–500 of the EpSB UHM3201 nifH sequence, respectively. The main cycling conditions consisted of an initial denaturation and enzyme activation step of 95 °C for 2 min, followed by 40 cycles of 95 °C for 5 s and 57 °C or 55 °C for 30 s for the LSU and nifH genes, respectively. Melting curves were analyzed to verify the specificity of the amplifications, and reactions containing Epithemia catenata DNA extract were included as negative controls. Reaction efficiencies were 104.23% and 95.15% for the LSU and nifH genes, respectively. The limit of detection for these assays was not empirically determined. gBlocks sequences, qPCR threshold cycle values, and conversion equations are provided in the Source Data file.Physiology experimentsThe daily patterns of N2 fixation were quantified for E. pelagica UHM3200 and E. catenata UHM3210 using two techniques: acetylene (C2H2) reduction to ethylene (C2H4) and argon induced dihydrogen (H2) production (AIHP). Both analyses were conducted using a gaseous flow-through system that quantified the relevant trace gas on the sample outlet line with a temporal resolution of 10 min73. To conduct the measurements, a 10-mL subsample of each Epithemia culture was placed in a 20-mL borosilicate vial and closed using gas-tight rubber stoppers and crimp seals. Separate bottles were used for H2 production and C2H2 reduction. During the experimental period, the temperature was maintained at 25 ± 0.2 °C using a benchtop incubator (Incu-Shaker; Benchmark Scientific) and light exposure was 200 μmol photons m−2 s−1 at wavelengths of 380–780 nm with a 12:12 h square light:dark cycle (Prime HD+; Aqua Illumination). To conduct the AIHP method, the sample vial containing the culture was flushed with a high purity gas mixture consisting of argon (makeup gas; 80%), oxygen (20%), and carbon dioxide (0.04%). In the absence of N2, all of the electrons that would have been used to reduce N2 to NH3 are diverted to H2 production, thereby providing a measure of Total Nitrogenase Activity (TNA). The C2H2 reduction assay also represents a measure of TNA. Our analytical set-up introduced C2H2 at a 1% addition (vol/vol) to the high purity air with a total flow rate (13 mL min−1) identical to the AIHP method. The gas emissions were analyzed using separate reductive trace gas analyzers that were optimized for the quantification of H2 and C2H4. To verify the observed daily patterns in N2 fixation, 15N2 assimilation measurements were conducted on triplicate samples of Epithemia cultures at targeted time points. Five milliliters of 15N-enriched seawater was added to the subsamples, which were subsequently crimp sealed and incubated for a 2 h period with the same light and temperature conditions as the daily gas measurements. At the end of the incubation, the contents of each vial were filtered onto a pre-combusted glass fiber filter. The concentration and isotopic composition (δ15N) of particulate nitrogen for incubated and non-incubated (i.e., natural abundance) samples was measured using an elemental analyzer/isotope ratio mass spectrometer (Carlo-Erba EA NC2500 coupled with a ThermoFinnigan Delta Plus XP). For each of the described analyses, cell-specific rates were calculated based on the average of triplicate cell concentration measurements, obtained from cell samples preserved at 4 °C with Lugol’s iodine solution and quantified within a week using a Sedgwick-Rafter counting chamber (Electron Microscopy Sciences, cat. # 68050-52). All rate measurement data is provided in the Source Data file.Reporting summaryFurther information on research design is available in the Nature Research Reporting Summary linked to this article. More

A workflow for the calculation of CSI is presented in Fig. 1c. For all the analyses, we used the R v. 4.0.3 software environment20 implemented in RStudio v. 1.4.1103. The scripts used for each methodological step are available at the Figshare repository21. After data download from primary sources (PaleoClim and WorldClim), specifically for the CSI-future map set we performed an initial step aimed to obtain individual bioclimatic variables for each future time period for the four SSPs (Fig. 1b). To achieve this, the median values of nine GCMs were calculated in functions compiled in raster R package22 for each individual bioclimatic variable (see a few exceptions of number of GCMs used in Table 2).Table 2 General circulation models (GCM) used to construct the future map sets.Full size tableThe standard deviation (SD) was estimated as a measure of the amount of variation or dispersion along time series, from which the resulting output maps showed the places where climate conditions remained constant or variable across the temporal periods considered (Fig. 1a,b). The SD, as a way to identify stable/unstable climatic areas, was previously used in other climatic or evolutionary studies4,14. To compute the SD output rasters, we applied the mosaic function setting “fun = sd” from raster R package, calculating the SD for each pixel in the 12 time period rasters for CSI-past and five times for CSI-future, independently for each variable. The mosaic function was also used for the range calculation, with “fun = min” and “fun = max” to obtain the minimum and maximum values of input rasters, respectively, with a further step for subtracting maximum to minimum values.Specifically, for CSI-past, as it includes several time periods with sea-level dropping below the present level (T1, T3, T5, T6, T7, T8, T9; Fig. 1a), we applied a mask of the current land surface, i.e. taking the T12 (Anthropocene) as a template. With this additional step, we were able to remove those pixels (grid cells) currently under the sea but that were once emerged. Most of these pixels, however, were only emerged during the LGM (ca. 21 ka), thus having values for bioclimatic variables for just a single time period (instead of the 12 routinely used for the variability estimation). The inclusion of these areas would result in highly climatically stable regions (low SD values; Supplementary Fig. 1), but this would be an obviously biased result. In contrast, we did not remove those areas affected by the sea-level rising periods, as only three periods contained “NoData” values (T2, T4, T10; Fig. 1a). However, to take this fact into consideration, we created a raster file in which these areas submerged during warm periods are indicated (see Supplementary Fig. 1). Finally, for both CSI-past and CSI-future, the resulting SD values were normalized to values between 0 and 1, with 0 representing completely stable areas and 1 the most unstable ones.The next step was focused on the selection of a relatively uncorrelated set of variables for each map set. We used the removeCollinearity function from virtualspecies R package23 that estimates the correlation value among pairs of variables from a given number of random sample points (10,000 in present case) according to a given method (Pearson for the present case) and a threshold of statistic selected (r  > 0.8 as a cut-off value). The function removeCollinearity returns a list of uncorrelated variables according to the settings specified, randomly selecting just one variable from groups of correlated ones (see Table 1 for a complete list of variables used for each map set). As we compiled estimates of variability independently for each variable and map set (e.g. SD bio1 past, SD bio2 past, etc.), each user can define his own CSI, selecting the more interesting variables according to the case of study.The final CSI maps were obtained by summing the SD values of the variables selected and the subsequent outputs normalized (0 to 1) (Figs. 2–4). Histogram plots were represented with ggplot2 R package24 and maps were exported with ArcGIS v.10.2.2 (Esri, Redlands, California, USA 2014). The histograms were computed for these final CSI maps, which represent the frequency and distribution of CSI values. We presented the final CSI maps with two different colour ramp schemes with ArcGIS. The first consisted of defining equal interval breaks from 0 to 1. The second was based on defining 32 categories with different value breaks for past and future map sets according to the value frequency shown by the histogram plot, i.e. the category with the highest CSI values (no. 32) was 0.71–1 in the past map set and 0.356–1 in the future map set.Fig. 2Maps of Climate Stability Index (CSI) values for the past map set from Pliocene (3.3 Ma) to present (1979–2013), at 2.5 arc-min grid resolution. Colours range from blue for low standard deviation (SD) values, which represents areas with low climatic fluctuations (i.e, low values of CSI) during the period Pliocene–present, to red for high SD values, which shows areas where high climatic fluctuations would have taken place (i.e., high values of CSI). On the upper map, the colour ramp shows equal interval breaks. The histogram with frequency and distribution of CSI values is also shown. On the lower map, the colour ramp has been manually adjusted to a defined set of break values (see details in the text).Full size imageFig. 3Maps of Climate Stability Index (CSI) values for the future conditions (Shared Socioeconomic Pathways: SSP1-2.6, SSP2-4.5, SSP3-7.0, and SSP5-8.5) from present (1970–2000) to future (2100), at 2.5 arc-min grid resolution. Colours range from blue for low standard deviation (SD) values, which represents areas with low climatic fluctuations (i.e, low values of CSI) from present to future, to red for high SD values, which shows areas where high climatic fluctuations would have taken place (i.e., high values of CSI). The colour ramp shows equal interval breaks. The histogram with frequency and distribution of CSI values is also shown for each future scenario.Full size imageFig. 4Maps of Climate Stability Index (CSI) values for the future conditions (Shared Socioeconomic Pathways: SSP1-2.6, SSP2-4.5, SSP3-7.0, and SSP5-8.5) from present (1970–2000) to future (2100), at 2.5 arc-min grid resolution. Colours range from blue for low standard deviation (SD) values, which represents areas with low climatic fluctuations (i.e, low values of CSI) from present to future, to red for high SD values, which shows areas where high climatic fluctuations would have taken place (i.e., high values of CSI). The colour ramp has been manually adjusted to a defined set of break values (see details in the text).Full size image More

Study siteAn open field small-scale release of a GM strain of Anopheles mosquitoes was carried-out in July 2019 in the village of Bana in Western Burkina Faso (see Supplementary Fig. 5). The study was granted regulatory authorisation from the National Biosafety Agency (NBA) (order No. 2018-453/MESRSI/SG/ANB of 10 August 2018 authorising the controlled release of genetically modified sterile male mosquitoes) and institutional ethical permission from the Institutional Ethics Committee for Research in Health Sciences: CEIRES (No. A-003/2019-CEIRES granted on January 9th 2019) and a programme of engagement established community acceptance. Details of the extensive stakeholder and communication processes and activities that were conducted in preparation of this release will be published elsewhere. The village of Bana is located in Western Burkina Faso (12°36′00″N, 3°28′59″W), 23 km west of the city of Bobo-Dioulasso.Bana has two main inhabited agglomerations of similar size: Bana Centre (administrative area) and Bana Market (economic area), separated by a 1.5 km unpopulated land band, crossed by a small semi-permanent river and a forest (see Supplementary Fig. 5). In its entirety, the village comprises about 130 compounds for about 759 inhabitants (local census, IRSS 2014). This region is characterised by two seasons: a wet season from June to September and a dry season from November to April. The mean annual rainfall in the village is about 800 mm and the mean temperature is about 27 °C (22–32 °C)52.Study designThe study design followed the format of an MRR experiment with an intensive period of recaptures followed by several months of monitoring to confirm the disappearance of the transgene. Both the period (July) and design (MRR-like experiment) were informed by previous baseline entomological studies and MRR experiments conducted in the same village41,52. Given the low population size expected in July and to avoid over-sampling, a lower recapture effort (reduction of daily swarm sampling number) was implemented than in previous MRR studies performed in the same area.41 The month of July corresponds to the start of the rainy season, when regular rains and cooler weather promote mosquito survival, and the target population of A. coluzzii is at a much lower level than later in the rainy season41,52. In July, plant coverage is still sparse and males tend to seek refuge inside houses and can be captured in good numbers via indoor sampling52.GM sterile strain maintenanceThe mosquito strain used in the experiment was the genetically modified mosquito Anopheles coluzzii sterile male strain referred to as Ac(DSM)2 (for Anopheles coluzzii Dominant Sterile male strain 2). This strain is the product of local introgression (series of backcrosses) of the original Ag(DSM)2 (dominant sterile male on Anopheles gambiae G3 mosquitoes strain 2) with a local A. coluzzii wild-type (WT) colony (female DSM-carrier crossed with male WT)34. The importation of Ag(DSM)2 in Burkina Faso, introgression with local wild type background and maintenance were conducted under regulatory authorisation from the National Biosafety Agency (N°000002/MRSI/SG/ANB of October 21th 2016). The wild-type A. coluzzii strain used for introgression and maintenance of Ac(DSM)2 was colonised in July 2014 from gravid female adults collected in village 7 of the Kou valley (VK7) in western Burkina Faso. Both colonies were maintained in a dedicated ACL2 (Arthropod Containment Level 2) insectary located within the IRSS main campus at Bobo-Dioulasso, Burkina Faso.For general stock-keeping purposes, Ac(DSM)2 was reared in a dedicated and highly secured climate-controlled room at a temperature fixed at 27.4 °C (±0.2, 95% Confidence intervals) and a relative humidity of 76.3% (±3.2, 95% CIs). Rearing rooms have natural light via windows and were supplemented with an artificial lighting regime of LD 12/12 h photoperiod, including dusk (1 h) and dawn (1 h). Larvae were reared in plastic trays (20 × 30 cm) with 1 l of deionized water and fed with an optimised larvae diet regime53. When mosquito larvae reached their level 3 instar (L3) larvae stage they were sorted manually between transgenic and non-transgenic mosquito larvae using a fluorescent stereomicroscope (Olympus SZX7, 2-8 Honduras street, London, United Kingdom) and put in separated trays to continue their development till pupation. At the pupal stage a second round of sorting occurred to separate male and female (sexing) from both strains. The sexing was done under a basic stereomicroscope (Olympus SZX7 basic, 2–8 Honduras street, London, United Kingdom) using a thin soft brush. Pupae from each strain and sex were placed in small plastic cups inside separate fresh adult cages to emerge. Adults were kept in 30 × 30 × 30 cm insect cages (produced locally) and continuously supplied with 10% (w/v) glucose solution (made with deionized water). Each generation, adult female transgenic mosquitoes were mated with male mosquitoes from the wild-type colony and blood-fed with fresh rabbit’s blood, using a membrane feeder (Hemotek® feeder, Hemotek Ltd, Blackburn United Kingdom). Gravid females were allowed to oviposit in plastic Petri dishes containing a wet sponge covered with filter paper. Eggs were collected and hatched in plastic trays. First instar larvae (L1) were then redistributed into several trays to keep similar larvae abundance (about 250 L1 larvae per tray).In accordance with Mendelian inheritance, stock-maintenance crosses between Ac(DSM)2 females and wild type colony males are expected to generate ~50% hemizygous transgenic male and female progeny referred to as Ac(DSM)2 and 50% non-transgenic sibling with a wild-type phenotype referred to as WT-Ac(DSM)2. That the actual phenotypic proportions matched the expected ratio was checked at each generation a part of standard procedures of colony maintenance.Production, sexing, marking and transport of release mosquitoesReleased males were derived from the 41st backcross generation from strain importation. Assuming Mendelian inheritance, the proportion of residual non-local genetic background after so many generations would be negligible (= 0.541).In rearing the release mosquito cohort, some changes were made in the stock-keeping procedure to maximise the fitness of male mosquitoes to be released. These changes aimed to minimise male mosquito handling during the entire process (rearing, sorting, marking and transport). Crucially, no transgenic versus non-transgenic sorting was done at larval stage resulting in a mix of transgenic and non-transgenic sibling males in the release generation. Additionally, to minimise the number of transgenic female mosquitoes released during the study, male versus female sexing was done at both pupae (initial) and adult (complementary) stages leading to a very high sorting accuracy (over 99.5%). Pupae sexing followed the procedure described for stock maintenance. Next, adult sexing focused on removing the few females resulting from errors in pupal sexing. It consisted of removing those rare females from male mosquito cages through inspection by eye of cages and in using a heat source to attract females. Once spotted, these were removed from male release cages using a mouth aspirator.After pupae sexing, male pupae were placed in 25 × 25 × 25 cm emergence cages (made locally and specially designed to fit dimensions of the secured coolboxes used for secure transportation) at a density of ~1400 pupae per cage. Following adult emergence, and over the following days, the cages were inspected by eye daily to check for and remove any females that had not been detected during the pupae sexing process. This procedure led to a total of 15,384 male mosquitoes aged 3–7 days have emerged in 15 cages and ready for marking and release purposes. Screening of ~50 males randomly picked from each emergence cage was conducted in the ACL2 insectary and revealed a slight bias in favour of WT-Ac(DSM)2 sibling males while Ac(DSM)2 male represented 43.3% (39.7–46.9, 95% CIs) of all emerged males. Based on this genotypic ratio, it was estimated that the male release cohort was equivalent to about 6659 transgenic male mosquitoes Ac(DSM)2 and 8725 non-transgenic sibling mosquitoes called WT-Ac(DSM)2 sibling. All males were kept untouched and in the same cages throughout the whole process until being released.The marking process was performed inside the ACL2 insectary facility, and was carried out the day before field release to allow enough time for mosquito recovery, rest and feeding. The environmental conditions were similar to those used during mosquito production. The mosquitoes were marked directly in their cages by using a cloud dye dusting technique. Aside from being fast, this highly efficient marking procedure (100% of mosquitoes successfully marked) was developed to allow the dust-marking of males in their original emergence cages, thereby avoiding male handling and damage during the marking process. This marking technique consisted of injecting pressurised red fluorescent colour powder (Bioquip® Gladwick Rancho Dominguez, CA 90220, USA; Ref: 1162R) into the cages by using a 5 ml syringe and needle to create a cloud of powder. The cages were wrapped with aluminium foil on all sides to prevent the dust from escaping through the meshed walls. Forceful injection of small amounts of powder from different sides of the cages through the aluminium cover and side netting created a dense cloud of fluorescent powder inside the cages to mark all the mosquitoes. Following marking, sugar-water was available ad-libitum to all marked mosquitoes until field release.About 2 h before the release time, the marked mosquitoes within the mosquito cages were transferred from the IRSS insectary to the release site in Bana village. Before leaving the IRSS insectary, the mosquito cages were covered by a second layer of mosquito net for security purposes. The cages were then wrapped with damp towels and placed in lockable cool boxes dedicated to their transport into the field. After having been secured, the cool boxes containing marked mosquitoes were transported to the release site. The entire process complied carefully with all regulatory requirements related to the permissions received for maintenance, handling and the release of these genetically modified organisms in Burkina Faso.Release phaseAll marked mosquitoes were released on the same day at around 5 pm (about one hour before swarming) in the centre of Bana village by opening the travel cages and allowing free exodus. Mosquitoes that did not leave were counted and subtracted from the released total (n = 534, 3.5%). Taking into account mortality and based on the ratio of Ac(DSM)2 and their siblings previously established, a total of 14,850 male mosquitoes were effectively released, with estimated numbers of 6428 hemizygous transgenic male A. coluzzii mosquitoes Ac(DSM)2 and 8422 non-transgenic WT-Ac(DSM)2 siblings.Recapture phaseMosquito recapture activities started the same day of release (about 2 h after mosquito release) and took place daily for a period of 20 days after release. Two different recapture methods were used: swarm collections using sweep nets (SWN) and pesticides spray catches (PSC) inside houses.Swarm sampling started on the evening of the release day using a well-established sweep net collection method47,54. Previous surveys in the same village41 had allowed mapping of swarm location or natural markers where swarming repeatedly occurs. To ensure sampling across the whole study area, a stratified randomised sampling procedure was used to select and sample 15 mosquito swarms daily at dusk using the sweep net collection method. The area of Bana village and Bana Marché were divided in six and four zones, respectively. Zone 1 and 2 in Bana Village are areas of high swarm abundance and the design ensured that these were not over-represented in swarm collections. Each evening, the teams of capturers set-out to collect up to five swarms per zones depending on swarm availability (swarms are fewer and smaller in early July than later in the month). All mosquitoes captured in the swarms were transported in their sweep nets to the field laboratory and frozen until the next morning for processing. At this stage, a random sample of 15 swarms each day was picked for dust screening and genetic analyses.Pyrethroid spray catches started the morning following the release and continued for 19 days. A set of 20 compounds were sampled each day. The sampling design followed that established in baseline studies leading to the release and in previous MRR studies41. Ten of the compounds were selected completely randomly and the other ten are a fixed set of compounds distributed regularly across the whole village. For each compound selected, a single room (1 sleeping room) within one of the house of compound was chosen for sampling. Although some compounds were selected more than once during the recapture period days, a different room (from a different house inside the same compound when applicable) was selected and no room was sampled twice during the survey period.Pyrethroid spray catches started the morning following the release and continued for 19 days. A set of 20 compounds were selected randomly each day. For each compound selected, a single room (sleeping room) was chosen for sampling. Although some compounds were selected more than once during the seven days, a different room (from a different house inside the same compound when applicable) was selected and no room was sampled twice during the survey period.Captured mosquitoes were identified morphologically in the field using adult anopheline morphological identification keys developed by Holstein55 and a field stereomicroscope (Perfex Sciences® Zoom Pro, Reference: S0852Z5 Toulouse, France). All An. gambiae s.l. mosquitoes were counted, checked for fluorescent dust marking using a Biofinder portable ultraviolet illuminator (Vansky, Shenzhen, China) and preserved in 80% ethanol. The identification of each marked mosquito was confirmed independently by two well-trained members of the staff before conservation in individual 1.5 ml storage microtubes for further analysis. The non-dusted wild Anopheles mosquitoes were pooled (10 individuals per tube) and stored in similar conditions. The location of each collection was recorded and mapped using a GPS (Garmin GPS) device, series GPSMAP®62.2.3. For all recaptured mosquitoes, we calculated the straight line distance from the release point to the recapture location using a Euclidean dispersal distance56. In the present case, the space was assimilated to a two-dimensional orthogonal axis system where xl and yl represent the coordinates of the release point and xr and yr represent the coordinates of the recapture point56. Calculation of the estimated flight distance of the mosquitoes then used the following formula:$${EFD}=sqrt{{left({x}_{r}-{x}_{l}right)}^{2}+{left({y}_{r}-{y}_{l}right)}^{2}}$$
(1)
Ac(DSM)2 male identificationMolecular analysis of recaptured marked mosquitoes was performed by PCR, to identify the Ac(DSM)2 strain and distinguish them from their non-transgenic WT-Ac(DSM)2 siblings. This PCR analysis consisted of detecting the integration of the eGFP::I-PpoI of the DSM transgene which characterised the transgenic mosquito strain Ac(DSM)2. In addition, a molecular species-diagnostic was performed concomitantly using the PCR technique based on the detection of SINE 200× locus57 and this PCR served as a control for DNA integrity. Each mosquito was split into two parts (abdomen and thorax) using forceps. The abdomen was used for the PCR and processed for DNA extraction using ‘squish’ buffer (PCR reaction buffer). The thorax was stored in 80% ethanol at −20 °C. For each mosquito analysed, the same DNA extract was used for both eGFP::I-PpoI transgene detection (identification of Ac(DSM)2 transgenic mosquito) and SINE 200X locus detection (for specie identification and DNA quality control). The Ac(DSM)2 construct was detected using the primers: pBacR-fwd [ATCGGTCTGTATATCGAGGTTTATT] and pBacR-Rev [CTCTAATATTTTGCCAAATGAAGTGCC] targeting the piggyBacR region required for insertion of the transgene. PCR reactions used the Gotaq® PCR kit (GoTaq® G2 Flexi DNA Polymerase, reference: M829B, Promega Corporation, 2800 Woods Hollow Road·Madison, WI 53711-5399, USA).Monitoring of Ac(DSM)2 non-persistenceMonthly mosquito collections were carried out using PSC and swarm sampling to confirm the disappearance of the Ac(DSM)2 transgene from the release site. Monitoring collections were conducted monthly for seven months. This period of monitoring was justified by the regulatory requirement of describing the Ac(DSM)2 disappearance through failure to detect the Ac(DSM)2 transgene for a minimum period of three consecutive months and with high statistical power. During each month of survey, a randomised selection of 20 houses (one room per house) and 20 swarms was sampled. All collected mosquitoes were identified morphologically using identification keys and a field stereomicroscope. Mosquitoes from A. gambiae complex were counted and preserved in 80% (v/v) ethanol for subsequent molecular identification. Each month, a representative sample of collected mosquitoes (up to 300 when available, from both PSC and swarm sampling) was analysed using the Ac(DSM)2-specific and species-specific PCR diagnostics described above to detect whether any A. gambiae s.l. mosquitoes were carrying the DSM transgene.Bayesian inference of mosquito survival, movement and population sizeWe fitted the recapture data to a diffusion model to further investigate dispersal and survival of the marked Ac(DSM)2 and their sibling males, and also to estimate the number of mosquitoes in the background population. This model assumes that the released mosquitoes tend to move in a random manner, meaning they repeatedly take short randomly directed flights that are independent of one another and of the environment. As described below, however, our estimation procedure does also allow for small additional movements where mosquitoes are attracted into nearby swarms at swarming time (dusk), or nearby houses for resting behaviour.We write the diffusion equation as$${partial }_{t}u=D{partial }_{x}^{2}u,$$
(2)
where (u(x,t)) is the probability density of the location of a single marked mosquito at location (x) and time (t), conditional on the individual being alive, and (D) is the diffusion coefficient. Assuming a point release at time (t=0), the above equation has solution$$uleft(r,tright)=frac{{e}^{-frac{{r}^{2}}{4Dt}}}{4,pi D,t}$$
(3)
where (r) is the distance from the release point. We next assume that the released mosquitoes have a constant survival probability of (s) per day, so that the expected number of extant released mosquitoes on day (d) is (R{s}^{d}) where (R) is the number that were released. The expected number of released mosquitoes in a small area ({dA}) is then given by$$qleft(r,dright)=R{s}^{d}frac{{e}^{-frac{{r}^{2}}{4Dd}}}{4,pi D,d}{dA}.$$
(4)
We take three further steps to convert this equation for (q(r,t)) into a likelihood function for the spatio-temporal distribution of recaptures of either Ac(DSM)2 or their sibling males. First, we pool the recaptures on a given day, and made by a given method (either swarm sampling or PSC), by partitioning the study area into annuli centred on the release location. These annuli are the recapture regions, and the expected number of extant marked mosquitoes in a given annulus is the integral of (q(r,d)) over that annulus. This step, therefore, averages out the expected number of marked mosquitoes from the inner to the outer radius of each annulus, and the annulus widths set the scale at which small movements towards swarms or houses, where mosquitoes may be recaptured, are assumed to occur in addition to random movements that underpin the diffusion model. We set the width of each annulus to 50 m, based on our judgement that this distance balances the capacity to separate recaptures at different distances (this capacity reduces with width), with the confidence that movements towards swarms or houses will largely remain within annuli (this confidence increases with width).Second, we assume the observation probability of mosquitoes in a given sample (representing an annulus, capture method, and day), is the number of unmarked mosquitoes in the sample divided by the (unknown) unmarked population size in that annulus. The unmarked population is assumed to have a uniform density, that we will infer alongside the mobility and survival parameters. Finally, we assume the number of marked mosquitoes in a given sample is Poisson-distributed around the expected number.For the data from each recapture method, we used the likelihood function to sample a posterior distribution for the diffusion coefficients and survival rates of the two types of released male mosquitoes, and the density of the unmarked population. We assumed uniform priors with respect to all five parameters and used a Markov chain Monte Carlo algorithm based on Metropolis-Hastings sampling to sample the posterior distribution directly from the log-likelihood. For each analysis (swarm or PSC), we sampled for 100,000 iterations, of which we discarded the initial 20,000 as a transient and thinned the remainder by 100, giving 800 samples in total.Statistical analysisData were analysed using the software JMP 14 (SAS Institute, Inc.). All data were checked for deviations from normality and heterogeneity, and analyses were conducted using parametric and non-parametric methods as appropriate. General linear modelling with Poisson distribution was used to describe male recaptures as a function of genotype and time. Kruskall-Wallis and Mann-Whitney test was used to describe respectively male participation in swarm and Euclidian dispersal distances. General linear modelling with Poisson distribution was used to describe male recaptures as a function of genotype and time. Estimates of population size, survival, and mobility were calculated using a Bayesian approach as described above.Reporting summaryFurther information on research design is available in the Nature Research Reporting Summary linked to this article. More

Van Marle, M. J. E. et al. Historic global biomass burning emissions for CMIP6 (BB4CMIP) based on merging satellite observations with proxies and fire models (1750–2015). Geosci. Model Dev. 10, 3329–3357 (2017).
Google Scholar 
Erb, K. H. et al. Unexpectedly large impact of forest management and grazing on global vegetation biomass. Nature 553, 73–76 (2018).
Google Scholar 
Cook-Patton, S. C. et al. Mapping carbon accumulation potential from global natural forest regrowth. Nature 585, 545–550 (2020).
Google Scholar 
Bastin, J. F. et al. The global tree restoration potential. Science 365, 76–79 (2019).
Google Scholar 
Bowman, D. M. J. S. et al. Fire in the Earth system. Science 324, 481–485 (2009).
Google Scholar 
Archibald, S. et al. Biological and geophysical feedbacks with fire in the Earth system. Environ. Res. Lett. 13, 033003 (2018).
Google Scholar 
Mills, B. J. W., Belcher, C. M., Lenton, T. M. & Newton, R. J. A modeling case for high atmospheric oxygen concentrations during the Mesozoic and Cenozoic. Geology 44, 1023–1026 (2016).
Google Scholar 
Lenton, T. M. in Fire Phenomena and the Earth System: An Interdisciplinary Guide to Fire Science (ed. Belcher, C. M.) 298–308 (Wiley, 2013).Pechony, O. & Shindell, D. T. Driving forces of global wildfires over the past millennium and the forthcoming century. Proc. Natl Acad. Sci. USA 107, 19167–19170 (2010).
Google Scholar 
Marlon, J. R. et al. Reconstructions of biomass burning from sediment-charcoal records to improve data-model comparisons. Biogeosciences 13, 3225–3244 (2016).
Google Scholar 
Archibald, S., Staver, A. C. & Levin, S. A. Evolution of human-driven fire regimes in Africa.Proc. Natl Acad. Sci. USA 109, 847–852 (2012).
Google Scholar 
Santín, C. et al. Towards a global assessment of pyrogenic carbon from vegetation fires. Global Change Biol. 22, 76–91 (2016).
Google Scholar 
Jones, M. W., Santín, C., van der Werf, G. R. & Doerr, S. H. Global fire emissions buffered by the production of pyrogenic carbon. Nat. Geosci. 12, 742–747 (2019).
Google Scholar 
Bird, M. I., Wynn, J. G., Saiz, G., Wurster, C. M. & McBeath, A. The pyrogenic carbon cycle. Annu. Rev. Earth Planet. Sci. 43, 273–298 (2015).
Google Scholar 
Hammes, K. & Abiven, S. in Fire Phenomena and the Earth System: An Interdisciplinary Guide to Fire Science (ed. Belcher, C. M.) 157–176 (Wiley, 2013).Schmidt, M. W. I. et al. Persistence of soil organic matter as an ecosystem property. Nature 478, 49–56 (2011).
Google Scholar 
Lavallee, J. M. et al. Selective preservation of pyrogenic carbon across soil organic matter fractions and its influence on calculations of carbon mean residence times. Geoderma 354, 113866 (2019).
Google Scholar 
Coppola, A. I. et al. Global-scale evidence for the refractory nature of riverine black carbon. Nat. Geosci. 11, 584–588 (2018).
Google Scholar 
Kuzyakov, Y., Bogomolova, I. & Glaser, B. Biochar stability in soil: decomposition during eight years and transformation as assessed by compound-specific 14C analysis. Soil Biol. Biochem. 70, 229–236 (2014).
Google Scholar 
Singh, B. P., Cowie, A. L. & Smernik, R. J. Biochar carbon stability in a clayey soil as a function of feedstock and pyrolysis temperature. Environ. Sci. Technol. 46, 11770–11778 (2012).
Google Scholar 
Masiello, C. A. & Druffel, E. R. M. Black carbon in deep-sea sediments. Science 280, 1911–1913 (1998).
Google Scholar 
Santos, F., Torn, M. S. & Bird, J. A. Biological degradation of pyrogenic organic matter in temperate forest soils. Soil Biol. Biochem. https://doi.org/10.1016/j.soilbio.2012.04.005 (2012).Zimmermann, M. et al. Rapid degradation of pyrogenic carbon. Glob. Change Biol. 18, 3306–3316 (2012).
Google Scholar 
Jones, M. W. et al. Fires prime terrestrial organic carbon for riverine export to the global oceans. Nat. Commun. 11, 2791 (2020).
Google Scholar 
Qi, Y. et al. Dissolved black carbon is not likely a significant refractory organic carbon pool in rivers and oceans. Nat. Commun. 11, 5051 (2020).
Google Scholar 
Pausas, J. G. & Paula, S. Fuel shapes the fire-climate relationship: evidence from Mediterranean ecosystems. Glob. Ecol. Biogeogr. 21, 1074–1082 (2012).
Google Scholar 
Archibald, S., Lehmann, C. E. R., Gómez-Dans, J. L. & Bradstock, R. A. Defining pyromes and global syndromes of fire regimes. Proc. Natl Acad. Sci. USA 110, 6442–6447 (2013).
Google Scholar 
Abatzoglou, J. T., Williams, A. P., Boschetti, L., Zubkova, M. & Kolden, C. A. Global patterns of interannual climate-fire relationships. Glob. Change Biol. 24, 5164–5175 (2018).
Google Scholar 
Brando, P. M. et al. Prolonged tropical forest degradation due to compounding disturbances: implications for CO2 and H2O fluxes. Glob. Change Biol. 25, 2855–2868 (2019).
Google Scholar 
Silva, C. V. J. et al. Drought-induced Amazonian wildfires instigate a decadal-scale disruption of forest carbon dynamics. Phil. Trans. R. Soc. B 373, 20180043 (2018).
Google Scholar 
Withey, K. et al. Quantifying immediate carbon emissions from El Niño-mediated wildfires in humid tropical forests. Phil. Trans. R. Soc. B 373, 20170312 (2018).
Google Scholar 
Pellegrini, A. F. A. et al. Fire frequency drives decadal changes in soil carbon and nitrogen and ecosystem productivity. Nature 553, 194–198 (2018).
Google Scholar 
Reisser, M., Purves, R. S., Schmidt, M. W. I. & Abiven, S. Pyrogenic carbon in soils: a literature-based inventory and a global estimation of its content in soil organic carbon and stocks.Front. Earth Sci. 4, 80 (2016).
Google Scholar 
Wei, X., Hayes, D. J., Fraver, S. & Chen, G. Global pyrogenic carbon production during recent decades has created the potential for a large, long-term sink of atmospheric CO2. J. Geophys. Res. Biogeosci. 123, 3682–3696 (2018).
Google Scholar 
Guimberteau, M. et al. ORCHIDEE-MICT (v8.4.1), a land surface model for the high latitudes: model description and validation. Geosci. Model Dev. 11, 121–163 (2018).
Google Scholar 
Thonicke, K. et al. The influence of vegetation, fire spread and fire behaviour on biomass burning and trace gas emissions: results from a process-based model. Biogeosciences 7, 1991–2011 (2010).
Google Scholar 
Yue, C. et al. Modelling the role of fires in the terrestrial carbon balance by incorporating SPITFIRE into the global vegetation model ORCHIDEE—Part 1: simulating historical global burned area and fire regimes. Geosci. Model Dev. 7, 2747–2767 (2014).
Google Scholar 
Abiven, S. & Santín, C. Editorial: From fires to oceans: dynamics of fire-derived organic matter in terrestrial and aquatic ecosystems. Front. Earth Sci 7, 31 (2019).
Google Scholar 
Santín, C., Doerr, S. H., Preston, C. M. & González-Rodríguez, G. Pyrogenic organic matter production from wildfires: a missing sink in the global carbon cycle. Glob. Change Biol. 21, 1621–1633 (2015).
Google Scholar 
Santín, C. et al. Carbon sequestration potential and physicochemical properties differ between wildfire charcoals and slow-pyrolysis biochars. Sci. Rep. 7, 11233 (2017).
Google Scholar 
Andela, N. et al. A human-driven decline in global burned area. Science 356, 1356–1362 (2017).
Google Scholar 
Arora, V. K. & Melton, J. R. Reduction in global area burned and wildfire emissions since 1930s enhances carbon uptake by land. Nat. Commun. 9, 1326 (2018).
Google Scholar 
Mouillot, F. & Field, C. B. Fire history and the global carbon budget: a 1° × 1° fire history reconstruction for the 20th century. Global Change Biol. 11, 398–420 (2005).
Google Scholar 
Gibson, D. Grasses and Grassland Ecology. Annals of Botany (Oxford Univ. Press, 2009).Dixon, A. P., Faber-Langendoen, D., Josse, C., Morrison, J. & Loucks, C. J. Distribution mapping of world grassland types. J. Biogeogr. 41, 2003–2019 (2014).
Google Scholar 
Bond, W. J. Ancient grasslands at risk. Science 351, 120–122 (2016).
Google Scholar 
Retallack, G. J. Global cooling by grassland soils of the geological past and near future. Annu. Rev. Earth Planet. Sci. 41, 69–86 (2013).
Google Scholar 
Leys, B. A., Marlon, J. R., Umbanhowar, C. & Vannière, B. Global fire history of grassland biomes. Ecol. Evol. 8, 8831–8852 (2018).
Google Scholar 
Alvarado, S. T., Andela, N., Silva, T. S. F. & Archibald, S. Thresholds of fire response to moisture and fuel load differ between tropical savannas and grasslands across continents. Glob. Ecol. Biogeogr. 29, 331–344 (2020).
Google Scholar 
Buisson, E. et al. Resilience and restoration of tropical and subtropical grasslands, savannas and grassy woodlands. Biol. Rev. 94, 590–609 (2019).
Google Scholar 
Rodionov, A. et al. Black carbon in grassland ecosystems of the world. Glob. Biogeochem. Cycles 24, GB3013 (2010).
Google Scholar 
Haberl, H., Erb, K. H. & Krausmann, F. Human appropriation of net primary production: patterns, trends and planetary boundaries. Annu. Rev. Environ. Resources 39, 363–391 (2014).
Google Scholar 
Medan, D., Torretta, J. P., Hodara, K., de la Fuente, E. B. & Montaldo, N. H. Effects of agriculture expansion and intensification on the vertebrate and invertebrate diversity in the Pampas of Argentina. Biodivers. Conserv. 20, 3077–3100 (2011).
Google Scholar 
González-Roglich, M., Swenson, J. J., Villarreal, D., Jobbágy, E. G. & Jackson, R. B. Woody plant-cover dynamics in Argentine savannas from the 1880s to 2000s: the interplay of encroachment and agriculture conversion at varying scales. Ecosystems 18, 481–492 (2015).
Google Scholar 
Satir, O. & Erdogan, M. A. Monitoring the land use/cover changes and habitat quality using Landsat dataset and landscape metrics under the immigration effect in subalpine eastern Turkey. Environ. Earth Sci. 75, 1118 (2016).
Google Scholar 
Şekercioĝlu, Ç. H. et al. Turkey’s globally important biodiversity in crisis. Biol. Conserv. 144, 2752–2769 (2011).
Google Scholar 
Schierhorn, F. et al. Post-Soviet cropland abandonment and carbon sequestration in European Russia, Ukraine and Belarus. Glob. Biogeochem. Cycles 27, 1175–1185 (2013).
Google Scholar 
Jaglan, M. S. & Qureshi, M. H. Irrigation development and its environmental consequences in arid regions of India. Environ. Manage. 20, 323–336 (1996).
Google Scholar 
Joshi, A. A., Sankaran, M. & Ratnam, J. ‘Foresting’ the grassland: historical management legacies in forest-grassland mosaics in southern India, and lessons for the conservation of tropical grassy biomes. Biol. Conserv. 224, 144–152 (2018).
Google Scholar 
Huang, F., Wang, P. & Zhang, J. Grasslands changes in the Northern Songnen Plain, China during 1954–2000. Environ. Monit. Assess. 184, 2161–2175 (2012).
Google Scholar 
Zhou, Y., Hartemink, A. E., Shi, Z., Liang, Z. & Lu, Y. Land use and climate change effects on soil organic carbon in north and northeast China. Sci. Total Environ. 647, 1230–1238 (2019).
Google Scholar 
Williams, N. S. G. Environmental, landscape and social predictors of native grassland loss in western Victoria, Australia. Biol. Conserv. 137, 308–318 (2007).
Google Scholar 
Dowling, P. M. et al. Effect of continuous and time-control grazing on grassland components in south-eastern Australia. Aust. J. Exp. Agric. 45, 369–382 (2005).
Google Scholar 
DeLuca, T. H. & Zabinski, C. A. Prairie ecosystems and the carbon problem. Front. Ecol. Environ. 9, 407–413 (2011).
Google Scholar 
Ceballos, G. et al. Rapid decline of a grassland system and its ecological and conservation implications. PLoS ONE 5, e8562 (2010).
Google Scholar 
Haugo, R. et al. A new approach to evaluate forest structure restoration needs across Oregon and Washington, USA. For. Ecol. Manage. https://doi.org/10.1016/j.foreco.2014.09.014 (2015).DeLuca, T. H. & Aplet, G. H. Charcoal and carbon storage in forest soils of the Rocky Mountain West. Front. Ecol. Environ. 6, 18–24 (2008).
Google Scholar 
Walker, X. J. et al. Increasing wildfires threaten historic carbon sink of boreal forest soils. Nature 572, 520–523 (2019).
Google Scholar 
Bellè, S. L. et al. Key drivers of pyrogenic carbon redistribution during a simulated rainfall event. Biogeosciences 18, 1105–1126 (2021).
Google Scholar 
Abney, R. B., Jin, L. & Berhe, A. A. Soil properties and combustion temperature: controls on the decomposition rate of pyrogenic organic matter. Catena 182, 104127 (2019).
Google Scholar 
Bradstock, R. A., Hammill, K. A., Collins, L. & Price, O. Effects of weather, fuel and terrain on fire severity in topographically diverse landscapes of south-eastern Australia. Landsc. Ecol. 25, 607–619 (2010).
Google Scholar 
Rogers, B. M., Soja, A. J., Goulden, M. L. & Randerson, J. T. Influence of tree species on continental differences in boreal fires and climate feedbacks. Nat. Geosci. 8, 228–234 (2015).
Google Scholar 
Coppola, A. I. & Druffel, E. R. M. Cycling of black carbon in the ocean. Geophys. Res. Lett. 43, 4477–4482 (2016).
Google Scholar 
Stenzel, J. E. et al. Fixing a snag in carbon emissions estimates from wildfires. Glob. Change Biol. 25, 3985–3994 (2019).
Google Scholar 
Murphy, B. P., Prior, L. D., Cochrane, M. A., Williamson, G. J. & Bowman, D. M. J. S. Biomass consumption by surface fires across Earth’s most fire prone continent. Glob. Change Biol. 25, 254–268 (2019).
Google Scholar 
Brando, P. M. et al. Droughts, wildfires and forest carbon cycling: a pantropical synthesis. Annu. Rev. Earth Planet. Sci. 47, 555–581 (2019).
Google Scholar 
Appezzato-da-Glória, B., Cury, G., Soares, M. K. M., Rocha, R. & Hayashi, A. H. Underground systems of Asteraceae species from the Brazilian Cerrado. J. Torrey Bot. Soc. 135, 103–113 (2008).
Google Scholar 
Belcher, C. M. et al. The rise of angiosperms strengthened fire feedbacks and improved the regulation of atmospheric oxygen. Nat. Commun. 12, 503 (2021).
Google Scholar 
Barbero, R., Abatzoglou, J. T., Larkin, N. K., Kolden, C. A. & Stocks, B. Climate change presents increased potential for very large fires in the contiguous United States. Int. J. Wildl. Fire 24, 892–899 (2015).
Google Scholar 
Stephens, S. L. et al. Managing forests and fire in changing climates. Science 342, 41–42 (2013).
Google Scholar 
Trenberth, K. E. Changes in precipitation with climate change. Clim. Res. 47, 123–138 (2011).
Google Scholar 
Prein, A. F. et al. The future intensification of hourly precipitation extremes. Nat. Clim. Change 7, 48–52 (2017).
Google Scholar 
Abatzoglou, J. T., Williams, A. P. & Barbero, R. Global emergence of anthropogenic climate change in fire weather indices. Geophys. Res. Lett. 46, 326–336 (2019).
Google Scholar 
Silveira, F. A. O. et al. Myth-busting tropical grassy biome restoration. Restor. Ecol. 28, 1067–1073 (2020).
Google Scholar 
Strassburg, B. B. N. et al. Global priority areas for ecosystem restoration. Nature 586, 724–729 (2020).
Google Scholar 
Schmidt, H. P. et al. Pyrogenic carbon capture and storage. GCB Bioenergy 11, 573–591 (2019).
Google Scholar 
Fu, Z. et al. Recovery time and state change of terrestrial carbon cycle after disturbance. Environ. Res. Lett. 12, 104004 (2017).
Google Scholar 
Zhu, D. et al. Improving the dynamics of Northern Hemisphere high-latitude vegetation in the ORCHIDEE ecosystem model. Geosci. Model Dev. 8, 2263–2283 (2015).
Google Scholar 
Zhu, D. et al. Simulating soil organic carbon in Yedoma deposits during the Last Glacial Maximum in a land surface model. Geophys. Res. Lett. 43, 5133–5142 (2016).
Google Scholar 
Krinner, G. et al. A dynamic global vegetation model for studies of the coupled atmosphere-biosphere system. Glob. Biogeochem. Cycles 19, GB1015 (2005).
Google Scholar 
Yue, C., Ciais, P., Cadule, P., Thonicke, K. & Van Leeuwen, T. T. Modelling the role of fires in the terrestrial carbon balance by incorporating SPITFIRE into the global vegetation model ORCHIDEE—Part 2: carbon emissions and the role of fires in the global carbon balance. Geosci. Model Dev. 8, 1321–1338 (2015).
Google Scholar 
Hantson, S. et al. The status and challenge of global fire modelling. Biogeosciences 13, 3359–3375 (2016).
Google Scholar 
Hantson, S. et al. Quantitative assessment of fire and vegetation properties in simulations with fire-enabled vegetation models from the Fire Model Intercomparison Project. Geosci. Model Dev. 13, 3299–3318 (2020).
Google Scholar 
Li, F. et al. Historical (1700–2012) global multi-model estimates of the fire emissions from the Fire Modeling Intercomparison Project (FireMIP). Atmos. Chem. Phys. 19, 12545–12567 (2019).
Google Scholar 
Forkel, M. et al. Emergent relationships with respect to burned area in global satellite observations and fire-enabled vegetation models. Biogeosciences 16, 57–76 (2019).
Google Scholar 
Parton, W. J., Stewart, J. W. B. & Cole, C. V. Dynamics of C, N, P and S in grassland soils: a model. Biogeochemistry 5, 109–131 (1988).
Google Scholar 
Singh, N. et al. Transformation and stabilization of pyrogenic organic matter in a temperate forest field experiment. Glob. Change Biol. 20, 1629–1642 (2014).
Google Scholar 
Viovy, N. CRUNCEP Version 7—Atmospheric Forcing Data for the Community Land Model (Research Data Archive at the National Center for Atmospheric Research, Computational and Information Systems Laboratory, 2018); https://doi.org/10.5065/PZ8F-F017Mckee, T. B. T. et al. The relationship of drought frequency and duration to time scales. In Proc. Eighth Conference on Applied Climatology 179–184 (American Meteorological Society, 1993).The NCAR Command Language, Version 6.6.2 (UCAR/NCAR/CISL/TDD, 2019).Freeborn, P. H., Wooster, M. J., Roy, D. P. & Cochrane, M. A. Quantification of MODIS fire radiative power (FRP) measurement uncertainty for use in satellite-based active fire characterization and biomass burning estimation. Geophys. Res. Lett. 41, 1988–1994 (2014).
Google Scholar 
Giglio, L. MODIS Collection 5 Active Fire Product User’s Guide Version 2.5 (Science Systems and Applications, 2013).Huang, N. et al. Spatial and temporal variations in global soil respiration and their relationships with climate and land cover. Sci. Adv. 6, eabb8508 (2020).
Google Scholar 
Warner, D. L., Bond-Lamberty, B., Jian, J., Stell, E. & Vargas, R. Spatial predictions and associated uncertainty of annual soil respiration at the global scale. Glob. Biogeochem. Cycles 33, 1733–1745 (2019).
Google Scholar  More

All of the statistical analyses were conducted using the R programming language version 4.0.5 within the RStudio IDE version 1.4.124,25. Data visualization and processing were performed with the ‘tidyverse’ collection, ‘foreach’ and ‘doParallel’ R packages26,27,28. Geographical Information System (GIS) operations on raster and vector files were conducted using the ‘sf’, ‘exactextractr’ and ‘raster’ R packages29,30,31.Data sources and pre-processingBiodiversity dataWe used the North American Breeding Bird Survey dataset as our source of biodiversity data due to its long temporal coverage and spatial extent14,32. The BBS is composed of bird species abundance records collected since 1966 from over 4,000 survey routes across the countries of Mexico, USA and Canada. For this study we focused solely on routes in the USA, due to their longer time dimension. Data collection follows public access roads that are 24.5 miles long (circa 39.2 km) using a point count protocol whereby routes are surveyed every half-mile (800 m) for a total of 50 stops. At each stop, observers stand for 3 min and record the species and the abundance of every bird seen or heard within 400 m of their location. The routes are surveyed by volunteers with experience in bird observation, and surveys are conducted from late April to July to capture the peak of the breeding season.We selected the years 2001 and 2016 as the two timepoints of our analysis. This 15-year timeframe corresponded to the longest possible timespan for which land cover data products were available at high spatial resolution18. Before analysis, we subset the BBS dataset by removing routes that had incomplete survey lengths (less than 50 point count stops, indicated by the RouteTypeDetailID field value being less than 2 in the extracted BBS dataset) or that were surveyed under adverse weather conditions such as high wind and rain (as indicated by the Run Protocol ID field being equal to 1), which could affect bird occurrence and detectability. Following this filtering process, the total number of BBS routes analysed was 960 (Extended Data Fig. 1).For higher precision when inferring the relationship between avian diversity and environmental variables, we subdivided each route into five segments of equal length, consisting of 10 count locations each. This approach was motivated by the need to more closely associate bird communities with the land cover composition in the area in which they are found. To minimize the spatial autocorrelation between adjacent segments and avoid overlaps in landscapes analysed, we filtered the data to keep only the first, third and fifth segment of each route. These segments therefore formed our sampling unit used in all analyses.We recognized that environmental conditions and stochastic trends in populations could introduce variability in biodiversity calculated from bird community data. We therefore extracted, for each segment and each species, the average population count across a 3-year period centred on our two timepoints (2000, 2001, 2002 and 2015, 2016, 2017)33. We then calculated the mean abundance of each species across these 3 years.The effect of observer experience34,35,36 was accounted for by sourcing the observer ID responsible for each route at each timepoint and including it as a random effect in the legacy model (see ‘Model development’ section). We also controlled for the time of day as it is plausible to expect visibility and avian species activity patterns to vary between early morning and later parts of the day. Time of day for each segment was calculated by averaging across the start and end time data entries associated with each route, and then including this as a covariate in both the legacy and equilibrium models (see ‘Model development’ section). However, we did not model detectability issues associated with traffic noise and disturbance for two reasons. First, all BBS surveys are conducted along public access roads with a vehicle, so the disturbance is expected to be similar across sites. Second, previous studies have found no clear evidence for noise being the main cause for reduced bird abundance near roads37.Following these procedures, our processed BBS dataset included entries of mean abundances of each species for a total of 2,880 segments, corresponding to segment 1, 3 and 5 of 960 routes (Extended Data Figs. 1 and 2). For each segment, at each timepoint we calculated different measures of alpha diversity following the Hill numbers framework38. We then selected to use the effective number of species at q = 1, calculated as the exponential of the Shannon–Wiener Index38. The effective number of species at q = 1 sits at the theoretical half-way point between the classic species richness measure that accounts only for the absolute number of species (q = 0) and the Berger-Parker dominance index (q = infinity), which instead only reflects the most common species. Thus, the effective number of species is a robust alternative to species richness, which does not take account of species rarity or detectability and can thus lead to biased biodiversity estimates16,17.Land cover and environmental dataLand cover data for the US for our focal years of 2001 and 2016 were sourced from the open-access NLCD CONUS products developed by the US Geological Survey (USGS)18,39. The NLCD products are high-resolution (30 m pixel dimensions) classified raster files covering the land area of the whole USA. This dataset provides us with the opportunity to look at finely gridded spatio-temporal changes in a landscape over a relatively long timeframe of 15 years, while utilizing data collected and analysed with the same methods (for example, land use classification algorithms).To reduce the number of potentially collinear explanatory variables included in our models, we aggregated the land cover variables provided by the NLCD dataset. We summarized these to five land cover categories: ‘urban’ (an aggregate of the Developed-Open Space (subclass 21), Developed-Low Intensity (22), Developed-Medium Intensity (23) and Developed-High Intensity classes); ‘forest’ (an aggregate of the Deciduous Forest (41), Evergreen Forest (42) and Mixed Forest (43) classes); ‘grassland’ (an aggregate of the Shrub (52), Grassland/Herbaceous (71) and Pasture/Hay (81) classes); ‘cropland’ (cultivated Crops (82) subclass) and ‘wetland’ (an aggregate of the Woody Wetland (90) and Herbaceous Wetland (95) classes). The Perennial Ice/Snow (12), Open Water (11) and Barren Land (31) classes were excluded from the analysis as they were very uncommon in our dataset. The distribution and total area of the land cover categories across the US are shown in Supplementary Figs. 1 and 2. Temperature data were sourced from the 30 arc-seconds gridded PRISM climate database19 and were extracted as the mean across May and June for each group of years from which bird abundances were taken.We first sampled the landscape surrounding each segment using a range of buffer shapes and sizes, and then selected the buffer type on the basis of the capacity of each buffer type to explain the response variable. The types of buffers that we explored were: a circular buffer around the centroid of the polygon defined by the vertices of each segment (4,000 m radius) and a series of three buffers around the segment line (500 m, 2,000 m and 4,000 m radius). The best fit was given by the smallest buffer size of 500 m, shown in Extended Data Fig. 2, which also coincides with the BBS protocol effective counting distance of 400 m and more closely reflects the size of bird territories14. Land cover variables were computed as a percentage of the total buffer area. Change in percentage points for each land cover type between the 2 years was computed by subtracting the values at the two timepoints. A change product is also provided by the USGS databases40, but it does not meet our needs because it considers land cover changes based on a ranking. Nonetheless, a comparison of urban land cover change between the timepoints showed a similar result (Supplementary Fig. 4). Land cover data were processed geospatially using the NAD 83 Conus Albers Coordinate Reference Systems projection, EPSG 5070.Model developmentTheoretical backgroundWe developed a statistical model that conceptualized extinction debts and colonization credits by combining the following two concepts: (1) the settled biodiversity of avian communities in a given landscape composition (that is, a system at equilibrium) and (2) the lagged response in the species diversity in a given landscape due to recent land cover changes (that is, a system moving to a new equilibrium). We reasoned that, given enough time, and with no further changes in land cover, the effective number of species at a given location would eventually equilibrate. The equilibrium distribution of the effective number of species emerges with the waning of the legacy effect of previous landscape compositions in encouraging or impeding the recruitment and survival of particular species. We did not model these ecological mechanisms directly, but instead expressed the equilibrium of the effective number of species, and the rate of approach to this equilibrium, as empirical functions of environmental covariates. It is important to keep in mind that during a finite time interval following environmental change, it is possible that our observations of effective number of species represent a system in a transitory state towards its new equilibrium. Yet, environmental changes may occur at rates that never allow the system to equilibrate. Although the equilibration processes are latent (that is, not amenable to direct observation), the combination of equilibrium and temporal legacy components into an integrated model, applied to a dataset with extensive environmental replication (due to spatial expansiveness), has allowed us to retrieve distributions for all relevant model parameters (see below).Model overviewThe observed effective number of species Rs,t at site s in year t for t = t1,t2 is modelled as a normally distributed variate with mean μs,t and standard deviation σ$$R_{s,t} approx mathrm{Normal}left( {mu _{s,t},sigma } right)$$
(1)
We assume that, under landscape change, the system is in a state of flux and that the data are from observations witnessing the transition between two (unattained) equilibria. The expected state of the system at any given point in time, μs,t, was formulated as a mixture of past and future equilibrium distributions (that is, a weighted average of the two distributions, where the weights are given by the complementary proportions ω and 1 − ω)$$mu _{s,t} = fleft( {x_{s,t_2};beta } right)omega left( {{Delta}x_{s,t_1,t_2};gamma } right) + fleft( {x_{s,t_1};beta } right)left( {1 – omega left( {{Delta}x_{s,t_1,t_2};gamma } right)} right)$$
(2)
Here, the function f describes the equilibrium distribution of the effective number of species as a function of the configuration of the local environment, captured in covariates xs,t. The weighting function ω depends on covariates ys,t derived from the difference in the local land cover between 2016 and 2001 (that is, it is a function of the land cover change that has taken place). The mixture weights ω and (1 − ω) determine the relative importance of the two equilibrium distributions (past or current). If ω = 1, the interpretation is that the new equilibrium distribution has been completely attained, and thus the current (2016) effective number of species is entirely explained by the current (2016) land cover. Conversely, if ω = 0, the current effective number of species is entirely explained by the past (2001) land cover. The vectors of parameters β and γ, presented in equation (2), are inferred from model fitting.We also augmented equation (2) with a function g of static covariates and random effects z that we expect to have an impact on the effective number of species. Thus, the model comprised equilibrium components, a temporal legacy component and static covariates:$$mu _{s,t} = fleft( {x_{s,t_2};beta } right)omega left( {{Delta}x_{s,t_1,t_2};gamma } right) + fleft( {x_{s,t_1};beta } right)left( {1 – omega left( {{Delta}x_{s,t_1,t_2};gamma } right)} right) + gleft( {z_s;alpha } right)$$
(3)
in which (fleft( {x_{s,t};beta } right)) are the equilibrium components for the two timepoints, (omega left( {Delta x_{s,t_1,t_2};gamma } right)) is the temporal legacy component, and (gleft( {z_s;alpha } right)) is the function that captures the static covariates and random effects, with α being the estimated static covariates parameter effects.Equilibrium componentsWe defined the equilibrium distribution of the effective number of species at a given timepoint as a function (fleft( {x_{s,t};beta } right)) of land cover. This function describes the expected effective number of species at location s, given sufficient time for the community to adapt to the given land cover composition. We now describe this function in more detail.The equilibrium component was formulated as a log-linear model comprising a total of I = 5 environmental covariates (the percentage cover of five landscape classes: urban, forest, grassland, wetland and cropland), using 2nd-order polynomial terms, captured by the coefficient j, to account for optima in effective number of species along each of the five environmental dimensions:$$fleft( {x_{s,t}} right) = {mathrm{exp}}left( {beta _0 + mathop {sum }limits_{i = 1}^{I = 5} mathop {sum }limits_{j = 1}^{J = 2} beta _{i,j}x_{i,s,t}^j} right)$$
(4)
In equation (4), the β parameters capture the effect of covariates on the equilibrium and are assumed to be the same for each environmental composition. A simplifying assumption necessary for the application of this model is that the effective number of species had equilibrated at the first timepoint. As data become available for more years in the future, the influence of this assumption on the model results will diminish and more accuracy will be achievable with multiple timepoints.To allow for conditionality in the effects of one land cover variable on the response of the effective number of species to another land cover variable, we extended this function with pairwise interaction terms k between all the linear terms for land cover variables and pairwise linear-quadratic terms, as follows:$$fleft( {x_{s,t}} right) = {mathrm{exp}}left( {beta _0 + mathop {sum }limits_{i = 1}^{I = 5} mathop {sum }limits_{j = 1}^{J = 2} beta _{0,i,j,}x_{i,s,t}^j + mathop {sum }limits_{i = 1}^{I = 4} mathop {sum }limits_{k = i + 1}^{K = 5} beta _{1,i,k}x_{i,s,t}x_{k,s,t}} right)$$
(5)
Temporal legacy componentThe main covariates, ({Delta}x_{i,s,t_1,t_2}), for the part of the model that captures the temporal legacy, (omega left( {{Delta}x_{s,t_1,t_2};gamma } right)), are derived from the change in land cover (({Delta}x_{i,s} = x_{i,s,t_2} – x_{i,s,t_1})) between the two timepoints$$x_{i,s,} = left{ {begin{array}{*{20}{l}} {x_{1,i,s} = left| {{Delta}x_{i,s}} right|,} hfill & {x_{2,i,s} = 0,} hfill & {{mathrm{if}},{Delta}x_{i,s} < 0} hfill \ {x_{1,i,s} = 0,} hfill & {x_{2,i,s} = {Delta}x_{i,s},} hfill & {{mathrm{otherwise}}} hfill end{array}} right.$$ (6) where ({Delta}x_{s,t,z}) is a vector of the ith environmental change variable (that is, urban, forest, grassland, wetland, cropland) at site s and for directionality z. The effect of these covariates on the mixture weights is given by:$$omega left( {{Delta}x_{s,t_1,t_2};gamma } right) = {mathrm{exp}}left( {mathop {sum }limits_{i = 1}^{I = 5} - gamma _{i,z}{Delta}x_{z,s,i}} right)$$ (7) This formulation weights the contribution that the environmental variables at the two timepoints have on the current effective number of species, as a function of the magnitude and directionality of change in each type of land cover covariate. The γ parameters, and subsequently the temporal legacy component, are allowed via the inclusion of the environmental change data ({Delta}x_{s,t,z}), to account for the distance between the land cover at the two timepoints, therefore quantifying how far the initial community would need to travel to reach equilibrium in 2016 as a function of the type, magnitude and directionality of change. It should be noted that our model, in equation (3), is only implicitly related to the speed with which the effective number of species reacts to environmental changes. Instead, it quantifies how much further it would still have to travel to reach the expected equilibrium associated with the current configuration of the landscape.Static covariatesAs described in model equation (3), we included a function of static covariates to which we can expect the effective number of species to respond without lags relating to the past landscape. We added a linear and quadratic fixed effect for temperature in 2016 to control for any difference in the effective number of species related to climatic characteristics and to allow for a parabolic relationship to be expressed (optima either at mean or extremes values). We also controlled for the heterogeneity of a landscape by including the effective number of land cover types, computed in the same way as the effective number of species, as a fixed effect40. A fixed effect for time of day, reflecting the time at which each segment was surveyed, was included to correct for differences in species detectability between early morning and later parts of the day41. An observer-level random effect was also added to control for variation between observers35,36 and partly account for between-route variation, given that we would expect observers who collect data from multiple routes to do so within a relatively small area. Spatial autocorrelation of the effective number of species was tested for all segments at once and by different radiuses for neighbour inclusion (500 m, 1,000 m, 5,000 m, 10,000 m, 100,000 m), using the Moran’s I statistic42. Spatial autocorrelation was not corrected for because Moran’s I was not significant at any spatial scale (P  > 0.05). Pseudo-replication between neighbouring segments was avoided by considering segments 1, 3 and 5, whose land cover buffers did not overlap (Extended Data Fig. 2).Model fittingThe model was fitted within a Bayesian framework using a Hamiltonian Markov chain Monte Carlo algorithm implemented in the STAN programming language43 version 4.3.0 and the ‘cmdstanr’ R package version 2.26.144.We ran 4 chains, sampling for 1,000 iterations with a burn-in period of 500 iterations each. These numbers of iterations were sufficient to achieve chain convergence. The STAN sampling was run on four parallel threads on a multi-core Intel i7 – 8750H processor with a maximum clock speed of 4.1 GHz.For the purposes of Bayesian inference, all slope parameters associated with the equilibrium component equation (5) and the static additive terms were assigned an unbiased prior (beta _{i,j} approx Nleft( {0,1} right),{mathrm{and}},z_s approx Nleft( {0,1} right),) where N is normal, with the aim of shrinking the parameter estimated towards 0 (that is, no covariate effect). A gamma distributed prior, with shape and rate 0.001, was assigned to the standard deviation of the random effect. For the following known and expected relationships, we also truncated the range of parameter values by bounding the upper or lower limits of the prior/posterior distributions. Intercept and standard deviation of the observer random effect were bounded below by 0. Linear effects for the environmental covariates and temperature were bounded below at 0, while their quadratic counterparts were bounded above at 0. Interaction terms were not limited. The temporal legacy component parameters were given a uniform (U) prior (gamma _isim Uleft( {0,1} right)), bounded between 0 and 1 to act as a weighting proportion between the present and the past. The upper bound on the gamma parameters to 1 does not bias us towards an increased contribution of the past land cover, but instead provides a more conservative approach.Model diagnostics were conducted by assessing chain convergence visually through trace plots, as well as statistically by employing the Gelman-Rubin test, which compares the estimated between-chain and within-chain variances45. Chain autocorrelation and the associated effective sample size were also monitored. In the case of low effective sample size, the chains were extended until the effective sample size exceeded a threshold value of 400. The marginal posterior distribution for each parameter was visualized via a density plot to check for multimodality.Model selection was conducted to inform choice of the size and shape of the land cover buffer around each sampled segment. We did so by comparing values of the Watanabe-Akaike Information Criterion leave-one-out (WAIC)-loo information criterion46 of four different models, each computed using land cover data calculated with two different buffer options of various sizes: a circular buffer around the centroid of the polygon defined by the vertices of each segment (4,000 m radius) and a series of buffers around the segment line (500 m, 2,000 m and 4,000 m radius). This approach was implemented through the ‘loo’ R package version 2.1, which provides an improvement on the original WAIC by including diagnostic measures around the point-wise log-likelihood value estimated around each sample draw47.Visualization of model predictionsA map of the USA (Fig. 1) was produced to represent the predicted extinction debts and colonization credits (that is, positive or negative distance in the effective number of species from the expected equilibria). The map was produced on a hexagonal grid at a spatial resolution of 10 km vertex-to-opposite-vertex, with each hexagon covering a total of 86 km2. Values of extinction debt and colonization credit were calculated by subtracting the predicted effective number of species produced by the model (equation 3) from the predicted effective number of species at equilibrium in 2016 (that is, when the legacy component equals 1). To correctly propagate and represent uncertainty in the extinction debts and colonization credits presented, this process was repeated 1,000 times for predictions originating from different draws from the posterior distribution. Uncertainty in the form of the geometric coefficient of variation, calculated as (root {2} of {{e^{left( {mathrm{log}left( {sigma + 1} right)^2} right)} – 1}}) where σ is the standard deviation, is mapped in Extended Data Fig. 4a. Extended Data Fig. 4 also includes a copy of Fig. 1 (Extended Data Fig. 4b) for reference, alongside upper (Extended Data Fig. 4c) and lower (Extended Data Fig. 4d) credible intervals.Over/underestimation values of biodiversity that could arise by neglecting debts and credits were computed as the difference between the effective numbers of species predicted by the equilibrium model and the legacy model, multiplied by 100 and then divided by the predicted effective number of species under the legacy model. This calculation results in a percentage measurement of the extent to which (in relative terms) the current effective number of species under- or overestimates the diversity that a given landscape can sustain at equilibrium.To further validate our predicted extinction debts and colonization credits, we compared the direction of the expected changes with the recorded difference in effective numbers of species between 2016 and 2019 (the latest year for which data are available). To do so, we sourced bird abundances from the North American BBS dataset14,32 for the year 2019 and conducted the same data processing as described above for the other two timepoints. We then conducted a Pearson correlation test to assess how well the observed change followed the model-predicted one. We are nevertheless aware that a 3-year timeframe is unlikely to be large enough for debts and credits to fully manifest.Plots were also generated to describe the behaviour of the mixture weight, ω (equation 7), which captures the contribution (weighting) of the landscape composition in determining the effective numbers of species at the two timepoints (Fig. 2 in the main text). Values of ω across the whole spectrum of plausible land cover change values (that is, from −100 to +100) were simulated by averaging over 10,000 draws from the posterior distribution of each γ parameter. Credible intervals were measured by taking the 95% range of the 10,000 draws.Explaining spatial variation in debts and creditsThe extinction debts and colonization credits predicted for the contiguous USA states were further modelled to identify which past land cover changes were the main drivers of the delayed biodiversity changes in USA bird communities. We considered the values of debts or credits associated with the 92,000 individual 86 km2 hexagons (Fig. 1) as a response variable. We then specified a Gaussian linear model including the magnitude of each land cover change as explanatory covariates. Positive and negative changes in each covariate were treated as separate linear components to differentiate their effects. The model was fitted to 1,000 sets of debts and credits, each originating from predictions based on independent draws from the posterior distribution. For each generalized linear model (GLM) fit, we then subsequently sampled each parameter distribution another 1,000 times and extracted the summarized parameter estimates from a total of 100,000 values. Model coefficients and their resulting uncertainty from the above process are presented in Fig. 4 and in more detail as part of Supplementary Table 3.Reporting SummaryFurther information on research design is available in the Nature Research Reporting Summary linked to this article. More

Pollen-inferred landscape change and pre-industrial demographyRecently, data derived from tree rings or ice cores have been employed to approximate changes in human economic activity related to past epidemics, as well as to warfare and climatic variability46,47. However, none of these proxies is directly related to human demography or provides a basis to estimate variation in the Black Death’s mortality on a regional scale across Europe (to date only a single archaeological study using pottery as a proxy for demographic change on the national level, focusing on just a single country—England—has appeared48).In recent years, pollen data have been proven to be closely related to demographic variability. Most importantly, detailed comparisons of historical documentary data on population trends and landscape changes as revealed by pollen data have been carried out on a local scale and a close link between changes in European pollen data and changes in European local demography over the past millennium has been demonstrated on multiple occasions, that is, during the period and region of our concern here49,50. A strong link between long-term demographic trends as visible in regional settlement numbers and macro-changes in land cover (deforestation/afforestation) have also been confirmed for ancient Greece51. Additionally, a recent publication successfully employed pollen data to test the extent of the mortality associated with the sixteenth-century Spanish and Portuguese empires’ colonization of tropical regions in the Americas and Asia52. However, as of now there is no method to quantify past demographic trends in absolute numbers based on palaeoecological data. Consequently, we also focus in this paper on relative changes in historical societies’ populations and test the now common idea that the Black Death caused enormous mortality across Europe (with many scholars now arguing for a mortality exceeding 30% and upwards of 50% of the population within a few years) (see also Fig. 1). Using our BDP approach, we conclude this hypothesis is not maintainable. Our evidence for demography-related landscape changes (or lack thereof) negates it.Our main indicator is cereal pollen. In pre-industrial economies, rural labour availability (hence rural population levels) and the spatial scale of cereal cultivation were directly related. An increase in the extent and intensity of cereal cultivation—as reflected in pollen data—would have required not only a predilection and demand for cereals, but also greater availability of labour and thus population growth or significant immigration. The maintenance of existing agricultural activity, in turn, would have required relatively stable population levels53,54,55. The uniform ~50% mortality postulated for the Black Death across Europe should have resulted in a large and significant decline of cereal cultivation and parallel forest regrowth across Europe, as previously demonstrated for mid-fourteenth-century Sweden26 and singular sites in some regions of western Europe56. This result agrees with the fact that Black Death mortality could be high among people at productive age, as illustrated for England57,58. Moreover, even in the case of England, a comparatively commercialized and adaptive rural economy in mid-fourteenth-century Europe, the loss of 50% of the population led to a significant decline in the total area under cultivation (as documented by heterogeneous written sources)59. In Italy, another well-developed economy at that time, the expansion of large estates following the Black Death also did not compensate for the general loss of cereal productivity60. This effect, high mortality driving arable contraction, must have been yet more pronounced in more subsistence-oriented and less adaptive economies, with limited surplus production, such as in regions of the Iberian Peninsula, Germany, Sweden and particularly east-central Europe. Importantly, palaeoecological evidence for arable contraction may be indicative, to some extent, of not only rural population decline but also urban population decline in the region, as there is evidence in some areas, following the pandemic, of rural-to-urban migration, of country-dwellers repopulating urban centres10. Possibly less common was intraregional rural migration, as marginal lands were abandoned for better quality soils, which were more likely to remain under cultivation26,61.Therefore, cereal pollen remains our most potent pollen indicator related to demographic changes in pre-industrial European societies. Other pollen indicators, reflecting rewilding and reforestation (secondary ecological succession) of cereal fields abandoned as a result of significant mortality, or the transformation of cereal fields into pastures, which required less rural labour and thus also could have been a response to high plague mortality, play a secondary role in our analysis and provide further support for our conclusions.BDP data collectionExisting online palynological databases (the European Pollen Database (EPD)62 www.europeanpollendatabase.net, and the Czech Quaternary Palynological Database (PALYCZ)63, https://botany.natur.cuni.cz/palycz/), as well as personal contacts of the study authors and a systematic publication search were employed to identify palynological sites in Europe reaching the required chronological and resolution quality for the study of the last millennium. In order to enable statistical analysis, we included only sites clustered in well-defined historical-geographical regions, excluding isolated sites even if the quality of a site’s data was very good. Data of sufficient quality and amount from regions for which the Black Death is well-studied, notably central and northern England and the Low Countries, is not presently available; to the best of our knowledge, for each of these regions there currently is not more than a single isolated site56, which does not allow for the application of statistical approaches.In total, 261 pollen records with the average temporal resolution of 58 years and 14C-age control (or varve chronology), have been collected. The age–depth models of the sequences have been provided by authors in original publications, by the EPD or developed through the Clam package (version 2.3.4) of R software for the purpose of this study. The analytical protocol for pollen extraction and identification is reported in the original publications. The Pollen Sum includes all the terrestrial taxa with some exceptions based on the selection done in the original publications. The full list of sequences, exclusions from the Pollen Sum, age-depth models and full references are reported in Supplementary Data 1.The taxa list has been normalized by applying the EPD nomenclature. In this respect, the general name Cichorioideae includes Asteraceae subf. Cichorioideae of the EPD and PALYCZ nomenclatures, which primarily refers to the fenestrate pollen of the Cichorieae tribe64. Ericaceae groups Arbutus unedo, Calluna vulgaris, Vaccinium and different Erica pollen types, whereas deciduous Quercus comprehends both Q. robur and Q. cerris pollen types65. Rosaceae refers to both tree and herb species of the family. Finally, Rumex includes R. acetosa type, R. acetosella, R. crispus type, Rumex/Oxyria and Urtica groups U. dioica type and U. pilulifera.BDP summary pollen indicatorsIn order to connect changes visible in the pollen data to human demographic trajectories, we assembled four summary pollen indicators that describe specific landscapes related to human activity. They reflect different degrees of demographic pressure on the landscape (cereal cultivation, pastoral activities, which are less-labour intensive than cereal cultivation, abandonment and rewilding) as well as different durations of land abandonment that might have occurred post-Black Death. Our indicators account for the fact that Europe is a continent rich in natural heritage, with a wide range of landscapes and habitats and a remarkable wealth of flora and fauna, shaped by climate, geomorphology and human activity. In order to ensure uniform interpretation of the indicators, we relied on criteria that can be applied to all European landscapes regardless of their local specificity. Cereals and herding are directly related to human activities and are barely influenced by spatial differences. More complex is the succession of natural plants with their ecological behaviour and inter-species competition. For this reason, we relied on existing quantitative indicators of plant ecology.The Ellenberg L – light availability indicator66 provides a measure of sunlight availability in woodlands and consequently of tree-canopy thickness, reflecting the scale of the natural regeneration of woodland vegetation after cultivation or pasture activities61. Nonetheless, ecological studies have suggested that geographic and climatic variability between different European regions can influence the Ellenberg indicator system67,68,69,70,71. The original indicators were primarily designed for Central Europe58, but several studies developed Ellenberg indicators for other regions, reflecting the specific ecology of the selected taxa (British Isles;72 Czech Republic;73 Greece;74 Italy;75 Sweden76). Plants with L values between 5 and 8 are listed in the fast succession indicator, the ones with L values ranging from 1 to 4 are included in the slow succession indicator. The result is the following list:1) Cereals: only cultivated cereals have been included: Avena/Triticum type, Cerealia type, Hordeum type, Secale. 2) Herding includes pastoral indicators linked to the redistribution of human pressure: Artemisia, Cichorioideae, Plantago lanceolata type, Plantago major/media type, Polygonum aviculare type, Rumex, Trifolium type, Urtica, Vicia type. 3) Fast Succession comprises indicators of relatively recent reforestation of cultivated land after abandonment: Alnus, Betula, Corylus, Ericaceae, Fraxinus ornus, Juniperus, Picea, Pinus, Populus, deciduous Quercus, Rosaceae. 4) Slow Succession includes indicators of secondary succession established after several decades of abandonment: Abies, Carpinus betulus, Fagus, Fraxinus, Ostrya/Carpinus orientalis, Quercus ilex type.In order to validate the indicators overcoming the regional limits of Ellenberg values, a different subdivision has been provided following the Niinemets and Valladares shade tolerance scale for woody species of the Northern Hemisphere77. The subdivision of taxa in the Fast and Slow succession indicators remains the same with only three changes: Fraxinus ornus and Picea move from Fast to Slow succession and Fraxinus from Slow to Fast succession. Extended Data Figs. 1 and 2 show that the two groupings yield the same results, which confirms the reliability of our indicators. There is only one clear exception (Russia), with one more region where smaller-scale diversion occurs for only one indicator, Slow Succession (Norway). The different indicator behaviour results from the different attribution of Picea in our two sets of succession indicators: at high latitude, Picea characterizes the final stage of the ecological succession and hence its different attribution results in different summary indicator values in Russia for the two stages of ecological succession, fast and slow.Please note our summary indicators are not designed to reflect the entirety of the landscape and reconstruct all of its different components. Rather, they are a means of approximating changes in the landscape related to the types of human activities, and their intensity, as much as they relate to demographic changes in human populations using and inhabiting these landscapes.BDP analytical statistical and spatial methodsTo control for local specificity, pollen percentages of every taxon from each pollen site were standardized. From the taxa percentage in a given year the arithmetic mean calculated for the observations from the period 1250–1450 was subtracted and the result divided by the standard deviation for the 1250–1450 period. Standardized taxa results were assembled for each site into four BDP summary indicators. Since each indicator has different numbers of taxa, the sum of standardized taxa values calculated for a given year and site was divided by the number of taxa in the indicator. For the purposes of replication, this standardized pollen dataset, comprising the four indicators for each sample from each site, is available as Supplementary Data 2.This dataset has been analysed in two ways, statistically and spatially.For the statistical approach, standardized regional indices of landscape transformation were created for each region by calculating the average value for all sites within the region, for each of the subperiods analysed in the study (1250–1350 and 1351–1450; 1301–1350 and 1351–1400; 1325–1350 and 1351–1375). Differences between means for each subperiod were measured by the use of the bootstrapping based on 10,000 resamples. The 90% and 95% confidence intervals were estimated with the bias-corrected and accelerated method (BCa)78. These results are visualized in Fig. 5 for the comparison of the subperiods of 1250–1350 versus 1351–1450, and in Supplementary Figs. 4 and 6 for the comparison of the subperiods of 1300–1350 versus 1351–1400 and 1325–1350 versus 1351–1375, respectively.For the spatial approach, we employed the Bayesian model AverageR developed within the Pandora and IsoMemo initiatives (https://pandoraapp.earth/) to map the distribution of pollen indices across Europe. AverageR is a generalized additive model that has been described previously79. It relies on a thin-plate regression spline80 to predict new, unseen data using the following model:$$Y_{i} = {g}( {{{{mathrm{longitude}}}},{{{mathrm{latitude}}}}} ) + {varepsilon}_{i}$$Where Yi is the independent variable for site i; g(longitude, latitude) is the spline smoother; and εi ∼ N(0, σε) is the error term.The spline smoother can be written as X × β where X is a fixed design matrix and β is the parameter vector. Surface smoothing is controlled by employing a Bayesian smoothing parameter estimated from the data and trades-off bias against variance to make the optimal prediction75. This parameter β is assumed to follow a normal distribution: β ~ N(0, 1 /δ × λ × P), where P is a so-called penalty matrix of the thin plate regression spline, which penalizes second derivatives81. The δ parameter is by default set to 1 but this can be adjusted to suit smoothing needs for each application. In our study δ was set at 0.9 to match the preferred spatial scale of analysis for our dataset (approx. 250 to 500 km).AverageR was employed to generate smoothed surfaces for three sets of temporal bins (1250–1350 versus 1351–1450, as well as 1300–1350 versus 1351–1400 and 1325–1350 versus 1351–1375) and for the four BDP indicators (Supplementary Figs. 3, 5 and 7). For the same indicator the difference between the two temporal bins was plotted (Fig. 5; Supplementary Figs. 4 and 6).Reporting SummaryFurther information on research design is available in the Nature Research Reporting Summary linked to this article. More

Erwin, D. H. Curr. Biol. 25, R930–R940 (2015).CAS 
Article 

Google Scholar 
Rabosky, D. L. & Hurlbert, A. H. Am. Nat. 185, 572–583 (2015).Article 

Google Scholar 
Hughes, M., Gerber, S. & Wills, M. A. Proc. Natl Acad. Sci. USA 110, 13875–13879 (2013).CAS 
Article 

Google Scholar 
Novack-Gottshall, P. M. et al. Nat. Ecol. Evol. https://doi.org/10.1038/s41559-021-01656-0 (2022).Article 

Google Scholar 
Wright, D. F. Sci. Rep. 7, 13845 (2015).
Google Scholar 
Deline, B. in Elements of Paleontology hcxz (Cambridge Univ. Press, 2021).Ausich, W. I., Kammer, T. W., Rhenberg, E. C. & Wright, D. F. Palaeontology 58, 937–952 (2015).Article 

Google Scholar 
Zamora, S. & Rahman, I. A. Palaeontology 57, 1105–1119 (2015).Article 

Google Scholar 
Sheffield, S. L. & Sumrall, C. D. Palaeontology 62, 163–173 (2018).Article 

Google Scholar 
Deline, B. et al. Curr. Biol. 30, 1672–1679 (2020).CAS 
Article 

Google Scholar 
Novack-Gottshall, P. M. Paleobiology 33, 273–294 (2007).Article 

Google Scholar 
Anderson, P. S. Paleobiology. 35, 321–342 (2009).Article 

Google Scholar 
Cole, S. R. & Hopkins, M. J. Sci. Adv. 7, eabf4027 (2021).Article 

Google Scholar 
Hopkins, M. J. & St. John, K. Syst. Biol. 70, 1163–1180 (2021).Article 

Google Scholar 
Blomberg, S. P., Garland, T. Jr & Ives, A. R. Evolution 57, 717–745 (2003).Article 

Google Scholar 
Wright, A., Wagner, P. J., & Wright, D. F. in Elements of Paleontology hcx2 (Cambridge Univ. Press, 2021).Cole, S. R., Wright, D. F. & Ausich, W. I. Palaeogeogr. Palaeoclimatol. Palaeoecol. 521, 82–98 (2019).Article 

Google Scholar  More

Isolation and cultivationIn total, 39 mixotrophic flagellates were investigated (Table 1). Thirty-three isolates were enriched and isolated from euphotic zone samples at Station ALOHA (22° 45′ N, 158° 00′ W) in February and May 2019. To select for mixotrophic grazers of Prochlorococcus, whole seawater was amended with K medium [28] (1/20 final concentration), and live Procholorococcus (MIT9301) was added as prey (~5 × 106 cells mL−1 final concentration). Enriched seawater samples were incubated under ~70 µmol photons m−2 s−1 irradiance on a 12 h:12 h light:dark cycle and monitored by microscopy daily up to five days. Each day, samples were serially diluted to extinction (9 dilution steps, 12 replicates per dilution) in 96-well plates in nutrient-reduced K medium (1/20 concentration) with a constant background of Prochlorococcus cells. Wells at the highest dilution showing growth of putative grazers were subjected to 3–6 further rounds of dilution to extinction. Four additional mixotrophs were isolated in full K medium using water from earlier cruises, and two (dictyochophyte strains UHM3021 described in [24] and UHM3050) were enriched in K minus nitrogen medium (K-N) without Prochlorococcus enrichment. All isolates were rendered unialgal, but not axenic, and maintained at 24 °C in K-N medium (~0.2 µM N) amended with Prochlorococcus prey, under the same light conditions as above. Dense Prochlorococcus cells grown in Pro99 medium [29], were harvested and concentrated through gentle centrifugation at 2000 RCF for 5 minutes and resuspended in fresh K-N medium to minimize nutrient carryover. To ensure their long-term accessibility, the isolates used in this study are being transferred to the National Center for Marine Algae and Microbiota at the Bigelow Laboratory for Ocean Science, East Boothbay, ME, USA (ncma.bigelow.org).Table 1 Mixotrophic flagellates investigated in this study.Full size table18S rRNA gene sequencing and phylogenetic analysisCells were harvested by centrifuging 25–50 mL dense cultures at 3000 RCF for 10 min at 4 °C. Genomic DNA was extracted from the pellets using the ZymoBIOMICS DNA Kit (Zymo Research, Irvine, CA, USA). A near-full-length section of the eukaryotic small-subunit ribosomal RNA (18S rRNA) gene was amplified by PCR with the Roche Expand High Fidelity PCR System (Sigma-Aldrich, St. Louis, MO, USA) using either forward primer 5′-ACCTGGTTGATCCTGCCAG-3′ and reverse primer 5′-TGATCCTTCYGCAGGTTCAC-3′ [30], or Euk63F 5′-ACGCTTGTCTCAAAGATTA-3 and Euk1818R 5′-ACGGAAACCTTGTTACGA-3′ [31]. Amplicons were purified using spin columns (DNA Clean & Concentrator-25; Zymo Research, Irvine, CA, USA) and sequenced (Sanger) using the same PCR amplification primers and an additional reverse primer 1174R, 5′-CCCGTGTTGAGTCAAA-3′ [32], when necessary to connect two ends. For phylogenetic analyses, similar sequences were retrieved from the PR2 database [33] based on BLAST similarity, and two environmental homologs (GenBank Acc. FJ537342 and FJ537336) were retrieved from NCBI GenBank for the undescribed haptophyte taxon, which was not affiliated with any reference sequence from the PR2 database. Sequence alignments including 39 isolates, 29 reference and 2 outgroup taxa were created with MAFFT v7.450 using the G-INS-i algorithm [34] in Geneious R11.1.5 (http://www.geneious.com) [35]. Terminal sites that lacked data for any of the sequences were trimmed and any sites with greater than 25% gaps were removed from the alignment, which generated a total sequence length of 1617 bases. Phylogenetic analysis was performed using MrBayes v3.2.6 in Geneious R11.1.5 [36] with two runs of four chains for 1,000,000 generations, subsampling every 200 generations with burn-in length 100,000, under the GTR substitution model. The Bayesian majority consensus tree was further edited within iTOL v5 [37]. All 18 S rRNA gene sequences were deposited in GenBank with accession numbers MZ611704–MZ611740; MN615710–MN615711.Microscopic observationThe average diameter of flagellates in the exponential growth phase (n = 20 cells per strain) was measured by transmitted light microscopy using image analysis software (NIS-Elements AR, Nikon, Minato City, Tokyo, Japan) calibrated with a stage micrometer. Equivalent spherical diameter (ESD) and biovolumes were calculated assuming spherical cells. Chloroplasts were visualized by autofluorescence under epifluorescence microscopy. An average ESD of 0.64 µm was used for Prochlorococcus prey [24].Visual evidence of phagocytosis was obtained by adding fluorescent beads (0.5 μm YG Fluoresbrite Microspheres; Polysciences) to each culture. Samples post incubation (~2 h) were fixed with an equal volume of 4% ice-cold glutaraldehyde, and subsamples (20 µL) were mounted on a glass slide under a coverslip. Paired images captured using epifluorescence and transmitted light microscopy (Olympus BX51 with Leica DFC 7000 T color digital camera) were overlain to identify cells with ingested beads.Grazing experimentsLong-term grazing experiments were conducted for all 39 grazers, and 31 were used to quantify grazing rates based on rates of disappearance of Prochlorococcus cells, which persist but do not readily grow in K-N medium [24]. Rates were not calculated for eight isolates (seven Florenciella and one DictyX) because they were sampled at a lower frequency. Fifteen isolates representing all genera (or approximately genus-level clades) were examined in more detail by replicating grazing experiments two times (marked in bold in Supplementary Table S1), while the remaining sixteen isolates were tested once to survey within- and across- genus variation. Prior to the experiments, all grazer cultures were maintained/acclimated in the experimental medium (K-N with prey). Experiments were initialized by inoculating late-exponential-phase grazers into fresh K-N medium at a final concentration of ~103 flagellates mL−1, and adding live, unstained prey at a final concentration of 2–3 × 106 Prochlorococcus mL−1. Grazers were incubated for 3–8 days in total, depending on how fast prey were ingested. To minimize carryover of dissolved nutrients and prey growth in the grazing experiments, Prochlorococcus were grown to stationary phase in Pro99 medium, then pelleted (2000 RCF for 3–5 min) and resuspended in fresh K-N medium prior to addition to control and experimental cultures. Control cultures of grazer without added prey and prey without grazers were included during each grazing experiment to confirm that grazer growth and prey removal were attributable to grazing. Cell concentrations of prey and grazers were measured every 12–24 h by flow cytometry of glutaraldehyde-fixed samples at final concentration of 0.5% (Attune NxT; Thermo Fisher Scientific, Waltham, MA, USA, and CytoFLEX; Beckman Coulter Life Sciences, Indianapolis, IN, USA). Populations were distinguished based on light scatter and pigment autofluorescence and occasionally confirmed with DNA fluorescence (stained post-sampling with DNA stain SYBR Green I). Ambient bacteria concentrations, monitored using DNA stains, were ≤10% of the added Prochlorococcus at the start and during most periods of all grazing experiments.Grazing rates and biovolume conversion efficiencyWe calculated ingestion rates, I (prey grazer−1 h−1) for each grazer as:$$I = frac{{P_t – P_{t + 1}}}{{G_{{{{{rm{avg}}}}}}(T_{t + 1} – T_t)}}$$
(1)
where (P_t) and (P_{t + 1}) are the prey abundance at sampling interval t and t + 1 (cells mL−1), Gavg is the arithmetic mean grazer abundance (cells mL−1) over the time interval, and ((T_{t + 1} – T_t)) is the time (h) between two sampling intervals. Clearance rates (C, nL grazer−1 h−1) were calculated by dividing the ingestion rate by average prey concentration over the same interval, and specific clearance rate (body volume grazer−1 h−1) was calculated by dividing the clearance rate by cellular biovolume (µm3) of each grazer. Equation (1) uses a linear approximation of prey and grazer trajectories over the sampling interval, which was appropriate for our data where change could be relatively linear, concave-up, or concave-down. Other commonly used ingestion rate calculations assume exponential prey decline [38] and/or exponential grazer growth [39].Clearance rates over time in each experiment were assessed visually to obtain a representative series of rates that minimized potential influence of modest prey growth/decline observed in control cultures (Supplementary Fig. S1), as well as potential slowing of ingestion as the grazer neared carrying capacity or depleted prey to a low concentration. For each experiment a contiguous set of relatively constant rates were used to calculate a mean clearance rate. This assessment sometimes excluded the first 12–24 h, but not when removal rates were particularly fast. Intervals when the grazer neared carrying capacity were also often excluded, if grazing rates slowed down. To assess whether clearance rates increased as prey were depleted we plotted clearance rate and ingestion rate as a function of prey concentration (Supplementary Figs. S2 and S3). In general, there was no relationship between clearance rate and prey concentration, and ingestion increased linearly with prey concentration. These patterns imply that the prey concentrations in this experiment did not saturate the ingestion rates of these grazers. Under non-saturating prey concentrations the average clearance rate over an experiment should be a good estimate of the maximum clearance rate (Cmax). Consistent with this interpretation, functional responses fit to these experiments yielded Cmax estimates that were similar to the reported average clearance rates. Because these experiments were not designed with a sufficient range of prey density to esimate functional responses, we do not report the Cmax estimates, but we note here that our reported average clearance rates may be useful as approximate Cmax numbers in future work.Six grazers representing three classes (three dictyochophytes, two haptophytes, and one chrysophyte) were further investigated to determine functional grazing responses using a wide range of initial prey densities (105–107 cells mL−1). Functional responses were modeled using the Holling type II curve, (I = frac{{I_{{{{{rm{max}}}}}}P}}{{P + frac{{I_{{{{{rm{max}}}}}}}}{{C_{{{{{rm{max}}}}}}}}}}), where I is the ingestion rate over a sampling interval (Eq. 1), Imax is the maximum ingestion rate, Cmax is the maximum clearance rate and P is the arithmetic mean prey density between two sampling points. This curve was fit to ingestion rate data using maximum likelihood with R package bbmle [40].For 31 isolates we calculated the amount of grazer biovolume created per prey biovolume consumed, using data from the same grazing experiments used to calculate grazing rates. It was calculated based on the following formula:$$E = frac{{(F_{{{{{rm{f}}}}}} – F_{{{{{rm{i}}}}}})}}{{(P_{{{{{rm{i}}}}}} – P_{{{{{rm{f}}}}}})}}frac{{B_{{{{{rm{F}}}}}}}}{{B_{{{{{rm{p}}}}}}}}$$
(2)
where Ff and Fi are the final and initial flagellate concentrations, Pi and Pf are initial and final prey concentrations in each culture, and BF and BP are the cellular biovolume of prey and grazer. We refer to the quantity E as the biovolume conversion efficiency, and we use it as an indicator of physiological differences among diverse mixotrophs. Note that biovolume conversion efficiency can be greater than 1, if prey have greater nutrient:biovolume than the grazer.Quantitative PCRReal-time, quantitative PCR (qPCR) was performed to quantify the 18S rRNA gene abundances of representative mixotroph groups discriminated at approximately the genus level, including Florenciella, Rhizochromulina and another undescribed clade within the class Dictyochophyceae; Chrysochromulina and another undescribed clade within the division Haptophyta; clade H in the class Chrysophyceae; and Triparma eleuthera and Triparma mediterranea in the class Bolidophyceae. Primers (Supplementary Table S2) were designed to target a short region (95–176 bases) of the 18S rRNA gene and meet basic criteria (≤2 °C difference in melting temperature between members of a pair, %G + C content between 45 and 65%, ≤1 degenerate position per primer, no predicted primer dimers). Sequences considered targets for a given primer set had ≤1 mismatch across both primers, which included all or most known members within the corresponding targeted clade. Members in the nearest non-targeted clade had ≥3 mismatches distributed across both primers. Efficiency and specificity of the synthesized primers (IDT Inc., Coralville, IA, USA) was tested by ensuring there was specific amplification (qPCR followed by melting curve analysis and gel electrophoresis) when using DNA from cultures within the targeted group and no amplification when using DNA from cultures close to, but outside of the targeted group (Supplementary Table S3). Empirical observations of amplification success using control cultures were used to infer whether species known only by environmental sequences were likely to amplify with a given primer set (Supplementary Fig. S4).In situ gene abundances were quantified in water samples collected from Station ALOHA at 5, 25, 45, 75, 100, 125, 150, and 175 m, during HOT cruise numbers 259 (Jan), 262 (Apr), 264 (Jul), and 266 (Oct) of 2014. Seawater (ca. 2 L) was filtered through 0.02 μm pore-size, aluminum oxide filters (Whatman Anotop, Sigma-Aldrich, Saint Louis, MO, USA) and stored at −80 °C. Genomic DNA of both grazer cultures and environmental samples was extracted (MasterPure Complete DNA and RNA Purification Kit; Epicentre) as described elsewhere [41]. Four replicated PCR reactions (10 μL) were carried out for each sample except for Triparma (duplicates) and consisted of 5 μl of 2× PowerTrack SYBR Green Master Mix (Thermo Fisher Scientific, USA), 10 ng environmental DNA, 500 nM of each primer, and nuclease-free water. Reactions were run on an Eppendorf Mastercycler epgradient S realplex2 real-time PCR instrument. Each run contained fresh serial dilutions (1–6 log gene copies) of target-specific, 750-bp synthetic standards (gBlocks, IDT) prepared in triplicate. The cycling program included an initial denaturation step of 95 °C for 2 min, followed by 40 cycles of 95 °C for 5 s and 55 °C for 30 s. Specificity of amplification was checked with a melting curve run immediately after the PCR program and occasionally, by gel electrophoresis. Amplification efficiencies ranged from 95% to 106% for all the primers.To convert gene copies to cell numbers, 18S rRNA gene copy number per cell−1 was determined for representative isolates in the seven targeted genera/clades. Known quantities of cultured cells (106–107 cells) from each isolate with 2–8 replicates were pelleted at 4000 RCF for 15 min at 4 °C. DNA was extracted from the pelleted cells (MasterPure Complete DNA and RNA Purification Kit, Lucigen), quantified by fluorometry (Qubit, Invitrogen) and the extract volume adjusted to achieve a DNA concentration of 10 ng µL−1. The expected number of eukaryotic cells µL−1 of extract was calculated as the difference between the total cells in the sample prior to centrifugation and in the supernatant afterward (as determined by flow cytometry) divided by the final extract volume. Copy number of the 18S rRNA genes µL−1 of extract was determined by qPCR with the appropriate group-specific primers. The resulting value of gene copies µL−1 was divided by the equivalent number of eukaryotic cells µL−1 in the extract (assuming 100% extraction efficiency) to derive minimum estimates of gene copies cell−1. An average value for representatives within each genus/clade (1–5 isolates) was used to calculate in situ cell concentrations for the genus. These derived in situ abundances were compared to flow cytometric counts of total photosynthetic picoeukaryotes at Station ALOHA obtained from the Hawai’i Ocean Time-series Data Organization and Graphical System (https://hahana.soest.hawaii.edu/hot/hot-dogs/).Global distribution revealed through Tara Oceans 18S rRNA metabarcodesTo estimate the relative abundance of the OTUs closely related to our diverse isolates on a broader geographic scale, we searched the 18S rRNA-V9 sequence data from the 0.8–5 µm fraction of surface water sampled at 40 stations by the Tara Oceans project (http://taraoceans.sb-roscoff.fr/EukDiv/). Reads for ‘Tara lineages’ with highest similarity (E-value < 10−15) to each of our targeted clades (Supplementary Table S1) were expressed as a fraction of total reads excluding dinoflagellates but included all other Tara Oceans phytoplankton ‘taxogroups’: Bacillariophyta, Bolidophyceae, Chlorarachnea, Chlorophyceae, Chrysophyceae/Synurophyceae, Cryptophyta, Dictyochophyceae, Euglenida, Glaucocystophyta, Haptophyta, Mamiellophyceae, Other Archaeplastida, Other Chlorophyta, Pelagophyceae, Phaeophyceae, Pinguiophyceae, Prasino-Clade-7, Pyramimonadales, Raphidophyceae, Rhodophyta and Trebouxiophyceae. Dinoflagellates were excluded because of the difficulty in assigning phototrophic vs. heterotrophic status to all taxa, and because nearly all dinoflagellate reads were from a single, poorly annotated OTU that was also highly abundant in larger size fractions. More