Integrating fossil data in ecological niche models to improve predictions of future habitat of Caribbean corals

Abstract

Ecological niche models (ENMs) are used to assess the abiotic preferences of species by linking their occurrences to the environmental conditions in which they live. We developed a fossil‐informed ENM framework that integrates mid‐Holocene and modern occurrences to test niche stability and reconstruct abiotic niche characteristics for four critical reef‐building Caribbean coral species (elkhorn coral [Acropora palmata], staghorn coral [Acropora cervicornis], boulder brain coral [Colpophyllia natans], and mustard hill coral [Porites astreoides]). Given evidence of niche stability, we used fossil‐improved niche estimates to predict area and location of habitat for future climate scenarios in 2050 and 2100. We built species distribution models with environmental predictors and compared models trained with modern‐only versus combined fossil and modern occurrences to evaluate differences in niche breadth, model performance, and projected habitat distributions under future climate scenarios. Including mid‐Holocene fossil data in ENMs broadened niche estimates, resulting in a larger area of predicted habitat than models based solely on modern data (up to 114,559 km² more in 2100). Although our models showed that suitable habitats existed for most corals in 2100, the amount declined dramatically (45–100% decrease in area from the present day), there was a significant restriction of lower latitude habitat suitability, and marine protected areas did not overlap the majority of predicted future suitable habitat (8–20% overlap by 2100). Fossil‐informed models expanded niche estimates in environmental space and incorporated environmental conditions not represented in modern data, resulting in broader projections of future habitat. Our results suggest that actions to reduce emissions and expand protected areas in the northern Caribbean are imperative to prevent significant degradation and that using fossil occurrences in niche estimation can improve the reliability of conservation forecasting, an approach that is transferable across taxa and regions.

INTRODUCTION

Global biodiversity is declining rapidly (Díaz et al., 2019), and climate change is driving shifts in species’ geographic distributions as organisms track suitable environmental conditions (Chen et al., 2011); therefore, predicting where suitable habitat will persist or emerge is increasingly important for conservation planning and management (Pecl et al., 2017). To address this challenge, ecological niche models (ENMs) are widely used to forecast species’ potential distributions under changing environmental conditions (Elith & Leathwick, 2009). These models quantitatively estimate abiotic niche preferences by correlating species occurrences with the environmental conditions in which they reside.

Although the goal of certain ENM algorithms might be to reconstruct the full range of environmental tolerances of a species (i.e., fundamental niche [Hutchinson, 1957]), ENMs built exclusively on present‐day data are limited by environmental conditions that exist on Earth today (existing niche [Peterson et al., 2011]) and the number of observed species occurrences that can be reduced by dispersal limitations, species sampling, and biotic interactions. Thus, present‐day ENMs reconstruct a time‐limited niche estimate that is closer to Hutchinson's (1957) realized niche than to the fundamental niche. This is a commonly recognized but significant constraint of ENMs across their application (Jiménez et al., 2019; Kearney, 2006; Soberón, 2007).

Although realized niches shape current distributions, models based only on modern data can be biased when projected into novel climates. Including broader climatic variation better approximates the breadth of species’ abiotic requirements and improves model robustness under nonanalog conditions (Jones et al., 2019; Lima‐Ribeiro et al., 2017; Peterson et al., 2011; Williams & Jackson, 2007). This challenge is particularly relevant for coral reef ecosystems, which are experiencing rapid global decline in coral cover, diversity, and ecosystem function (Carpenter et al., 2008; Couce et al., 2013; Dee et al., 2019; Frieler et al., 2013; Pandolfi et al., 2011; Veron et al., 2009). Degradation is especially problematic in the Caribbean because these reefs’ corals have high levels of endemism (Souter et al., 2021) and low functional redundancy (McWilliam et al., 2018), which leads to increased vulnerability (Jackson et al., 2014; Micheli et al., 2014; Roff & Mumby, 2012). Warming also has a significant effect on Caribbean coral health. Record marine heatwaves caused the worst ever coral reef bleaching in the Northern Hemisphere in 2023 (Goreau & Hayes, 2024). Couce et al. (2012, 2013), Freeman (2015), Freeman et al. (2013), and Jones et al. (2019) used ENMs in reef systems. However, these studies were done at the ecotype level, and only data from present‐day observations were used (e.g., the past decade).

Caribbean corals have persisted for hundreds of thousands to millions of years under changing environmental conditions (van Woesik et al., 2012), which provides an opportunity to expand the environmental space used to train ENM. Individual ENMs have been built for Last Interglacial and modern reef coral occurrences at the ecotype level and showed that combining the predictions of these time‐specific models can expand estimates of niche occupation and alter predictions of global reef coral distributions (Jones et al., 2019). Building on this concept, we developed a transferable multitemporal approach at the species level and tested whether it improved niche estimation when applied to individual Caribbean reef‐building corals. This allowed us to evaluate how the expansion of environmental niche estimates varied among species and to demonstrate how fossil‐informed models can address conservation questions.

We characterized suitable environmental conditions for four Caribbean coral species: elkhorn coral (Acropora palmata, Lamarck, 1816), staghorn coral (Acropora cervicornis, Lamarck, 1816), boulder brain coral (Colpophyllia natans, Houttuyn et al., 1773), and mustard hill coral (Porites astreoides, Lamarck, 1816). We used these conditions to project potential distributions today and in the future based on climate model predictions of environmental change. Acropora palmata and A. cervicornis have been dominant reef builders in the Caribbean since the Plio‐Pleistocene (Wallace, 2012), and their branching morphology provides structural complexity for reef ecosystems (González‐Barrios & Álvarez‐Filip, 2018). In the last century, these species have declined due to the synergistic effects of land‐use changes (Cramer et al., 2012), white band disease (Sutherland et al., 2011), and bleaching (Williams et al., 2017) and are now considered critically endangered by the International Union for Conservation of Nature (IUCN) (Crabbe et al., 2022; Miller et al., 2022). Colpophyllia natans is a massive and long‐lived coral that has been affected by stony coral tissue loss disease (Alvarez‐Filip et al., 2022), increased temperatures, and other human influences. This species is listed as vulnerable IUCN (Goergen & Vermeij, 2022). In contrast, P. astreoides is listed by IUCN as least concern (Kitahara et al., 2022). Porites astreoides is a weedy coral that does not provide much structural complexity in reefs (González‐Barrios & Álvarez‐Filip, 2018). Over the last several decades, there has been a shift to reefs dominated by P. astreoides and other weedy species (Aronson et al., 2002; Côté & Darling, 2010). These state shifts reduce habitat diversity and negatively affect associated reef communities (González‐Barrios & Álvarez‐Filip, 2018).

To improve predictions of future habitat suitability, we analyzed potential niche shifts in each species from the mid‐Holocene to the present day and the degree to which adding mid‐Holocene fossil records and environmental conditions improves understanding of future coral habitat suitability. Ultimately, the goal of improved prediction of suitable habitat in the future under this framework is to inform conservation efforts across different taxa, time intervals, and biomes. We specifically tested whether niches were stable from the mid‐Holocene to the present and whether including past warm periods incorporated more environmental niche space, thus producing a model estimate closer to these corals’ fundamental niche. Based on such forecasting, we predicted whether suitable habitat for these corals would decrease, shift northward, or continue to be supported by existing marine protected areas.

METHODS

Species occurrence data

We considered A. palmata, A. cervicornis, C. natans, and P. astreoides because they have ubiquitous Holocene fossil and present‐day records in the Caribbean. Present‐day and Holocene records were downloaded from the Global Biodiversity Information Facility (https://www.gbif.org/). These data were supplemented by field observations in Jamaica (2017–2024). Holocene occurrences were also downloaded from the Paleobiology Database (https://paleobiodb.org/#/) and supplemented with occurrence records in published literature (Adey et al., 1977; Aronson & Precht, 1997; Aronson et al., 2002; Banks et al., 2007; Cuevas et al., 2009; Gischler & Hudson, 2004; Greenstein & Moffat, 1996; Harmon et al., 1983; Hubbard et al., 1998; Lewis, 1984; Lighty et al., 1978, 1982; Ludwig et al., 1996; Macintyre et al., 1977; MacIntyre et al., 1983; O'Dea et al., 2020; Peltier & Fairbanks, 2006; Schellmann & Radtke, 2004; Stathakopoulos & Riegl, 2015; Toscano & Lundberg, 1999; Toth et al., 2019) (Appendix S1).

Taxon data were cleaned in Python by removing duplicates between data sources (matching latitude and longitude). Occurrence points outside the Caribbean were removed. The resulting species data included eight datasets: four species × two time bins (A. cervicornis, A. palmata, C. natans, and P. astreoides × mid‐Holocene, present day). The number of occurrences in each dataset was ≥16 (Appendices S2 & S3).

Environmental data

An ENM operates by comparing species geographic occurrences with the range of environments that exist in those geographic places (presence data) versus environments in geographic localities where the species do not live (background data) in a spatially explicit training region (defined by species’ dispersal information). Our environmental data were derived from present‐day and mid‐Holocene global climate layers. Present‐day and future climate layers were downloaded from Bio‐ORACLE (https://www.bio‐oracle.org/) (Assis et al., 2018; Tyberghein et al., 2012). Future environmental layers from Bio‐ORACLE (Assis et al., 2018; Tyberghein et al., 2012) were downloaded for two different time bins and representative concentration pathway (RCP) scenarios. The RCP scenarios represent a broad range of plausible future climate conditions considering human modification of several variables (e.g., emission, socioeconomic, and biophysical assumptions) that ultimately influence global atmospheric carbon concentrations (Moss et al., 2010). Newer climate projections based on CMIP6 shared socioeconomic pathways (SSPs) incorporate updated data; however, these SSP datasets were not available during model development. To account for potential differences, we evaluated spatial patterns of projected habitat change across scenarios (see “DISCUSSION” & Appendix S15).

We focused on RCP 4.5, a mid‐range scenario in which emissions peak in the mid‐21st century, and RCP 8.5, the highest emissions scenario in which emissions continue throughout the 21st century. We chose ENM geographic projections for 2050 and 2100 to capture the ecological changes before and after emissions are curtailed in RCP 4.5. Mid‐Holocene paleoclimate data were derived from MARSPEC (https://marspec.weebly.com/) (Braconnot et al., 2007; Sbrocco & Barber, 2013). We used the mid‐Holocene ensemble model from Paleo‐MARSPEC, which represents the mean of all PMIP2 general circulation models after downscaling via the delta method. Each model's monthly anomalies (paleo–preindustrial control) were interpolated to a 5‐arc‐minute grid and added to the modern MARSPEC climatology from which the mean, minimum, maximum, and range were calculated per grid cell (Sbrocco & Barber, 2013).

Environmental data were constrained to conditions that are hypothesized to significantly affect coral species’ distributions, including maximum, minimum, and range of sea surface salinity (SSS) and sea surface temperature (SST). In Bio‐ORACLE, these variables are derived from monthly climatologies in which the minimum and maximum correspond to the lowest and highest monthly values, respectively, in each grid cell; the mean represents the average of all 12 monthly values; and the range is calculated as the difference between maximum and minimum monthly values (Assis et al., 2018; Tyberghein et al., 2012). SST and salinity are two of the most important determinants of distributions along coastlines, and predictions of distributions based on these variables can be highly accurate (Belanger et al., 2012). The balance of correlation threshold with the number of environmental variables retained for analyses is important for model accuracy and is an area of active study (Graham, 2003). Environmental layers used here were tested for variable correlation. Coefficients >0.836 were removed to minimize the effects of model bias (Graham, 2003). Variables retained in the model had a maximum correlation coefficient of 0.58 and included maximum and range of SSS and SST. Finally, environmental data from deep regions (i.e., >150 m), where zooxanthellate corals cannot survive (Lesser et al., 2009), were excluded by masking these areas in the SSS and SST rasters with bathymetry layers from MARSPEC and Bio‐ORACLE to remove locations lacking suitable shallow‐water habitat for photosynthetic corals.

The ENMs were trained with occurrences from the geographic region where populations of these coral species survive and reproduce. The choices of training region size and location have implications for model output (see “Ecological niche models”) and are based on the geographic area that species can explore given their dispersal capacity (the M region of Peterson [2011] and Barve et al. [2011]). Acropora palmata, A. cervicornis, and C. natans are endemic to the Caribbean, thus the training region (M) was the Caribbean, which incorporated areas in which these species are known to occur or to have the potential to access in their recent history. Environmental layers for the present‐day and mid‐Holocene time bins were masked to this region. Due to incomplete sampling in Cuba, Hispaniola, and Jamaica for the mid‐Holocene, these islands were removed from the mid‐Holocene training region. The ENMs for P. astreoides were also trained in this region to provide comparable results to the endemic species. Because P. astreoides also exists in the southern Atlantic and Pacific regions, interpretation of ENM results must include consideration of potentially missing environmental information from these areas.

Differences between mid‐Holocene and present day in niche space

To determine whether a coral's niche changed from the mid‐Holocene to the present day and the potential expansion of the model‐based niche, we compared environmental spaces between the present‐day and mid‐Holocene time bins with procedures implemented in the Ecospat R package (Broennimann et al., 2012; di Cola et al., 2017). Ecospat uses observation records of species to extract environmental information, conduct a principal component analysis (PCA), and transform the first two components of the PCA into densities by kernel smoothers (Broennimann et al., 2012). Ecospat measures niche overlap along the PCA gradient with Schoener's D (Schoener, 1970), which varies from no overlap to complete overlap (0–1). To determine whether there was a difference between the predicted mid‐Holocene niche and the niche estimated in the present day, we tested for niche equivalency and niche similarity (Warren et al., 2008). Niche equivalency indicates whether niches are less equivalent than they would be by chance by randomly permuting the occurrences between entities. Niche similarity is less strict and differs from niche equivalency in that it is used to evaluate niche overlap relative to null expectations derived from the available environmental background, rather than being used to test for complete equivalency under a constrained null model. If the overlap is >95%, both entities occupy environments that are more similar or equivalent than expected by chance.

We determined the proportion of the present‐day niche that was not occupied in the mid‐Holocene (hereafter niche expansion); the proportion of the present‐day niche that also occurred in the mid‐Holocene prediction (hereafter stability); and the proportion of the mid‐Holocene niche that was not occupied in the present day (hereafter unfilling) (di Cola et al., 2017). Unfilling values delineated portions of environmental space that were suitable but are not observable today, possibly reflecting biotic interactions or environmental factors not captured. We tested the significance of these proportions against 1000 permutations via one‐tailed tests (ecospat.niche.similarity.test function from Ecospat with default parameters).

Ecological niche models

To determine potential future locations of suitable habitat for these corals, we applied an ENM framework. We used the maximum entropy algorithm (MaxEnt) to predict the geographic extent of species’ suitable habitat based on abiotic suitability. MaxEnt performs well with presence‐only data (Phillips et al., 2006) and small sample sizes (Thibaud et al., 2014). Analyses were done with the kuenm R package (Cobos et al., 2019) with some modifications to fit our specific hypothesis tests (R Core Team, 2022) (R code in Appendix S4).

Candidate models were created from a combination of 15 regularization multipliers (0.1–5 at intervals of 0.5) and two feature classes, quadratic and a combination of linear quadratic and product. We also applied fivefold cross‐validation. The best model was selected based on likelihood criteria (i.e., the Akaike information criterion [AIC] and the Akaike information criterion corrected for small sample sizes [AIC_c] [Warren & Seifert, 2011]). Statistically significant differences from a random prediction were assessed with the partial receiver operating characteristic (partial ROC) that corrects for the lack of specificity data in presence‐only models (Peterson et al., 2008). Eight best models were produced given four coral species × two time bins. These best models were used to create climatic suitability projections under mid‐Holocene, present‐day, and future RCP scenarios. Extrapolation was allowed when species’ response curves decreased toward extremes (Appendix S5). Suitable areas were identified using the maximizing the sum of sensitivity and specificity (maxSSS) threshold to convert continuous suitability to binary projections of suitable versus unsuitable areas (0). The maxSSS threshold is considered one of the best thresholding approaches for presence‐only models and small sample sizes (Liu et al., 2013, 2016). For each coral species, thresholded models trained from the mid‐Holocene and the present day were combined to create a time‐averaged prediction of habitat suitability across all time bins (mid‐Holocene, present‐day, and future RCP scenarios). Where extrapolation risk was identified (Appendix S6), multivariate environmental similarity surface (MESS) analyses were used to identify regions containing environmental conditions outside the training area range in both time bins (Elith et al., 2011; Thuiller et al., 2009). The MESS analyses informed extrapolation risk in binary projections and were interpreted only for variables with response curves that did not decline toward environmental extremes (Appendix S5).

To test our hypothesis of reduced habitat suitability for Caribbean corals in the future, area estimates of binary suitability between scenarios were calculated. Binary models were converted to geographic area by counting the number of suitable habitat cells × area of each cell for each species among the mid‐Holocene, present‐day, and future scenarios. A finding of less geographic area through time would support the hypothesis. To test the hypothesis that species experienced northward migration of future suitable habitat, we compared the geographic area of suitable habitat in 5° latitude bins for the mid‐Holocene, present‐day, and future RCP scenarios. A trend of increasing latitude for the largest suitable habitat area through time would support this hypothesis.

To test the final hypothesis, that areas of future suitable habitat will not be in current marine protected areas, we used the World Database on Protected Areas (UNEP‐WCMC & IUCN, 2025), which provides spatially explicit polygon and point data for global marine protected areas. We extracted all marine protected areas in the Caribbean basin (5–32° N, 98–55° W) and retained only those designated as marine that had established, designated, or inscribed protection status. A 1367‐km² area was used for points lacking a recorded area. This number corresponds to the average size of marine protected areas (Guarderas et al., 2008). For each time bin and species, we estimated the percentage of suitable habitat in the marine protected areas; a declining proportion would support this hypothesis.

RESULTS

Differences in species abiotic niche from the mid‐Holocene to present day

It is impossible to compare niche occupancy in regions that were not environmentally available to both species. Fortunately, most of our present‐day niches and mid‐Holocene niches were in shared environmental space, which allowed for high confidence in our comparisons (Figure 1). Quantitatively, stability values ranged from 0.53 to 0.71 across species (A. cervicornis, p = 1.000; A. palmata, p = 1.000; C. natans, p = 0.924; P. astreoides, p = 0.998) (binomial standard error in Appendix S8); unfilling ranged from 0.03 to 0.27. Acropora cervicornis and A. palmata exhibited the highest unfilling (0.27, p = 0.962, and 0.17, p = 0.999, respectively), suggesting more divergence between mid‐Holocene and present‐day conditions for these species, whereas C. natans and P. astreoides showed greater overlap (unfilling = 0.03, p = 1.000, and 0.13, p = 0.929, respectively).

FIGURE 1.

Mid‐Holocene and present‐day abiotic niches of (a) A. cervicornis, (b) A. palmata, (c) C. natans, and (d) P. astreoides (stability, overlap of mid‐Holocene and present‐day niches; unfilling, proportion of mid‐Holocene niche not occupied in the present day; expansion, proportion of the present‐day niche not occupied in the mid‐Holocene; circles, direction in principle component space each environmental variable contributes most to).

Niche analyses showed high overlap between mid‐Holocene and present‐day niches in A. cervicornis, P. astreoides, and C. natans (Schoener's D = 0.40–0.46). This overlap was greater than expected under a niche similarity test (1000 permutations) (right‐tailed, p = 0.05, 0.03, and 0.01, respectively). For C. natans, niche overlap was also greater than expected under a niche equivalency test (1000 permutations) (D = 0.54, right‐tailed, p = 0.03), indicating strong conservation of environmental space through time (Appendix S7).

The four coral species showed high amounts of niche stability with some niche expansion. Niche unfilling was the most variable and was consistently less prevalent. Unfilling values were greatest for A. palmata and least for C. natans (Appendix S8).

Reduction of habitat suitability in future scenarios

All species were projected to lose habitat under every RCP scenario; losses increased under higher emissions and from 2050 to 2100 (Figures 2 & 3; Appendices S3 & S9). Colpophyllia natans had the most pronounced habitat loss. Despite the loss of substantial habitat, P. asteroides retained more habitat than the other species in 2100 for both RCP 4.5 and 8.5 (Figure 3; Appendix S3).

FIGURE 2.

Habitat for (a–c) A. cervicornis, (d–f) A. palmata, (g–i) C. natans, and (j–l) P. astreoides (top to bottom rows) for the present day (a, d, g, j), for representative concentration pathway (RCP) 4.5 for 2100 (b, e, h, k), and for RCP 8.5 for 2100 (c, f, i, l) (diagonal hatching, present‐day‐trained or mid‐Holocene‐trained data used to predict habitable space; dark cross‐hatching, areas with overlap between models).

FIGURE 3.

Total combined (i.e., mid‐Holocene and present‐day trained) habitat area for A. cervicornis, A. palmata, C. natans, and P. astreoides from the mid‐Holocene, present day, and under two representative concentration pathways (RCP) in 2050 and 2100.

These projections of suitable habitat (Figure 2; Appendix S9) were broadly robust across species and scenarios, though some regions require cautious interpretation due to extrapolation risks illustrated by MESS analysis (Appendix S10). Because many response curves showed ecologically realistic, unimodal, normal relationships with environmental variables (Appendix S5), suitability in these areas should still be interpreted cautiously. For A. palmata, this is especially important. Under RCP 8.5 in 2100, all predicted habitat for this species occurred in nonanalog environmental space (Appendix S10), meaning those estimates carried significant uncertainty. Portions of the RCP 8.5 2050 and RCP 4.5 2100 predictions—including the Florida Keys and the southern coast of Cuba—also were entirely in nonanalog conditions and should also be treated with caution (Appendix S10). Porites astreoides showed similar limitations. The RCP 8.5 2100 habitat was generated by the Holocene‐trained model, which made predictions exclusively for nonanalog areas (Appendix S10). Likewise, the RCP 8.5 2050 predictions for the southern coast of Cuba came only from the Holocene model and were nonanalog; thus, we can be confident only in the regions predicted by the present‐day model for that area. Outside Cuba, however, most predictions occurred in analog environmental space. In the northern Yucatán, nonanalog sections were corroborated by overlapping present‐day model predictions, which strengthened confidence in this region. In contrast, other species’ models carried far less extrapolation risk. All predictions for A. cervicornis were in analog conditions, and the projections for C. natans were dominated by the present‐day trained model, which did not rely on nonanalog extrapolation. Projections in analog conditions substantially increased confidence in those outcomes.

Regardless of the niche stability identified, the training region time bin influenced these results. The above description refers to the combined training datasets. However, all species showed some geographically unique suitable habitat areas when trained in the mid‐Holocene versus models trained with only present‐day data. Generally, present‐day‐trained models were more conservative (less predicted habitat area) than mid‐Holocene‐trained models for future projections. We expected this given warming training‐region conditions in the mid‐Holocene (Figure 2; Appendix S3). Acropora palmata and P. astreoides projected models for RPC 8.5 2100 trained only on present‐day data showed no habitat (Appendix S3).

Future northward shift of suitable habitat

The most extreme habitat loss occured at lower latitudes, and losses were always higher for the RCP 8.5 versus RCP 4.5 scenarios (Figure 4). All species lost suitable habitat around 10° N under RCP 8.5 2100, and only A. cervicornis retained limited area near 15° N. By contrast, most remaining suitable habitat occurred near 20° N across species under the warmer scenarios. For some corals under RCP 8.5, there was no (C. natans) or very little (A. palmata) habitat by 2100 (Figure 3; Appendix S3).

FIGURE 4.

Predicted habitat binned by 5° latitude for the mid‐Holocene, present day, representative concentration pathway (RCP) 4.5, and RCP 8.5 for 2050 and 2100 for (a) A. cervicornis, (b) A. palmata, (c) C. natans, and (d) P. astreoides.

Suitable habitat in marine protected areas

Our combined binary models predicted about 35–40% of suitable habitats are presently in marine protected areas (Table 1; Appendices S11 & S12). The percentage of suitable habitat within marine protected areas changed under future scenarios and differed among species. For C. natans, the percentage of suitable habitat in marine protected areas increased under RCP 4.5 and RCP 8.5 but collapsed entirely by 2100 under RCP 8.5. For P. astreoides, the percentage of suitable habitat in protected areas remained steady under RCP 4.5 but decreased into 2100 for RCP 8.5. For A. palmata, the percentage of suitable habitat in protected areas increased under RCP scenario 4.5, but under RCP scenario 8.5, by 2100 none of the remaining suitable habitat was in protected areas. For A. cervicornis, the percentage of suitable habitat in marine protected areas showed increasing overlap with marine protected areas across all scenarios (Table 1).

TABLE 1.

Percentage of modeled coral habitat in existing marine protected areas (MPAs), including standardized 1367‐km² area ^a around point‐only MPAs, under two representative concentration pathways (RCP) in 2050 and 2100.

Species	Present day	RCP 4.5 2050	RCP 4.5 2100	RCP 8.5 2050	RCP 8.5 2100
P. astreoides	37.2	39.99	42.43	39.82	27.48
C. natans	38.23	46.53	54.06	47.93	None b
A. cervicornis	34.15	40.03	45.78	38.35	47.14
A. palmata	32.99	39.53	38.71	39.74	0.0

^a Area surrounds point‐only marine protected areas (MPAs) (i.e., lacking defined boundaries), corresponding to mean global MPA area.

^b No suitable habitat projected.

DISCUSSION

Merits of incorporating the fossil record in ENMs

There is promising work indicating that combining paleontological and neontological information provides better ENM predictions (Descombes et al., 2015; Jones et al., 2019). Many periods are considered analogs for predicted present‐day global warming, such as the Eocene (Descombes et al., 2015), the Last Interglacial (Greenstein et al., 1998; Kiessling et al., 2012), and the mid‐Holocene (Yin & Berger, 2015). We tested the utility of mid‐Holocene conditions to improve predictions in future scenarios. We found that for reef‐building corals, the fossil record can help fill these gaps to inform conservation decisions and priorities.

Our results confirmed that the mid‐Holocene fossil record provided data to estimate a broader portion of a species’ niche in environmental space (Jones et al., 2019) and thus improved future predictions of suitable coral habitat. In all four coral species, mid‐Holocene‐trained models suggested additional suitable habitat in future climate scenarios than the predictions from present‐day‐trained models alone (Figure 2). Similarly, in environmental space, unfilling values indicated that mid‐Holocene species–environment data contributed additional information to estimates of the species’ niche (Appendix S8). As such, sampling in the Holocene was complementary: gaps in present‐day environmental space were filled in by Holocene‐trained models’ climate space (Appendix S13). The amount of geographic and environmental space added to these models by mid‐Holocene occurrences was related to the number of fossil occurrence points included in the model. Across species, inclusion of mid‐Holocene records expanded the modeled geographic range by approximately 10–40% relative to models built with present‐day occurrences alone (Appendix S14). For example, C. natans had a less extensive fossil occurrence record than the other species investigated, and the C. natans mid‐Holocene‐trained model contributed less unique suitable area in geographic and environmental space (Figure 1; Appendix S3). The lower amount of mid‐Holocene occurrence data was most likely a consequence of incomplete sampling and preservation bias rather than true ecological rarity (Antell et al., 2024). Regardless, as with all ENM studies, additional data, such as the mid‐Holocene records included here, often improve model fidelity (Cobos et al., 2023; Peterson et al., 2011), and data that incorporate new or different environments will have an outsized influence on improved, more accurate predictions.

One concern when using the fossil record for future predictions is the potential that a species’ environmental niche is not stable over time, and the niche estimated from fossil occurrences is not a good representation of current species tolerances and expected responses. Our results (Appendix S7) indicated that for three of the four coral species, C. natans, A. cervicornis, and P. astreoides, niche estimates between time bins were more similar than by chance; thus, their niche was stable over time (Figure 3; Appendix S3). Acropora palmata showed nonsignificant tests of niche equivalency and similarity. However, like the other three taxa, this species showed considerable overlap and stability across time (Appendices S7 & S8). Therefore, we considered that time‐averaged model assessments (incorporating both mid‐Holocene‐ and present‐day‐training data) represented reasonable predictions of species tolerances in the near future. This is consistent with the high levels of overlap and stability between mid‐Holocene and present‐day niches seen in P. astreoides previously (de Oliveira et al., 2024).

On the coral ecotype scale, reef coral niches from the Last Interglacial to the present day were more equivalent than predicted by chance (Jones et al., 2019). Future research should explore the temporal scale of niche stability, that is, the degree to which niche stability is characteristic among populations within a species over thousands of years versus over the lifetime of the species. This may also clarify the question of whether niche estimation of genera in deep‐time studies is a reasonable proxy for species‐level niche stability (e.g., Purcell & Stigall, 2021; Waterson et al., 2016). Our results suggested niche stability may vary between species, even within the same genera (i.e., the acroporids [Appendix S7]), potentially reflecting not only differences in environmental tolerance but also competition among closely related species. Such data were important for generating the most accurate predictions for biological conservation. Ultimately, adding the fossil record provided two endmember hypotheses of future habitat suitability, and the reality is probably between them. In one endmember, coral species can only survive in current conditions and will go extinct by the turn of the century. In the other, coral habitat is not entirely lost, and these vulnerable coral species may survive.

Transferability of the approach

Our approach offered a novel framework that can be used to investigate other taxa of conservation interest. Fossil‐informed models have been applied across diverse organisms including marine mollusks, reptiles, and mammals (Davis et al., 2014; Saupe et al., 2014; Waterson et al., 2016) but without the use of fossil time series to improve model fidelity across the lifetimes of the species. Comparable fossil occurrence data suitable for this workflow are widely available for many taxa through global syntheses, such as the Paleobiology Database, Neotoma (Williams et al., 2018), and New and Old Worlds (Žliobaitė et al., 2023). This workflow can be extended to additional past intervals because it is scalable to any conservation priority taxon with multitemporal occurrence data (Ingenloff & Peterson, 2021; Nogués‐Bravo et al., 2008). Other warm periods, such as the Last Interglacial, can reduce extrapolation into novel climates and strengthen projections (Jones et al., 2019). Environmental data extending into geological time are increasingly available from Global Climate Models across the Phanerozoic (e.g., Burls et al., 2021; Haywood et al., 2020; Otto‐Bliesner et al., 2021; Steinig et al., 2024). Layers processed for modeling are also available for marine and terrestrial systems from the Last Glacial Maximum to the present (MARSPEC; Sbrocco & Barber, 2013; CHELSA; Karger et al., 2023).

Application of these techniques to other reef systems in particular is quite feasible. Global syntheses, such as the PaleoReefs database, document reef occurrences worldwide; these data provide particularly dense coverage in the Cenozoic and Quaternary and resolution from species to higher taxonomic levels (Kiessling et al., 2019). Applying this framework to Indo‐Pacific reefs, including the Coral Triangle, is especially promising because marine species distribution model syntheses are strongly biased toward the North Atlantic and Mediterranean, and there is limited coverage of tropical and Southern Hemisphere regions (Klein et al., 2024; Melo‐Merino et al., 2020; Robinson et al., 2017).

Caribbean reefs in the 21st century

Although coral reef loss has been reported over the last several decades and is predicted for the near future (Chaudhary et al., 2023; Chen, 2021), a high‐resolution view of how coral species will respond in the Caribbean was not clear. Examining changes at the species level is important (Raja et al., 2021), and trait‐based models should focus on reef corals rather than reef loss as a habitat proxy because the two processes can be decoupled in both timing and mechanism. Understanding species‐specific predictions is vital as Caribbean coral reefs are especially vulnerable to climate change and habitat loss. The region has low functional redundancy in corals (McWilliam et al., 2018), and climate predictions indicate Caribbean SST will experience increases of 2–3°C across the 21st century, surpassing regional bleaching thresholds by midcentury (Lawman et al., 2022).

Regional‐scale models address the need for more localized information for conservation efforts (van Hooidonk et al., 2016). Models of Caribbean corals often focus on present‐day conservation priorities, predicting current suitable habitat in existing conservation areas (Egan et al., 2021; van Woesik et al., 2020; Yuen et al., 2023). Although valuable, these approaches provide limited insight into how environmental conditions may shift in coming decades or how conservation priorities and protected area designations may change.

Keeping global temperature increase below 2°C, as stated in the most recent IPCC report (Allan et al., 2023), is critical for coral reef health (Couce et al., 2023). Our results strongly support a need to reduce global emission outputs. Models projected that by 2100 there would be substantial loss of Caribbean coral habitat under the RCP 4.5 scenario and near‐complete loss under the RCP 8.5 scenario (Figure 2). Given that RCP 4.5 reflects a likely overly optimistic view of human intervention in projected climate change, it is likely that reefs in the future will look vastly different from those of the past.

Our models were limited to SST and salinity conditions, which may differ from the benthic conditions directly affecting the corals. Moreover, conditions such as nutrification, turbidity, and acidification are also important for coral health, and without these dimensions, these models should be interpreted as a simplified estimate. Coral disease (e.g., white band disease in A. cervicornis and A. palmata [Sutherland et al., 2011], stony coral tissue loss disease in C. natans [Alvarez‐Filip et al., 2022]) and biotic interactions, such as a continued loss of herbivores (Hughes et al., 1985; Jackson et al., 2001), will add biological stress to these corals into the future, further reducing suitable habitat. These chronic local stressors have contributed to the degraded state of many Caribbean reefs (Jackson et al., 2014). Some of the areas predicted suitable here, such as the south coast of Cuba and the Florida Keys, will be nonanalog environmental conditions in the future. This represents an area of model uncertainty, particularly for the species we modeled without complete unimodal response curves (e.g., A. palmata and P. astreoides under RCP 8.5 in 2100). Even though these areas were predicted suitable, they may not be habitable in the near future; therefore, our models (Figure 2) are likely overestimates of future suitable habitat.

These limitations suggest that our projections were conservative and future conditions are potentially more severe than modeled here. Indeed, newer climate models released after these analyses (SSP) project warmer and more saline sea‐surface conditions across the Caribbean. Comparisons of RCP‐ and SSP‐based climate projections in the Caribbean indicate that SSP models predict warmer, more saline, and more variable sea‐surface conditions across the Caribbean relative to RCP models (Appendix S15). Regions that retained suitable habitat under RCP scenarios—particularly the Florida Keys, Cuba, and the north coast of the Yucatán Peninsula—showed consistently higher temperature and salinity under SSP projections. This pattern was somewhat less pronounced for the Yucatán, suggesting that this region may retain suitable habitat longer than others (Appendix S15).

The persistence of P. astreoides under warming conditions, combined with its relatively high dispersal ability (Serrano et al., 2016), supports predictions that reefs will increasingly be dominated by weedy species (González‐Barrios & Álvarez‐Filip, 2018; Green et al., 2008). Porites astreoides provides less reef structural complexity than the other species we analyzed (González‐Barrios & Álvarez‐Filip, 2018), and shifts toward dominance by low‐relief taxa are expected to reduce carbonate production and long‐term framework maintenance (Perry et al., 2012). Thus, reefs that are predominantly composed of P. astreoides will lack habitat dimensions that support the diversity of organisms that people rely on to survive (González‐Barrios & Álvarez‐Filip, 2018).

Model predictions of habitat loss at lower latitudes

Many models that predict future coral reef habitat indicate a shift poleward in concert with a decline in habitat at the equator (Freeman et al., 2013; Jones et al., 2019). The poleward migration of coral reef habitats has been documented in the fossil record during warming intervals, such as the Last Interglacial (Jones et al., 2019; Kiessling et al., 2012) and the Eocene (Descombes et al., 2015). Our results for specific reef corals did not indicate a significant northward shift of Caribbean coral species (Figure 4). This difference may reflect reef persistence dependent on other taxa or community shifts beyond the species modeled here. In addition, the models’ limited ability to extrapolate means these results should be interpreted with caution. The lack of northward shift for the Caribbean is suggested by other studies (e.g., Chaudhary et al., 2023), and these corals may track suitable habitat southward, a pattern that warrants further study. Our analyses indicated there will be a more substantial loss of suitable habitat at lower latitudes, corroborating models of global‐scale reef patterns (Descombes et al., 2015; Kiessling et al., 2012).

Dispersal ability to track suitable habitat area is a predictor of species ability to survive climate change (Trakhtenbrot et al., 2005). Latitudinal and bathymetric shifts have been witnessed in response to global climate change (Chen et al., 2011). It will be important for coral species to be able to track suitable habitat into the future. Porites astreoides can disperse over long distances within the Caribbean, with connected communities over 1700 km away (Serrano et al., 2016). In contrast, the two Acropora species in our study have limited dispersal (Hemond & Vollmer, 2010) compared with P. astreoides, so the latter may be better able to keep up with shifting habitats. For instance, in A. palmata, a dispersal distance of only 1 km has been documented previously (Japaud et al., 2019). The dispersal ability of C. natans is poorly constrained, but estimating these kinds of data is critical to generating reliable predictions about their future distributions.

Habitat suitability is species specific

Niche estimates for these four coral species showed within‐species stability (Appendix S7), but this stability was not seen between species as is particularly evident looking at their differing projections of suitable habitat. The specificity of species niche dimensions has real conservation consequences. The four species we studied are in different threat categories. Acropora cervicornis and A. palmata are critically endangered (Crabbe et al., 2022; Miller et al., 2022), C. natans is vulnerable (Goergen & Vermeij, 2022), and P. astreoides is a species of least concern (Kitahara et al., 2022). That said, these categories are not fully consistent with our results of this study. Colpophyllia natans had the greatest reductions of habitat, whereas A. cervicornis and P. astreoides sustain the least loss of suitable area by 2100 under both RCP scenarios. The mismatch with threat status may indicate that C. natans is more at risk than previously assumed. The IUCN assigns species categories of extinction risk looking at population trend, size, structure, and geographic range (IUCN Standards & Petitions Committee, 2024). Those criteria are likely not reflected in abiotic changes as modeled here. Thus, our results provide additional information and improve vulnerability estimates.

Marine protected areas and future coral reefs

Around 73% of modern coral reefs are outside marine protected areas (UNEP‐WCMC & IUCN, 2025). Our results corroborated this, with about 65% of predicted suitable habitat falling outside of current marine protected areas in the Caribbean (Table 1; Appendix S15). Under RCP 8.5, the proportion of suitable habitat within protected areas declines for P. astreoides, A. palmata, and C. natans by 2100, indicating the need to expand protected regions. In contrast, the percentage of suitable habitat in protected areas increases across all scenarios for A. cervicornis, potentially indicating that current areas are better placed for this species. Protected areas should be established in the northern Caribbean in localities predicted to have suitable habitats in the near future. There are nature reserves near predicted future suitable habitat, such as the Sian Ka'an Biosphere Reserve in Mexico and the Florida Keys National Marine Sanctuary.

The decision to create or expand marine protected areas is not as simple as targeting areas where species will be likely to thrive. Complicating factors, such as community buy in, funding, and local policies, are important in all conservation decisions (Voyer & Gladstone, 2018). Despite these difficulties, our results suggest that expanded marine protected areas in suitable coral habitats will be crucial to support reefs into the future.

Caribbean reefs are vital resources, especially for people who live near them. Shallow water coral reefs provide essential ecosystem services, including income (Brander et al., 2007), protection from storm damage (Ferrario et al., 2014), and vital nutrition from associated seafood (Grafeld et al., 2017; Hicks et al., 2019). In the ENMs we developed, the largest areas of future suitable habitat for reef corals exist in the northern parts of the Caribbean (Figure 2). Regions that rely on coral reefs, such as the Lesser Antilles and Jamaica, are likely to be without suitable habitat for these reef corals by 2100 (Figure 3).

These findings echo calls for climate‐smart conservation (Brito‐Morales et al., 2022), which emphasizes anticipating spatial shifts in biodiversity. For Caribbean reefs, this underscores the urgency of flexible protection strategies that can accommodate range contractions. Our results highlight the northern Caribbean as likely future refugia for several shallow‐water coral species. However, the proportion and spatial configuration of suitable habitat in marine protected areas differ among emissions scenarios, underscoring that future refugia would have to be dynamic rather than fixed. Continued monitoring, modeling, and adaptive management will be essential to ensure that protection keeps pace with these shifting habitats. Expanding protected coverage strategically around such resilient zones and enhancing connectivity between existing marine protected areas, such as Sian Ka'an and the Florida Keys National Marine Sanctuary, would strengthen the persistence of coral ecosystems. Acting now to protect future refugia is critical to ensure that Caribbean reefs, and the livelihoods they support, endure beyond this century.

Supporting information

Supplementary Information

Supplementary Information

Appendix S3. Area calculations for each species in each iteration of modeling.

Supplementary Information

Supplementary Information

Appendix S6. Model chosen, AICc values of model and null, and directions to caution extrapolation for each iteration of modeling.

Appendix S7. Lesser and greater equivalency and similarity values.

Appendix S8. Expansion, stability, and unfilling between the mid‐Holocene and present‐day for four Caribbean coral species, with permutationbased p‐values and binomial standard errors for each metric.

Appendix S9. Combined models for all time periods and RCP scenarios

Appendix S10. Multivariate Environmental Similarity Surface (MESS) plots for mid‐Holocene trained models (top) and present‐day trained models (bottom row) in projected time periods where extrapolation was not allowed as indicated in supplemental table 4.

Appendix S11. Marine protected areas (MPAs) in the Caribbean basin as of 2024.

Appendix S12. Maps of marine protected areas (MPAs) in the Caribbean basin as of 2024 overlying combined binary suitable habitat for A. cervicornis, A. palmata, C. natans, and P. astreoides (top to bottom rows) for the 2100 RCP 4.5 and 8.5 models.

Appendix S13. Distributions of environmental conditions associated with mid‐Holocene and present‐day occurrences of A. cervicornis, A. palmata, C. natans, and P. astreoides based on maximum salinity, maximum temperature, salinity range, and temperature range.

Appendix S14. Comparison of suitable habitat pixels for Acropora cervicornis, A. palmata, Colpophyllia natans, and Porites astreoides across time intervals.

Appendix S15. Comparison of shared socioeconomic pathways (SSP) and representative concentration pathways (RCP) Bio‐ORACLE projections for sea‐surface temperature and salinity variables.

Williams, C. M. , Machado‐Stredel, F. , Martindale, R. C. , & Myers, C. E. (2026). Integrating fossil data in ecological niche models to improve predictions of future habitat of Caribbean corals. Conservation Biology, 40, e70323. 10.1111/cobi.70323