Main

Mycorrhizal fungi form nutritional symbioses with >80% of plant species and build extensive underground hyphal networks that can constitute >30% of the living microbial biomass of soils1. The fungi use their networks to forage in the soil for phosphorus, nitrogen and trace elements, and receive carbon from plant root systems in return. The diversity and activity of these fungal communities is fundamental to the functioning of Earth’s terrestrial ecosystems, with an estimated 3.6 billion tons of carbon annually allocated through plants to mycorrhizal fungi2,3. However, despite their importance, our understanding of the global distribution of mycorrhizal fungal biodiversity lags far behind plants and animals. As a result, it is difficult to develop appropriate measures to protect these organisms and the integrity of their symbiotic partnerships4,5.

Over the past decade, the emergence of comprehensive, large-scale molecular datasets on belowground fungi have made it possible to establish global baselines of mycorrhizal fungal biogeography6,7,8. These efforts have helped reveal the dominance of two types of mycorrhizal symbionts spread across Earth’s terrestrial biomes: ectomycorrhizal (EcM) fungi and arbuscular mycorrhizal (AM) fungi. These fungal types differ in their physiology, symbiotic behaviour and contribution to carbon cycling9,10. Although only about 2% of terrestrial plant species depend on EcM associations, these host species constitute >25% of global vegetation area and dominate in high-latitude forests11. AM fungi form associations with around 80% of plant species11 and dominate in tropical forests, grasslands and croplands globally. A third type—ericoid mycorrhizal (ErM) fungi—are common in high-elevation and high-latitude ecosystems but are rarer globally11. The scarcity of ErM fungi in large-scale fungal datasets has left a sizable gap in our understanding of their global diversity patterns or conservation opportunities.

Previous work has used plant genus as a proxy of mycorrhizal type to reveal fundamental insights into the distributions and functions of mycorrhizal symbioses12,13,14. However, efforts to explore mycorrhizal fungal biogeography from direct microbial observations, with methods such as DNA-based surveys of soil samples15,16,17,18,19, have shown that aboveground and belowground biodiversity patterns are often not correlated20. This finding means that patterns of plant communities may not accurately reflect distributions of mycorrhizal diversity belowground. Moreover, our ability to identify the scale of this mismatch is limited by geographical biases in the locations of mycorrhizal samples used for predictive modelling. Such large data gaps inevitably result in extensive model extrapolation into untrained environmental space21, which is particularly unreliable in under-sampled tropical regions where AM fungi are likely to dominate. Generating high-resolution global distributions of mycorrhizal fungal diversity from direct observations—and quantifying the extent of model uncertainty22—is essential to guide future conservation planning5. Protecting mycorrhizal fungal diversity will help avoid species loss across other taxonomic groups and help maintain critical ecosystem functions (for example, carbon sequestration) for realizing nature-based climate mitigation strategies4,5. To address these challenges, we took the following steps: (1) built machine-learning models to create high-resolution (1 km2) spatial predictions of mycorrhizal fungal richness and endemicity (rarity-weighted richness) at the global scale; (2) identified mycorrhizal fungal richness and endemism hotspots and assessed the extent of their current environmental protections; and (3) characterized spatial uncertainty and data limitations.

Calculating global fungal richness

First, we explored the distribution and range of the training data (Fig. 1). This dataset consisted of a globally distributed collection of nearly 25,000 soil samples containing >2.8 billion fungal DNA sequences from 130 countries compiled in the GlobalFungi, GlobalAMFungi and Global Soil Mycobiome consortium databases6,7,8. These are the largest datasets of fungal internal transcribed spacer (ITS) and small subunit (SSU) rRNA amplicon sequences assembled so far. These sequences were analysed using virtual taxa (VT) for AM fungi and 97% similar operational taxonomic units (OTUs) for EcM fungi. With these data, we estimated AM and EcM fungal richness using a rarefaction and extrapolation approach, and rarity-weighted richness (hereafter ‘rarity’; Extended Data Fig. 1), which is a metric of relative endemism used to guide conservation priorities23.

Fig. 1: Sample locations and mycorrhizal richness trends by biome.
figure 1

a,b, Distribution of sites and richness estimates for AM fungi (a) and EcM fungi (b). VT were created from SSU sequences for AM fungi, and 97% similar OTUs were created from ITS sequences and assigned to EcM fungi. Mycorrhizal richness patterns across terrestrial biomes are shown in boxplots (mangroves, flooded grasslands and rock and ice biomes not shown owing to low numbers of samples in both datasets). Global distributions of estimated richness are shown in density curves above biome-level boxplots. Boxplots indicate the median (centre line), first and third quartiles (lower and upper box edges) and 1.5× the interquartile range (IQR) (box whiskers). EcM richness data are shown as square-root transformed for visualization. Richness estimates were calculated using a rarefaction and extrapolation approach that incorporates sequencing depth per sample. Points shown here are all samples that passed quality-control checks and used as training data.

Full size image

Across both AM and EcM datasets, temperate forests were the most sampled biomes (32–61% of all samples), and mangroves and flooded grasslands were the least sampled biomes (0.1% of all samples). There were only a few AM fungal samples from desert and tundra biomes, and few EcM fungal samples from tropical dry forests and tropical conifer forests. Compared with other biomes, soil samples from montane grasslands, tropical conifer forests and temperate broadleaf forests showed the highest AM fungal richness estimates (Fig. 1a). Similarly, montane grasslands had the greatest AM fungal rarity estimates, followed by Mediterranean forests and tropical grasslands with the next highest in AM fungal endemism levels (Extended Data Fig. 1a). For EcM fungi, samples from coniferous forests (tropical and temperate) and temperate broadleaf forests had the highest richness estimates (Fig. 1b). EcM fungal endemicity was also highest in tropical conifer forests, whereas tundra and Mediterranean forests showed a greater concentration of rare EcM fungi than temperate forests (Extended Data Fig. 1b). We used these datasets to train random forest machine-learning models to predict global AM and EcM fungal richness and endemism patterns.

We wanted to know whether mycorrhizal fungi show clear global diversity patterns like plants and animals. To that end, we used maps of predicted richness to plot mean richness and rarity across latitudes. Predicted AM fungal richness was highest in ecosystems near the equator and gradually declined towards the poles. This result was in agreement with the classical latitude diversity gradient hypothesis observed across many taxa24,25 (Fig. 2a). By contrast, predicted EcM fungal richness was lowest near the equator and showed more species-rich communities across northern latitudes and in southern regions of South America and Australia, which created an inverse latitudinal diversity gradient (Fig. 2b and Extended Data Fig. 2). Latitudinal patterns of AM fungal endemism showed a similar pattern as AM richness, but EcM fungal endemism had a moderate increase near the equator compared with EcM richness (Fig. 3). These relationships probably emerged through a complex interplay of host vegetation14,26, degree of host specificity11,27 and plant–soil nutrient economies28,29. Notably, these analyses highlighted that plant diversity is not a straightforward indicator for mycorrhizal fungal biodiversity at global scales20,26. This finding is a problem for the assumption that Earth’s mycorrhizal fungi will be automatically protected under new biodiversity initiatives, as current conservation policy tends to rely on plant and animal diversity metrics for determining priority areas and ecosystem value5,30.

Fig. 2: Global predictions and latitudinal trends of mycorrhizal fungal richness.
figure 2

a,c, Richness maps show the predicted number of AM fungal VT (a) and EcM fungal OTUs (c) per 100 m2 (pixels approximately 1 km2). Predicted richness values are capped at 45 VT per 100 m2 (AM) and 70 OTUs per 100 m2 (EcM) for visualization. Crosshatches are superimposed over areas that are underrepresented by the training data (highly extrapolated) where model predictions should be interpreted with caution. Plots show mean richness trends across latitude (shaded area is ±2 s.e.m.). b,d, Bivariate maps show the combination of pixel-level uncertainty (orange gradient) and extrapolation (purple gradient) of AM (b) and EcM (d) fungi richness predictions. Uncertainty is measured as the coefficient of variation across n = 100 bootstrapped model predictions. Extrapolation reflects the degree of environmental difference and geographical distance from samples in the training dataset. Histograms on the charts show the frequency of pixels in different uncertainty and extrapolation levels visualized here. Masked areas (grey) are sparsely vegetated zones and dense urban areas based on global land-cover data.

Full size image
Fig. 3: Global predictions and latitudinal trends of mycorrhizal fungal endemism.
figure 3

a,c, Endemism maps show the predicted rarity-weighted richness of AM fungal VT (a) and EcM fungal OTUs (c) per 100 m2 (pixels approximately 1 km2). Rarity-weighted richness is a unitless metric, and predicted values are capped at 0.26 (AM) and 1.5 (EcM) for visualization. These spatial predictions reflect the simulated high-sampling scenarios to limit unequal sampling effects on rarity patterns. Crosshatches are superimposed over areas that are underrepresented by the training data (highly extrapolated) where model predictions should be interpreted with caution. Plots show mean rarity trends across latitude (shaded area is ±2 s.e.m.). b,d, Bivariate maps show the combination of pixel-level uncertainty (orange gradient) and extrapolation (purple gradient) of AM (b) and EcM (d) fungi rarity-weighted richness predictions. Uncertainty is measured as the coefficient of variation across n = 100 bootstrapped model predictions. Extrapolation reflects the degree of environmental difference and geographical distance from samples in the training dataset. Histograms on the charts show the frequency of pixels in different uncertainty and extrapolation levels visualized here. Masked areas (grey) are sparsely vegetated zones and dense urban areas based on global land-cover data.

Full size image

Mycorrhizal fungal hotspots

If plant diversity is not a direct indicator of mycorrhizal fungal diversity, then protecting underground biodiversity requires explicit consideration that is based on the development of new high-resolution identification and monitoring approaches. For effective underground conservation, the identification of global hotspots of mycorrhizal biodiversity is particularly important. We therefore used our spatial predictions to locate areas of highly species-rich and endemic mycorrhizal fungal communities (hotspots). Hotspots were defined as pixels in the upper 95th percentile of predicted richness and rarity values globally, as per previous definitions25,31. This analysis resulted in richness cut-off values of 39.9 VT for AM fungi and 60.0 OTUs for EcM fungi, and unitless rarity cut-off values of 0.24 for AM fungi and 1.27 for EcM fungi.

Sampling bias can affect the relative rarity of species, which means it is important to determine whether predicted endemism hotspots are biologically real or an artefact of uneven global inventories of mycorrhizal fungi. We trained the AM and EcM fungal rarity machine-learning models with additional covariate layers of global sampling intensity. These layers were created using kernel density interpolation from sample coordinates in the SSU and ITS training datasets (Methods). This approach enabled us to compare empirical rarity predictions under current sampling intensities against predictions that simulated universally ‘high-sampling’ scenarios, as previously reported32. For AM fungi, there was substantial overlap between the empirical and high-sampling rarity hotspot predictions (Extended Data Fig. 3a), which indicated that current sampling efforts are generally capturing the distribution of endemic AM fungal taxa. However, there was large divergence in the empirical and high-sampling predictions of EcM rarity hotspots, particularly in tundra and tropical forest biomes (Extended Data Fig. 3b). This result suggests that increasing sampling efforts in these habitats will probably reveal new or overlooked hotspots of rare and endemic EcM fungi. For both mycorrhizal types, we used the high-sampling predictions for subsequent spatial analyses.

We predicted major hotspots of AM fungal richness and endemicity across savannas of the Brazilian Cerrado, tropical forests across Southeast Asia and Guinean forests in West Africa (Figs. 2a and 4). Our models showed that these areas may contain more than 45 AM fungal species (VT) per 100 m2. Species-rich communities of AM fungi were likewise predicted in parts of East and North China in ecosystems that are rapidly undergoing significant anthropogenic land-use changes33. Smaller regions, like the montane grasslands in Ethiopia, dry forests throughout Central America and Mediterranean woodland and scrubland in Europe, were also predicted to contain highly diverse collections of AM fungi (Fig. 2a). However, these areas were not predicted to be endemicity hotspots. Instead, AM fungal endemicity hotspots were predicted to occur in tropical and subtropical forests in the Congo basin and eastern Amazon basin (Figs. 3a and 4). Although montane grasslands had the most species-rich AM fungal samples in the model training data, this biome did not emerge as a substantial AM fungal richness or rarity hotspot at the global level. This result may be due to the low number of samples from these habitats or the extent and location of montane grasslands compared with other biomes.

Fig. 4: Arbuscular mycorrhizal fungal hotspots and global protected areas.
figure 4

Predicted richness and endemism hotspots (95th percentile of predictions) for AM fungi. Coloured areas on the map show richness hotspots (green), rarity hotspots (purple) and the overlap of richness and rarity hotspots (yellow). Black areas indicate non-hotspots. Bar graphs show the total hotspot size and percentage overlap with protected areas by biome, and grey bars reflect different IUCN management categories (I, most strictly preserved habitats; NA, unassigned category). The dashed line at 30% hotspot area protected reflects the ambitions of 30 × 30 target goals under the Kunming–Montreal Global Biodiversity Framework. ND, no data.

Full size image

We predicted EcM fungal richness hotspots throughout northern forest ecosystems (Figs. 2b and 5). These included much of the Siberian and Canadian boreal forest regions, temperate coniferous forests across Western United States and Canadian mountain ranges and temperate broadleaf and mixed forests in Central Europe and the North American Great Lakes region. We predicted that these places could have underground ecosystems averaging >100 EcM fungal species (OTUs) per 100 m2. However, most of these northern forest biomes were not identified as EcM fungal endemism hotspots. Instead, we predicted that tundra ecosystems contained the greatest EcM rarity levels at high latitudes (Figs. 3b and 5). We also identified areas of global EcM diversity hotspots—especially rarity hotspots—in tropical biomes, such as tropical moist forests in China’s Yunnan–Guizhou plateau and mainland Southeast Asia, montane rainforests across Indonesia and Guianan highland forests in Venezuela. Of particular note are the predicted EcM fungal endemism hotspots in tropical conifer forests in Central America that contain the highest richness of EcM host plant lineages34. EcM fungal richness and rarity hotspots in the Southern hemisphere were predicted to be concentrated primarily in the Andes Mountain range and the temperate mixed forests in southern Chile, the east coast of Australia and New Zealand (that is, all habitats of the Nothofagaceae plant family).

Fig. 5: Ectomycorrhizal fungal hotspots and global protected areas.
figure 5

Predicted richness and endemism hotspots (95th percentile of predictions) for EcM fungi. Coloured areas on the map show richness hotspots (green), rarity hotspots (purple) and the overlap of richness and rarity hotspots (yellow). Black areas indicate non-hotspots. Bar graphs show the total hotspot size and percentage overlap with protected areas by biome, and grey bars reflect different IUCN management categories (I, most strictly preserved habitats; NA, unassigned category). The dashed line at 30% hotspot area protected reflects the ambitions of 30 × 30 target goals under the Kunming–Montreal Global Biodiversity Framework.

Full size image

Mycorrhizal hotspot protections

To assess the extent to which mycorrhizal diversity is currently protected, we overlaid mycorrhizal hotspot pixels with protected areas defined by the World Database of Protected Areas35. We then calculated the total extent to which the hotspots were within current protected area delineations. This analysis revealed that the centres of mycorrhizal biodiversity were unequally distributed across biomes; nearly all had low protective coverage, and the management stringencies for protected hotspots differed among the mycorrhizal types (Figs. 4 and 5). Globally, we predicted that only 9.5% of mycorrhizal richness hotspots occur in protected habitats, including 5.1% (about 280,000 km2) of AM hotspots and 13.9% (around 756,000 km2) of EcM hotspots. Mycorrhizal rarity hotspots were protected at roughly twice the rate (22.9% overall), with 22.6% (about 1.2 million km2) and 23.2% (around 1.3 million km2) of AM and EcM fungal rarity hotspots overlapping with protected areas, respectively. This result is probably because high levels of endemism often result from geographical barriers that cause insularity. Moreover, protected areas are more frequently established in remote, hard-to-access areas where human pressures are anticipated to remain low, regardless of conservation action36.

The protected areas considered here cover roughly 16 million km2 and contain everything from strict nature reserves to managed resource areas (International Union for Conservation of Nature (IUCN) management categories I–VI and not applicable (NA) or undefined (NA/undefined)). Analyses of hotspot protections by management category revealed that most protected AM fungal hotspots are in the least-strictly preserved habitats (Fig. 4). Specifically, 76.4% (richness) and 72.8% (rarity) of protected AM hotspots are under categories V, VI or NA/undefined. By contrast, protected EcM fungal hotspots are mostly under the strictest preservation levels, including nearly half of protected EcM richness and rarity hotspot area in IUCN categories I–III (Fig. 5).

At the continent scale, predicted mycorrhizal fungal hotspots in Asia have the lowest protection for both AM fungi (richness = 2.2% and rarity = 7.5% protected) and EcM fungi (richness = 11.3% and rarity = 17.2% protected), even though these hotspots are concentrated in different biomes for each mycorrhizal type (tropical moist forests compared with boreal and tundra systems; see the Zenodo archive (https://doi.org/10.5281/zenodo.14871588)37 for continent-level results). Europe has the highest protected AM fungal richness hotspots (19.6% protected), which we predicted to occur mainly in Mediterranean forests across the continent. Australia and Oceania showed the largest protected area overlap with AM fungal rarity hotspots (35.7% protected), primarily in tropical grasslands. Predicted EcM fungal richness hotspots are also the most well protected in Australia and Oceania (40.2% protected), which largely occur throughout temperate broadleaf forests. Both Australia and Oceania and South America have the largest protected area overlap with predicted EcM fungal endemism hotspots (46% protected).

Environmental predictors

Given the distinct hotspot locations and latitude–diversity relationships, we used shapley additive explanations (SHAP) to analyse which of the 24 environmental covariates were the most important predictors of mycorrhizal fungal richness and rarity. This method quantifies the additive contribution of each feature of a machine-learning model (for example, temperature and rainfall) on its predicted values, as opposed to measuring feature importance based on changes in model performance38. Therefore, SHAP is a useful method for interpreting the magnitude and directionality of predictors in our machine-learning models with respect to changes in mycorrhizal diversity. Moreover, in exploratory tests, the inclusion of spatial predictors (Moran Eigenvector layers) generally did not change the identity or order of important variables in either of the richness models. This result confirmed that the rankings of the most important predictors of mycorrhizal fungal richness are unlikely to be biased by spatially autocorrelated processes not included in the model (Supplementary Figs. 1 and 2).

Among the climate features, temperature (annual mean temperature and maximum temperature of warmest month) was generally the most important climatic predictor across all AM and EcM models (Extended Data Figs. 4 and 5). By calculating mean absolute SHAP values, these variables were two to five times more influential than other predictors, contributing ±0.07 VT to each AM fungal richness prediction and ±0.11 OTUs to each EcM fungal richness prediction on average. Specifically, higher temperatures predicted greater AM fungal richness and lower EcM fungal richness. Potential evapotranspiration was also an influential climate predictor for rarity models, with drier areas associated with higher AM and EcM fungal endemicity (Extended Data Figs. 4 and 5). However, the evapotranspiration trends were opposite for EcM diversity metrics, whereby drier areas were associated with lower EcM fungal richness. Using paleoclimate data, EcM fungal rarity was strongly and positively associated with climatic stability since the Pliocene (Extended Data Fig. 5; mean absolute SHAP value of ±0.04 OTUs), which indicates that current endemicity patterns reflect the evolutionary history of EcM fungal symbioses over millions of years. These findings add to recent work showing a link between climate and soil fungi at global scales17,39. Moreover, the results probably reflect joint climate tolerances of mycorrhizal fungi and their plant hosts, as well as climatic controls on soil organic matter decomposition13,15,17,28,39.

The importance of climatic controls in predicting mycorrhizal diversity was further supported by strong links between fungal richness and soil organic carbon. Greater soil organic carbon was associated with higher EcM richness and lower AM richness (Extended Data Fig. 4; mean absolute SHAP values of ±0.11 OTUs and ±0.08 VT, respectively). This result is consistent with past findings that soil organic carbon stocks are positively associated with EcM plant dominance14,40 and probably relates to metabolic differences between AM and EcM fungi in enzymatic capabilities and plant host litter chemistry9,41. In contrast to EcM, AM fungal richness peaked in the tropics, where climate and other biotic conditions favour rapid mineralization of plant litter that may drive AM fungal dominance over EcM symbiosis. Soil organic carbon was not among the top 10 most important predictors of EcM fungal endemism (mean absolute SHAP value of ±0.01 OTUs; Extended Data Fig. 5), which helps explain why we predicted substantial EcM rarity hotspots across tropical forests with generally lower soil carbon stocks than temperate or boreal forests12. Notably, soil pH (as a geospatial covariate layer) was not among the most important predictors identified in AM or EcM models, despite past work describing its importance in shaping overall soil fungal diversity patterns and AM fungal niche traits16,42.

In addition to climate, we were interested in testing how human activities affect diversity patterns. Given that 95% of the Earth’s terrestrial surface is affected by some level of human modification43, we used the summed percentage of cultivated and managed areas and urban and built-up habitats to measure anthropogenic land-cover (dense urban centres were masked from predictions; Methods). Anthropogenic land-cover emerged as an important and positive predictor of AM fungal richness (mean absolute SHAP value of ±0.07 VT; Extended Data Fig. 4). Although sampling bias in the SSU dataset from human-modified habitats may be partially responsible, disturbed ecosystems also tend to contain more ruderal, wind-dispersed AM fungal species that can increase local species richness at the expense of homogenizing regional-scale species composition44,45. Human-mediated land modification was a less important feature in the AM fungal rarity model (mean absolute SHAP value of ±0.004 VT; Extended Data Fig. 5) than in the AM fungal richness model, which could be interpreted as supporting the regional homogenization hypothesis. An open question is whether changes in AM fungal-associated vegetation, such as new crops and exotic plants, create more opportunities for diverse plant–AM fungal interactions to occur14,46. Different types of human disturbance are likely to have different effects on AM fungal diversity, and future research on AM fungi in human-modified habitats would benefit from analyses with more specific spatial layers on anthropogenic landscapes.

Aboveground plant biomass was the top predictor for EcM richness and rarity models (richness mean absolute SHAP value of ±0.13 OTUs; rarity mean absolute SHAP value of ±0.05), with higher biomass positively predicting higher EcM fungal richness and rarity. In general, this result reflects the dominance of EcM fungi in forest ecosystems, including estimates that 60% of all tree stems on Earth form EcM symbioses13. Elevation and slope were also among the most important predictors of EcM rarity (mean absolute SHAP values of ±0.04 and ±0.03, respectively; Extended Data Fig. 5), and we predicted hotspots of endemic EcM fungal communities across mountainous terrain in the Andes, Sierra Madre, Sierra Nevada and Cascade Mountain ranges (Fig. 5). For EcM fungi, aboveground biomass and topography may be more reliable indicators of richness and rarity than plant diversity. This finding is encouraging given that remote-sensing tools are increasingly able to detect complex forest structures of significant ecological value for protection47. Our results further underscore the critical planetary benefits of high-biomass forests that are supported by and foster endemic, species-rich EcM fungal communities.

Uncertainty analyses

For spatial predictions of biodiversity to be useful in conservation policy, it is important to measure and communicate the uncertainty associated with each prediction22. This is especially true for microbial biodiversity predictions that are more challenging to verify compared to plant and animal distributions48. To do so, we calculated two pixel-level metrics of uncertainty (Methods). In brief, we defined statistical uncertainty as the coefficient of variation in the predictions across n = 100 bootstrap samples of the training data to create a confidence interval around the prediction mean of each pixel. Next, we quantified extrapolation as the extent to which a given pixel departs from the sampling locations and the range of environmental conditions in the training data.

The uncertainty range of richness models were similar among mycorrhizal types. The coefficient of variation (the relative dispersion of pixel values around predicted means) ranged mostly between 0 and 0.06 for AM and EcM richness models (Fig. 2). AM fungal rarity models showed slightly higher uncertainties than for EcM fungal rarity interpolations (0.10–0.20 coefficient of variation; Fig. 3). This result is probably caused by differences in the characteristics of AM and EcM fungi that may affect local-scale sample variation or species detection (for example, host specificity and dispersal ability) and differences in the underlying datasets (for example, sample sizes and coverage across environmental gradients). The predictive accuracy of both fungal richness models was comparable (random cross-validation R2 = 0.61 for AM and R2 = 0.63 for EcM), which indicated that both spatial predictions are robust despite the differences in training dataset size. Spatial patterns of uncertainty were different for both mycorrhizal types and diversity metrics (Extended Data Fig. 6). For example, EcM fungal richness models showed approximately double the uncertainty in desert, grassland and tropical forest biomes compared with AM models. AM fungal model uncertainties were highest in areas with low predicted richness, such as desert and tundra regions (Fig. 2b).

Locations with high extrapolation reflected poor representation in the model predictors and were primarily in regions that are consistently underrepresented in global soil biodiversity data49 (Extended Data Figs. 7 and 8). For AM fungi, areas that require greater extrapolation included northern India, the Tibetan plateau, Indonesia and parts of the Amazon rainforest (Figs. 2 and 3). Therefore, extra caution is required when examining AM fungal predictions in these regions. Overall, the EcM predictions were less dependent on model extrapolation because they were built from a larger training dataset (over five times more ITS samples distributed across a wider range of environmental conditions than the SSU samples used for AM fungi). However, there were small regions with a high degree of EcM model extrapolation in Sub-Saharan Africa, northern Canada and Southeast Asia (Figs. 2 and 3).

Important critiques of machine-learning geospatial approaches have been recently noted21. We performed the following actions to address these concerns: (1) accounted for spatial autocorrelation variables in the model (Supplementary Figs. 1 and 2); (2) created spatial blocks to limit the distance between sample and prediction locations for additional cross-validation (Supplementary Fig. 3); and (3) provided clear data on prediction uncertainty (Figs. 2 and 3 and Extended Data Figs. 7 and 8). Moreover, we conducted multiple cross-validation procedures as there is currently no consensus on best practices for validating spatial models and measuring their inaccuracies (Methods). Overall, our models performed similar to (or better than) the accuracy of recent machine-learning predictive mapping approaches used for other soil organisms (Methods and Extended Data Fig. 9).

Conclusions

Our high-resolution, global-scale maps of AM and EcM fungal diversity suggest that <10% of mycorrhizal richness hotspots currently exist in protected areas. The same analysis for non-fungal taxa showed that both plant and animal biodiversity hotspots are much better protected. Replicating the analysis with spatial data of other taxonomic groups showed roughly 3-fold higher protection of richness hotspot areas for vascular plants (27% protected), trees (34% protected), ants (28% protected) and vertebrates (33–41% protected for amphibians, birds, mammals and reptiles), which is probably because threatened species in these groups drive most conservation interventions31,50. Endemic mycorrhizal fungi may be faring better: overall 22.9% of predicted rarity hotspot areas overlapped with protected areas, similar to the protection coverage of vertebrate rarity centres (19–21% protected for amphibians, birds, mammals and reptiles). To fully realize the 30 × 30 target goals of protecting Earth’s biodiversity under the Kunming–Montreal Global Biodiversity Framework, it is necessary to identify underground conservation priorities, set monitoring benchmarks and create specific restoration plans51. For example, our data can be used to help develop land-management strategies to protect and maintain mycorrhizal diversity reservoirs52, including identifying soil-management practices developed over millennia by local populations53. We also recommend that any use of these predictive biodiversity maps in a specific locale should be cross-referenced with the model uncertainty and extrapolation layers in the same area to maximize positive conservation outcomes22.

Our results should be seen as a first step towards understanding the baseline protected levels of Earth’s mycorrhizal fungal diversity. Overlaying protected areas on biodiversity hotspots does not tell us how successful these conservation strategies are at maintaining or improving mycorrhizal fungal biodiversity. Because IUCN management categories vary in strictness of habitat preservation, it is also important to analyse hotspot protections on the basis of these management categories, few of which are entirely free of human pressures. A critical next step will be to measure the effectiveness of protected areas to mitigate fungal biodiversity loss, as has been done using counterfactual analysis for other organisms54. This strategy will help determine which protected area conditions and objectives best deliver conservation benefits to mycorrhizal fungi—such as focusing on rare or threatened species, unique ecosystems and their entire communities, vital ecosystem services and/or locations of cultural and social significance. Conservation efforts will also be more effective by continuously monitoring mycorrhizal fungi, as they have known vulnerabilities to environmental change and disturbance19,52, and there is a need to track their recovery following conservation and restoration actions.

Owing to insufficient data, we were unable to robustly model ErM or orchid mycorrhizal fungal richness patterns. For instance, the average ErM fungal richness per sample was 0.6 species, and <3% of samples contained more than 4 ErM fungal species. These are both important fungal symbioses to understand in a global context because of their unique contributions to plant ecology and ecosystem functions11. Additional sequencing projects and data-mining efforts will facilitate large-scale analyses of these mycorrhizal types. A second issue is that using SSU data to understand the biogeography of AM fungal richness and rarity may be problematic owing to overly conservative species definitions that could underrepresent the true diversity of this mycorrhizal type (Methods). AM fungi have a unique genetic organization with high intragenomic variation that makes their sequencing and species definition more difficult than for EcM fungi, and only 332 AM fungal species have been formally described8. Alternative approaches, such as long-read sequencing of both environmental AM fungi and culture collections and the use of multiple metabarcoding target regions, may provide additional insights into the definition and distribution of AM fungal biodiversity55.

Although our models performed similar to (or better than) those for other soil organisms (for example, nematodes, springtails and earthworms; Methods), they are still an imperfect representation of mycorrhizal patterns in nature. In particular, 1 km2 is a relatively coarse spatial scale to quantify the average soil mycorrhizal richness expected per 100 m2. There can be large landscape heterogeneity in a pixel of this size, and soil fungal sampling is usually restricted to a smaller collection grid and sequencing <1 g of soil. Currently, 1 km2 is the highest spatial resolution possible for making reliable global predictions given the coverage of relevant covariate layers, but this technical limitation will be overcome as more global environmental datasets are created at 250-m and 30-m resolution, closer to the scale of field-sampling methods. We are also actively exploring how much within-pixel variation affects model accuracy as part of an ongoing effort to independently ground-truth these spatial predictions. Ground-truth validation is particularly important for the following regions: (1) under-sampled environments poorly represented in the model; (2) regions with high uncertainty despite good training data coverage; (3) areas with high landscape-related heterogeneity (for example, steep mountains); and (4) areas that are particularly vulnerable to rapid climate change and habitat loss from human activities.

Given the importance of these organisms for the productivity of ecosystems and the functioning of landscapes, it is surprising that mycorrhizal diversity remains highly underrepresented in conservation agendas. Explicit consideration of mycorrhizal biodiversity hotspots like those identified here can help direct and implement conservation strategies towards protecting the most diverse and endangered underground ecosystems.

Methods

Fungal data

Fungal occurrence records were generated from the data mining of published ITS and SSU sequencing studies collected for the GlobalFungi and GlobalAMFungi databases (see previous studies6,8 for details) and the ITS region from full-length sequences in the Global Soil Mycobiome consortium database7. For EcM occurrences used in this study, we considered only samples of the ITS2 barcode deposited in the GlobalFungi dataset as this marker is less biased by length variability compared with ITS1 (refs. 56,57). Moreover, ITS2 is better represented in the source database and ITS1 samples failed technical validation (see below). In brief, raw sequences and metadata from 255 ITS studies (representing the fourth release of the GlobalFungi database) were processed using an established bioinformatic pipeline that incorporates sequence quality checks, extraction of full ITS2 fungal regions using ITSx (v.1.1.2)58 and clustering into OTUs at 97% similarity level with subsequent exclusion of global singletons and chimeric sequences using USEARCH (v.11.0.0667)59. We used BLASTN searches against UNITE (v.8.3, released 10.5.2021) to assign putative taxonomy to non-singleton OTUs. Default BLAST parameters were used, and representative sequences were considered to belong to the closest BLAST hit genera in the case of >92% similarity and >95% coverage. Considering that the majority of EcM fungal genera do not include species belonging to other functional guilds, we found that these BLAST results criteria were reasonable. We excluded OTUs represented by sequences with e value > 10–50. The resulting OTU table with taxonomy assignments was compared against the FungalTraits database (v.1.1) to subset EcM fungi60.

In early tests, we used the ITS dataset to model AM fungi, but their low abundance detected from using this target region was leading to clearly erroneous spatial predictions (for example, zero AM species predicted in regions known to be dominated by AM vegetation). As a result, SSU data from the GlobalAMFungi database were used to estimate AM fungal species richness8. In brief, raw sequences and metadata from 45 SSU studies (representing the first release of the GlobalAMFungi database) were subjected to sequence quality checks, trimming of the sequences to the V4 region of the SSU and assignment of the sequences to VT from the MaarjAM database (v.5.6.2019 release)61. We also used BLASTN with a required sequence coverage of ≥98% and sequence similarity ≥97%, based on the most recent release of type sequences of VT. VT were defined on the basis of phylogenetic grouping of SSU sequences at roughly species-level diversity (which may result in conservative taxonomic designations of AM fungal taxa55,62). Molecular identification based on VT is typically used for AM fungal metabarcoding studies using the SSU region, and OTUs are more traditionally used for general fungal community sequencing (including EcM fungi) based on the ITS region. Both VT and OTUs are commonly used to approximate species-level assignments in sequencing studies of mycorrhizal fungi. The EcM OTU data used for this analysis is the result of clustering of GlobalFungi sequences followed by taxonomic assignment. By contrast, the AMF pipeline results in more conservative estimates of diversity, as SSU sequences are assigned to representative VT using stringent BLAST parameters. This results in the discarding of sequences without a good match to these known VT, which is necessary to filter non-fungal sequences (for example, Annelids) but also probably removes undescribed AM fungal taxa. In total, 2.8 billion fungal sequences were processed across 24,982 samples, which identified 332 VT of AM fungi (encompassing 86% of defined VT in the MaarjAM database) and 164,439 EcM fungal OTUs.

Richness estimates

We used analytical estimators for the rarefaction and extrapolation of species richness to measure the number of mycorrhizal species (OTUs or VT) in each sample63. This approach creates a sequencing depth-based rarefaction and extrapolation sampling curve (that is, a sample-specific species accumulation curve), with diversity estimates and 95% confidence intervals calculated at curve asymptotes (using the R package iNEXT64). Extrapolation end points were computed as twice the sequencing depth for each sample (default setting). Use of the rarefaction and extrapolation approach to estimate mycorrhizal richness enables a robust comparison of mycorrhizal patterns across studies of multiple sequencing technologies (although variation in error rates may persist), primer sets and sequencing intensities. The extrapolated richness estimates here are within similar ranges to past studies that estimated global diversity of mycorrhizal fungi26,39. Moreover, sample richness values were highly consistent among 96%, 97% and 98% clustering similarity thresholds (Supplementary Fig. 4).

We removed outliers in mycorrhizal richness estimates before spatial modelling. Two Australian studies with ITS samples from desert and Mediterranean biomes had unusually high EcM OTU richness estimates and standard deviations (two orders of magnitude higher than other Australian ecoregions65,66). Moreover, these studies have previously been identified as potentially inaccurate based on a recent database comparison6. All samples from these two studies were removed (n = 2,070). We then filtered samples by biome by removing estimated richness values that were more than five times the interquartile range higher than the biome-level median estimate. We only filtered values at the highest end of the distribution to avoid potentially removing ‘true’ estimates of low or zero mycorrhizal richness at a given location. See Supplementary Table 1 for a summary of outliers removed per biome.

We also calculated rarity-weighted richness to estimate the relative endemism of mycorrhizal fungal communities. Using samples from the outlier filtered dataset (see above), we created a species-level prevalence score as the number of samples for which each OTU or VT occurs and then summed the inverse prevalence score of all species present in each sample67. This rarity metric is useful for identifying site endemism or areas with a high concentration of rare taxa23. Rarity-weighted richness was originally used with a geographically unbiased dataset for which each pixel contained information on bird observations67, but has since proven useful in other contexts to predict patterns of relative species endemism32,68. To account for the possibility that geographically uneven sampling efforts affect fungal species prevalence scores and biases predictions of fungal endemism, we simulated a global high-sampling scenario for rarity spatial models (see below).

Geospatial modelling, validation and analysis

Spatial predictions of mycorrhizal fungal richness and rarity were created using a random forest modelling approach (modified from ref. 69). We first sampled a collection of >72 global environmental covariates at each of the locations in the dataset. To reduce overfitting and variance inflation, we removed highly correlated variables such that the final set of predictors comprised 24 environmental variables. These contained macroclimatic, soil texture and physicochemical information, vegetation, radiation and topographical indices and anthropogenic variables. Details of all predictors, including descriptions, units and source information, are previously described70 and provided in Supplementary Table 2. Variables describing soil structure and physicochemical properties were obtained from SoilGrids71, limited to the upper 5 cm of soil. Climate information (that is, mean annual temperature, annual precipitation, monthly maximum temperature and precipitation seasonality) was obtained from CHELSA72, and climate stability over palaeoecological timescales was obtained from the Climate Stability Index73. Spectral vegetation indices (that is, MODIS NPP product MOD17A3HGF V6.1, averaged annually) were obtained from the Google Earth Engine Data Catalogue. We used aboveground biomass data from CDIAC74. We used EarthEnv75,76 to collect consensus land cover information (including human development percentage) and eight topographical layers that capture a broad range of complex terrain features (for example, geographical isolation). The predicted plant diversity layer was obtained from a previous study77. The potential evapotranspiration layer was obtained from CGIAR78. Resolve Ecoregion classifications were used to categorize sampling locations into biomes79. All spatial covariate layers were reprojected and resampled to a unified pixel grid in EPSG:4326 (WGS84) at 30-arcsec resolution (approximately 1 km2 at the equator). Areas covered by permanent snow or ice (for example, the Greenland ice cap and glaciated mountain ranges, which were identified using SoilGrids71), barren land cover (sparsely vegetated regions defined in EarthEnv75,76) and highly urban and built-up areas were excluded from the analyses. Antarctic areas were excluded from analyses owing to limited coverage of covariate layers in the region.

To harmonize the data across the different experimental approaches of the original studies, we included six types of project-specific variables. This included two continuous variables (sample area and soil mass for DNA extraction) and the following four categorical variables: sequencing platform (Illumina, IonTorrent, PacBio, 454Roche and DNBSEQ-G400); target gene region (ITS1 and ITS2); sample type (soil, topsoil and rhizosphere); and primer set (37 different kinds). As an intermediate step, we created random forest spatial models (described below) to derive predicted values of mycorrhizal richness for comparison to observed data at the same geolocation. We filtered samples from the categorical project-specific variables that failed technical validation by comparing observed with predicted richness trends (failures were defined by zero correlation). This process removed all DNBSEQ-G400 samples (n = 3,738), all ITS1 samples (n = 33; these samples also had corresponding ITS2 sequence data that were retained) and 14 different primer sets (total n = 809), which were filtered before running the final models. To create spatial predictions, we harmonized the project-specific variables to the following most common levels: sample area = 100 m2; soil mass for DNA extract = 0.5 g; sequencing platform = Illumina; target gene = ITS2 (implicit because ITS1 samples were removed); sample type = soil; and primer set = ITS3–ITS4 for EcM. Before modelling, all categorical project-specific variables were transformed to binary variables (that is, one-hot encoding or dummy variables). See Supplementary Figs. 5–10 for more details on technical validation and data filtering.

After outlier removal and validation, the SSU dataset contained 3,234 samples (332 total VT) for AM fungal analysis, and the ITS dataset had 17,519 samples (41,086 total OTUs) for EcM fungal analysis. To create the training datasets and to reduce potential overfitting, we used only distinct observations of mycorrhizal fungal richness (that is, samples with non-identical richness or rarity values). This meant that when multiple samples fell within the same 1-km2 pixel, we removed duplicate richness estimates to ensure that each observation contributed equally to the learning process of the model (multiple identical observations caused the random forest models to overfit towards those combinations of richness levels and environmental variables). Using only distinct observations enabled us to retain important within-pixel variation rather than aggregating across multiple richness estimates per pixel. To deal with the zero values in the dataset, rarefied richness values were log10[x + 1] transformed before modelling. In the random forest models, we fixed the number of trees at 250 and explored different combinations of hyperparameters by varying the number of variables considered at each split (4, 6, 8, 10 or 12) and the minimum number of samples per leaf (2, 4, 6, 8, 10 or 12). In total, 30 hyperparameter combinations were tested for each AM and EcM response variable.

There is ongoing debate about the best methodology for validating spatial models80,81,82,83. As a result, we tested each model using both random cross-validation and k-fold nearest neighbour distance matching cross-validation (kNNDM)84. Random folds were assigned stratified per biome to ensure equal coverage of each fold. For both AM and EcM, the final predictions are an ensemble (mean) of the top ten best performing models based on coefficient of determination (R2) with kNNDM ten-fold cross validation. This approach resulted in the following predictive accuracy metrics for richness models: AM random cross-validated R2 = 0.61; AM kNNDM cross-validated R2 = 0.20; EcM random cross-validated R2 = 0.63; and EcM kNNDM cross-validated R2 = 0.28 (Extended Data Fig. 9a). The predictive accuracy metrics for rarity models were as follows: AM random cross-validated R2 = 0.56; AM kNNDM cross-validated R2 = −2.55; EcM random cross-validated R2 = 0.61; and EcM kNNDM cross-validated R2 = 0.19 (Extended Data Fig. 9b). Overall, these performance estimates were similar to (or exceeded) the predictive accuracy of global machine-learning models for other soil organisms69,85,86. The poor performance of the AM fungal rarity model based on kNNDM cross-validation suggests there may be spatial or environmental structures in the dataset that the model is unable to detect, which is possible if the VT dataset fails to sufficiently capture AM fungal endemism patterns because of conservative taxonomic assignments from the AM fungal reference database61. We used the SHAP algorithm to interpret the importance and directionality of environmental features38. Standard feature importance algorithms, like the ones implemented in random forest or boosted regression tree models, measure feature importance as a change in model performance and are more susceptible to feature collinearity. By contrast, SHAP handles feature collinearity relatively well and is based on the magnitude of feature contribution. We selected SHAP as the most suitable method here because of the possibility of feature collinearity among the environmental predictors and because we were interested in the effect of the features on the predicted value (as opposed to the effect on model error) to interpret the importance and directionality of the features.

As we expected the model performance to decrease with distance to training locations, we also performed spatially buffered leave-one-out cross-validation (SLOO-cv)80, a computationally intensive approach in which a separate model is trained for every unique location in the dataset and leaving out locations in a range of buffer sizes. At larger buffer sizes (that is, >500 km), SLOO-cv R2 values were in line with kNNDM cross-validation R2 values. To transform these findings into a visual representation, we plotted the R2 values against the distance to the nearest sampling location to create a spatial product (Supplementary Fig. 3). The resulting map was consistent with our previously described extrapolation map.

To generate a spatial understanding of our predictive accuracy, we created 100 bootstrap samples by resampling the training datasets with replacement using biome-based stratification. Leveraging the hyperparameter settings of the best performing random forest model, we generated 100 global prediction images. These were subsequently used to calculate a bootstrap coefficient of variation (derived by dividing the standard deviation by the mean) and 95% confidence intervals for each pixel (Extended Data Figs. 7 and 8). To locate environmental conditions and corresponding geographical regions that are underrepresented in the training data, we first converted the data into the principal component space. We then selected the first 13 and 14 axes that cumulatively accounted for 90% of the total variance for the models for AM and EcM, respectively. We evaluated whether raster pixels were associated with environmental conditions internal or external to the convex hull circumscribing the data points along each pairwise combination of principal component axes. We defined the degree of model extrapolation as the proportion of total pairwise combinations for which a pixel lies outside the corresponding convex hull. Finally, we combined this environmental extrapolation data with a map of geographical distance between sampling locations (distance and environmental extrapolation maps were combined in a 2:1 ratio, after scaling each from 0 to 1) to produce a global spatial assessment of the representativeness of our datasets (Extended Data Figs. 7 and 8). The extrapolation map was then used to mark pixels that were most underrepresented by the training data using a 5% extrapolation cut-off value (that is, areas with <95% coverage in training data space).

To explore possible spatial dependency in the data, we fit semi-variogram models per mycorrhizal type on richness estimates and model residuals using the R package automap (v.1.1-9)87. For AM fungi, we observed autocorrelation in the extrapolated richness data up to about 900 km; for EcM fungi this range reached about 550 km. This level of spatial autocorrelation in response variables is unsurprising given the scale of our georeferenced datasets, similar to a recent analysis showing that spatial weights of around 800 km were best at handling spatial autocorrelation in a global plant diversity dataset77. For model residuals, semi-variograms showed weak, although significant, spatial dependency for both AM and EcM fungal data (Supplementary Fig. 2). We then performed a Global Moran’s I-test with permutations calculated using the R package sfdep88, including five nearest points as neighbours and a Gaussian kernel function for building spatial weight matrices. For the richness estimates, spatial autocorrelation was positive and highly significant (Moran’s I of 0.71 for AM and 0.64 for EcM), whereas spatial autocorrelation was low in model residuals for both fungal types (Moran’s I of 0.11 for AM and 0.23 for EcM; Supplementary Fig. 1), similar to a previous study that reported limited effects of geographical distance on fungal alpha diversity39. To incorporate spatial processes into the prediction model, we applied spatial eigenvector mapping using distance-based Moran’s eigenvector maps (dbMEMs)89. We calculated dbMEMs using the R package adespatial (v.0.3-21)90, with the custom function quickMEM91. The inclusion of dbMEMs in the modelling approach showed mycorrhizal hotspots predicted in the same areas but reduced some finer-scale variation in richness predictions in the non-spatial models (Supplementary Fig. 10). These results led us to conclude that the models without dbMEMs capture the most variation; however, there might be some fine-scale spatial processes that are not explained for AM fungi here.

We evaluated how uneven sampling density might change rarity patterns by comparing model predictions that used current sampling intensity with those that used a simulated globally homogenous and high-sampling intensity, as previously described32. We created 1-km2 resolution sampling intensity layers for the SSU and ITS training datasets using kernel density interpolation from sample coordinates with a 5° radius and uniform decay rate in QGIS (v.3.40). These layers were rescaled 0−1 and included as covariates in the rarity random forest models to make empirical predictions of AM and EcM fungal rarity on the basis of the current distribution of sampling efforts. We then extrapolated rarity models under a global and equally high-sampling scenario by setting all pixels in the sampling intensity covariate layer to the maximum value of 1. Comparing patterns from the empirical and high-sampling rarity models showed where future sampling may reveal overlooked and under-realized hotspots of mycorrhizal fungal rarity32. For all analyses (for example, latitude trends and hotspot protections), we used the high-sampling rarity models to avoid biases in endemism predictions that were based on sampling intensity.

To measure how mycorrhizal richness and rarity varies with latitude, predicted richness and rarity pixel values were averaged at every 0.2 latitude degrees within 1°-wide longitude bands92 using functions in the R package Tidyverse (v.2.0.0)93. Correlations between latitude and estimated mycorrhizal richness and rarity of samples are also provided in Extended Data Fig. 2. Mycorrhizal hotspots were defined by setting a cut-off at the 95th percentile of predicted richness values. These hotspot pixels were then overlaid with the World Database of Protected Areas cropped to spatial predictions here (WDPA; www.protectedplanet.net)35 before calculating the total extent to which mycorrhizal hotspots in different biomes fall within current protected area delineations. Hotspot overlap with protected areas was calculated for each IUCN management category (terrestrial WDPA categories I–VI and unassigned). To understand mycorrhizal hotspot protections relative to other organisms, we compiled recently published spatial data on the species diversity of vascular plants25, trees94, ants32 and four groups of vertebrates (amphibians, birds, mammals and reptiles; IUCN95). The same hotspot analysis described above was applied to these spatial layers, and the total amount of richness and rarity centre overlap in protected areas was calculated for comparisons. All spatial analyses were based on raster data in Equal Earth projection at 1-km2 pixel resolution, and maps were visualized in Robinson projection.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.