Incorporating evolutionary and threat processes into crop wild relatives conservation
We applied a modified version of a planning framework for CWR conservation25,26 which has been used by numerous countries of Europee.g.29,63,64, Americae.g.65, Africa30 and Asia66,67. We addressed the following main steps of the toolkit (see Spanish version49): (i) CWR checklist, i.e., creating a list of CWR taxa distributed in an area (Supplementary Data 1), (ii) CWR inventory, i.e., taxa selection and collation of ancillary data, including taxonomic data (Supplementary Data 2), (iii) taxa extinction risk assessment (Table 1, Supplementary Data 3), and (iv) a systematic conservation planning assessment, i.e., spatial analyses to assess conservation areas (Fig. 1). We only provide a brief description of steps i-iii, as these are thoroughly described in Goettsch et al.2. Here, we focus on the systematic conservation planning assessment, introducing an approach in order to identify conservation areas for CWR that account for genetic differentiation in a spatially explicit way, through the use of proxies of genetic differentiation (Fig. 1).During the process -framed under the project “Safeguarding Mesoamerican crop wild relatives” (https://www.darwininitiative.org.uk/project/23007/)- more than 100 experts from academic, governmental, and non-governmental organizations from El Salvador, Guatemala, Honduras, Mexico, the UK, and IUCN participated in six workshops, shared data, and provided fundamental knowledge and feedback at each project stage to ensure accurate, reliable and robust information for next steps. The checklist, inventory and risk assessment were collaboratively developed between partners of El Salvador, Guatemala, and Mexico (hereafter, Mesoamerica; Goettsch et al.2). The spatial analysis to identify areas for in situ and ex situ conservation of CWR was done independently by each country.To assess conservation areas of CWR in Mexico, we developed proxies of genetic differentiation that account for evolutionary processes by including historical and environmental drivers of genetic diversity (see the Methods section ‘Proxies of genetic differentiation’). In addition, we used criteria such as information on taxon-specific tolerance to human-modified habitats and IUCN extinction risk category. We applied a systematic conservation planning approach and performed spatial analysis using the software Zonation50. We compared different scenarios to represent genetic diversity of CWR based on potential species distribution models (SDM) and proxies of genetic differentiation.Study areaMesoamerica is a cultural region encompassing the territories of Belize, Guatemala, El Salvador, the southern part of Mexico and parts of Honduras, Nicaragua and Costa Ricasee 2. In this study, we also included the dry areas of northern Mexico that are part of Aridamerica68 and the Nearctic biogeographic realm69 to account for the full extent of the geographic range of many taxa included in the extinction risk assessment2.For the assessment of conservation areas, we focused on Mexico, which is one of the most biodiverse countries in the world70. The Mexican territory covers 80% of the landscapes of the region called Mesoamerica. Its high biological diversity is attributed to its geographic, topographic, climatic, geological and cultural characteristics, which, among other factors, shaped the distribution of an extraordinary variety of ecosystems and species with high levels of endemism and species turnover among different regions32,71,72,73. In particular, the high genetic variation within populations of landraces and CWR is the result of past and ongoing sociocultural processes occurring in a wide range of distinct environmental conditions74,75.(i) CWR checklist and (ii) CWR inventoryThe compiled CWR checklist included ~3000 species and subspecies of 92 genera and 45 families of plants that belong to the same genus of a crop cultivated in Mesoamerica, or wild plant collected for food or other uses in the region (Supplementary Data 1).The first set of criteria were established in preparation for the first stakeholder workshop. The following criteria were applied at the genus level to compile the CWR inventory: (1) occurrence of wild relatives of cultivated plants or crops that were domesticated in Mesoamerica; (2) existence of research groups working on taxa that could support the extinction risk assessment; and (3) relation to a crop of economic and nutritional importance at local, national and regional levels, or cultivars known to require genetic improvement.To narrow the list for the inventory and extinction risk assessment, similar criteria were agreed upon in the same workshop and applied at the species level: (1) native distribution in Mesoamerica, incl. Aridamerica; (2) related to a crop of economic or social importance based on production and nutritional value; (3) related to a taxon for which Mesoamerica is the center of origin or domestication; (4) constitutes part of the primary or secondary gene pool, and in some cases the tertiary gene pool76. The primary gene pool consists of wild plants of the same species as the crop and thus their mating produces strong fertile progeny. The secondary gene pool is composed of wild relatives distinct from cultivated species but closely related as to produce some fertile offspring (same taxonomic series or section in the absence of crossing and genetic diversity information, see the ‘taxon group’ concept proposed by Maxted and collaborators77, Supplementary Note 5). The tertiary gene pool (same subgenus in the taxon group concept) corresponds to CWR that are more distant relatives to the taxa of the primary gene pool, but can have important adaptive traits which can be used with specific breeding techniques. This provided a preliminary list of 514 CWR taxa related to avocado, cotton, amaranth, cocoa, squash, sweet potato, chayote, chili pepper, cempasuchil, bean, sunflower, maize, papaya, potato, vanilla, and yuca (Supplementary Data 2).The list had to be further reduced due to time and funding restrictions to include those genera which when added together would include no more than 250 taxa, and that the taxonomic groups could be comprehensively assessed and their taxa evaluated throughout their entire range. Thus, not all species in the group necessarily met the criteria previously mentioned. See the final Mesoamerican CWR inventory in Supplementary Data 3; see summary in Table 1.(iii) Taxa extinction risk assessmentFull methodological details and results of this section are described in Goettsch et al.2. Summarizing, during the process 224 taxa were evaluated according to the International Union for Conservation of Nature, IUCN, Red List Categories and Criteria78. The IUCN Red List is a critical indicator to identify species most vulnerable to extinction considering a set of criteria, i.e., species’ population trends, size, structure, and geographic ranges. A Red List workshop with the participation of 25 experts from different project partner institutions and IUCN specialists was organized to assess the extinction risk of taxa. The threat analysis included not only species, but subspecies and subpopulations (i.e. races) for some groups (Supplementary Data 3, see summary in Table 1).(iv) Systematic conservation planning assessmentTo undertake the following spatial analyses we focused on the dataset of 224 CWR described above, which is representative of the CWR of the main crops of Mesoamerica (10 genera, Table 1).Species distribution modelingTo compile occurrence records, hundreds of data sources were consulted, including published and personal databases of the project participantse.g.79,80,81,82, the Agrobiodiversity Atlas of Guatemala (https://www.ars.usda.gov/northeast-area/beltsville-md-barc/beltsville-agricultural-research-center/national-germplasm-resources-laboratory/docs/atlas-of-guatemalan-crop-wild-relatives), the Global Biodiversity Information Facility (GBIF, https://www.gbif.org/), and Mexico’s Biodiversity Information System (SNIB, http://snib.mx/).To generate potential species distribution models (SDM), we used more than 13,000 occurrence records (Supplementary Data 4), that were standardized and curated by experts to generate the range maps of taxa as part of the extinction risk assessment, which were published in IUCN Red List (https://www.iucn.org/news/species/202109/threats-crop-wild-relatives-compromising-food-security-and-livelihoods). Spatial resolution of the SDM was 1 km2. SDM were obtained for taxa with more than 20 unique occurrence data in a 1 km2 grid covering the study extent to reduce uncertainty when using smaller sample sizes83. We used 19 bioclimatic variables and other climatic variables, such as annual potential evapotranspiration, aridity index, annual radiation, slope, and altitude84,85,86. Climate data represents annual and seasonal patterns of climate between 1950 and 2000. Also, we used a variable that described the percentage of bare soil and cultivated areas87. Collinearity between variables was assessed with the ‘corselect’ function of the package fuzzySim version 1.088, using a value of 0.8 and the variance inflation factors as criteria to exclude highly correlated variables.We used MaxEnt version 3.3.1, a machine-learning algorithm that uses the maximum entropy principle to identify a target probability distribution, subject to a set of constraints related to the occurrence records and environmental data89,90. Model calibration area for each taxon included those ecoregions where the taxon has been recorded; we used the terrestrial ecoregions dataset69. We did this based on the calibration area or ‘M element’ of the BAM diagram that refers to areas that have been accessible to the taxon via dispersal over relevant periods of time91,92. We randomly sampled 10,000 background localities from the selected areas.To reduce model complexity without compromising model performance, we built several models by varying the feature classes (FC) and regularization multipliers (RM) (see refs. 93,94,95) using R 3.6.096 and ‘ENMeval’ version 0.3.0 package97. FC determines the flexibility of the modeled response to the predictor variables, while the RM penalizes model complexity93. Occurrence records were randomly divided into 70% for model selection, and 30% of data was withheld for model validation. ENMeval carries out an internal partition of localities to test each combination of settings. Therefore, we selected the random k-fold method to divide localities into four bins. We build models with six FC combinations and varied RM values ranging from 0.5 to 4.0 in 0.5 increments. Optimal models were selected using Akaike’s Information Criterion corrected for small sample sizes (⍙AICc = 0). This method penalizes overly complex models and helps to choose those with an optimal number of parameters. However, it has been shown that the number of model parameters may not correctly estimate degrees of freedom98, and that model selection should not be selected solely with one measure99. Thus, we used 30% of the withheld data to test the area under the curve (AUC) of the receiver operating characteristic, and the omission error under a 10 percentile training threshold.We used the ten percentile or minimum training presence threshold to obtain binary maps of the presence and absence of suitable areas for species distribution. We asked experts of each taxonomic group who were also involved in the extinction risk assessment to select one of these two options and to indicate possible overestimated areas, which were then eliminated case by case using the information of Mexican ecoregions100 and watersheds101. Eight models were binarized with the minimum training presence threshold; for the other models we used the 10 percentile threshold. See MaxEnt performance and significance of SDM at Supplementary Data 5. AUC values ranged from 0 to 1; 0.5 indicated a model performance not better than random, while values closer to 1 indicated a better model performance; here we used SDM showing AUC values higher than 0.7. For Phaseolus and Zea, we used SDM that were previously generated by Delgado-Salinas et al.102, and Sánchez González et al.103, respectively. SDM for 116 taxa were validated by experts of each taxonomic group. See references and download links at Supplementary Data 6.For the conservation planning analysis of Mexico, we clipped the models to the Mexican territory, and trimmed the continuous SDM using the binary SDM to keep pixel values of areas with elevated probability of taxa presence. For taxa without SDM, we included the occurrence records of these taxa in the spatial analysis by using the information on observation location, i.e., coordinates (see Supplementary Data 3). This is done by enabling the function ‘species of special interest’ (SSI). See further details in the method section ‘Final conservation analysis’.Proxies of genetic differentiationTo identify proxies of genetic differentiation in an explicit, efficient, and repeatable way, we included environmental and historical drivers of genetic diversity. For this, we first divided Mexico into 27 Holdridge life zones (Supplementary Fig. 2, Supplementary Data 8), which we then subdivided according to phylogeographic studies that have found genetic differentiation among populations of several taxa (see division of each life zone into proxies in Supplementary Fig. 4; Supplementary Fig. 3 provides a general geographical overview of Mexico and main geographic references mentioned in Supplementary Fig. 4). The literature review was done searching for the words “phylogeography” and one of the following: (i) name of the Mexican biogeographic zones, (ii) “Mexico” + an ecosystem name (e.g. “Mexico” “rainforest”) or (iii) “Mexico” + lowlands/highlands. See list of references used in this study in Supplementary Data 9.In addition, we manually reviewed the citations to the most cited papers of the previous search. Reviews and meta-analyses were also included, although we excluded studies performed in CWR to show that our approach can be used without prior information on this group. As more studies on such taxa become available, they can be used to fine-tune the proxies of genetic differentiation. We focused on terrestrial species including plants, animals, and fungi (Supplementary Data 10) except to subdivide a life zone covering the coasts of the California Peninsula, where we could not find studies on terrestrial taxa so we included studies on fish species (see Supplementary Fig. 4).Since most of the life zones cover large territories, and complete phylogeographic congruence among different taxa is uncommon, we targeted to represent general trends that would likely occur across diverse species, instead of trying to represent fine idiosyncratic patterns of genetic differentiation. For instance, although distribution ranges of highland taxa shifted during the Pleistocene climate fluctuations, in general populations persisted (glacial-interglacial periods) within the main mountain ranges, while lowland populations were ephemeral (only glacial periods). So, gene flow among mountain ranges was more limited than within them. As a result, genetic differentiation among mountain ranges of different biogeographic provinces has been widely documented32, so we used this general pattern to subdivide the life zones that occur in highlands. These types of patterns are particularly relevant for a country like Mexico, due to its complex topography, tropical latitude, and geographic features of different ages, which promote population differentiation among the Mexican main geographic features. To translate the phylogeographic information into a spatial context, we used biogeographic regions, basins, topographic or edaphic data to split the life zones into different subzones using the best fitting cartography to represent the phylogeographic patterns (Supplementary Fig. 4).We obtained 102 proxies of genetic differentiation for Mexico (Supplementary Fig. 5). We validated our findings by using available genomic data of an empirical study of a wild relative of maize, the teosinte Zea mays subsp. parviglumis, which was not included in the literature review in order to test the usefulness of our approach regarding the lack of genetic data. The dataset includes ca. 1800 occurrence records and ca. 30,000 SNPs48. Sampling localities were not used for distribution modeling. Admixture groups per population were estimated for K1 to 60. According to the population analysis, Z. mays subsp. parviglumis is structured in 13 genetic clusters along a longitudinal gradient (Fig. 3a–c). We used the K = 13 for plotting based on the Cross-Validation error. The proportion of each genetic cluster was estimated by sampling locality and plotted using pie charts over the map (Supplementary Fig. 6). Then, using the data layer of the SDM subdivided by proxies of genetic differentiation, we extracted which was the proxy most frequent in a 5 km buffer for each sampling locality. The Admixture plot was ordered by all genetic clusters and subdivided by the proxy of genetic differentiation most frequent for each locality. In addition, we calculated a principal component analysis (PCA) and projected into a score plot the first three components. Individual samples were colored by the proxies where they fell in the 5 km buffer (Fig. 3c). To compare how genetic variation was represented by the different scenarios we plotted the proportion of the area of each proxy as given by the potential SDM according to two different scenarios (only considering SDM; combining SDM*PGD) considering 20% of Mexico’s terrestrial area (Fig. 3d). Analyses were run in R version 3.5.196 using the R packages pcadapt version 4.3.3104, ggplot2 version 2_3.3.3105, readr version 1.4.0106, gridExtra version 2.3107, ggnewscale version 0.4.5108, scatterpie version 0.1.5109, pophelper version 2.3.1110, raster version 3.4-5111, rgdal version 1.4-8112, rgl version 0.107.10113, and sp version 1.4-4114,115.Habitat preferenceWe considered habitat preference to refine the presence of CWR in the planning process; thus minimizing commission errors and highlighting areas that more probably contain taxa116. For each taxon, experts assessed its habitat preference (1: high preference; 0.5: low preference; 0.1: no preference) according to the following categories: (i) well-conserved vegetation (i.e. primary vegetation), (ii) human-impacted vegetation (i.e. secondary vegetation), (iii) less intensive rainfed and moisture agriculture, (iv) intensive rainfed and moisture agriculture, (v) irrigated agriculture, (vi) induced and cultivated grasslands and forests, and vii) urban areas (Supplementary Data 11). To spatially delimit these classes, we used the land use cover and vegetation map for Mexico117, and assessed seven main categories of land cover by grouping the map legend (Supplementary Fig. 9). To differentiate between less intensive and intensive cultivated areas, we followed Bellon et al.56, who associated the presence of native maize varieties of Mexico to occur in municipalities with average yields of less than or equal to 3 t ha-1 using agricultural production data from 2010 from the Information System of Agrifood and Fisheries (SIAP), and selected the municipalities with the established average maize yield. We combined the municipality layer with the land cover map to differentiate areas of high and low agricultural intensity. To generate taxon-specific habitat layers, we associated the habitat preference classes established by experts to the land cover map aggregated into seven major land cover categories, using R 3.6.096 and the following packages: raster version 3.4-5111 and rgdal version 1.4-8112. We obtained habitat maps for 116 taxa with SDM.Preliminary analysisWe generated five preliminary scenarios to explore different approaches to include conservation features for maximizing the representation of intraspecific diversity as given by taxa and proxies of genetic differentiation, i.e., representation of proxies within a taxa range (Supplementary Fig. 7): (i) “SDM” scenario, included 116 SDM, which we used as base scenario to examine the representation of taxa and proxies of genetic variability (n = 116); (ii) “SDM + LZ” scenario, included 116 SDM and 27 layers representing Holdridge life zones to consider environmental variation (n = 143); (iii) “SDM + PGD” scenario, included 116 SDM and 102 layers representing each proxy of genetic differentiation individually (n = 218); (iv) “SDM*PGD” scenario, included 5004 input layers representing the intersection of SDM and PGD (n = 5004; combining 116 SDM with 102 proxies resulted in 11,832 layers, but as some of the intersections produced empty outputs given the extension of SDM that do not cover all Mexico, for further analysis we used 5004 input layers with value data. To subdivide the layers, we used ArcGIS version 10.2.2118; to filter the layers, we used R 3.5.196.); (v) “SDM and PGD as ADMU” scenario, included 116 SDM as the main conservation features, while integrating one single layer of proxies of genetic differentiation to consider each of them as planning units by using the ‘Administrative units’ function. Analysis was done in Zonation50,119.We compared the results by assessing 20% of Mexico’s terrestrial area (Fig. 5b) to perform statistical analysis in R 3.5.196 using the following packages: purrr version 0.3.4120, ‘dplyr’ version 1.0.2121, ‘ggplot2’ version 2_3.3.3105, ‘raster’ version 3.4-5111, ‘scales’ version 1.2.0122, ‘sp’ version 1.4-4114,115, ‘tidyr’ version 1.0.2123, and ‘vegan’ version 2.6-2124. The area threshold was established based on Aichi target 11 and on comparisons of performance curves to efficiently represent taxa ranges delimited by SDM and proxies of genetic differentiation (Fig. 6). As using SDM combined with proxies of genetic differentiation showed the highest representation of genetic diversity (“SDM*PGD” scenario), we used this approach for the final analyses.Final conservation analysisWe identified areas of high conservation value for CWR in Mexico by using the software Zonation version 4.050,119, a systematic conservation planning tool that allows optimizing representation of species, taxa, or other conservation features, e.g., proxies of genetic differentiation, in a given study area. The program hierarchically ranks areas by removing cells of low conservation value, as given, for example, by a reduced number of taxa or occurrence of low weighted features, while considering multiple criteria such as the weighting of taxa and habitat preference of taxa. We applied the core-area zonation removal rule (CAZ) to maximize the representation of all conservation features in a minimal possible area51. Zonation generates two main outputs: (a) a hierarchical landscape priority rank map, that allows decision makers establishing different area thresholds to highlight areas of conservation interest; and (b) a representation curve showing species or conservation features range distribution in a given area. The curve also allows identifying how much area is needed to cover a certain taxon range or the distribution of a feature of conservation interest.For the conservation scenarios, we integrated the following inputs in the Zonation software: (1) 5,004 layers, i.e., SDM intersected with proxies of genetic differentiation (as described by “SDM*PGD” scenario, Fig. 4), (2) occurrence records of 98 taxa; only for those taxa without SDM, see Supplementary Data 3), (3) taxa specific habitat layers (according to Supplementary Data 11 and Supplementary Fig. 9), and (4) IUCN threat category (Supplementary Data 3) as an additional parameter to weight taxa differently to consider their vulnerability to extinction, see details below. See Zonation configuration at Supplementary Note 6.Data from different sources can be mixed in the same analysis, which is useful to not lose or omit information of any taxa of interest in the assessment. Here, we included information of a total of 214 taxa (see Supplementary Data 3). Distribution data of 116 taxa were represented by 5004 layers that resulted from combining 116 SDM and 102 PGD. This approach showed the highest proportion of area of taxa ranges (on average 41%) and highest representation of PGD within the area of each taxon (on average 76%; Fig. 4; see description in the main text). For some taxa, e.g. Cucurbita pepo, Physalis cinerascens, and Zea mays information on its distribution was assessed at subspecies level rather than at species level, explaining the difference in numbers of CWR taxa.In addition, we included occurrence data of 98 taxa without SDM to prevent missing important areas of taxa known distribution that are important to conserve (see Supplementary Data 3). We enabled the function ‘species of special interest’ (SSI) of Zonation, and included a SSI feature list file, listing the taxon names, as well as taxon-specific coordinate file for each of the 98 taxa that have been reviewed by the experts of each group. The spatial reference system was World Mercator projection. Occurrence data and SDM are treated similarly in the Zonation analysis, i.e., cells where taxa occur will be retained in the solution as long as possible to maximize its representation in the solution.We assigned weights to the 116 taxa with SDM by using IUCN threat categories (according to Supplementary Data 3), giving highest values to taxa with highest risk of extinction that urgently need management actions to further avoid genetic erosion. By including conservation feature weights, Zonation estimates the conservation value of a cell not only based on the presences of a taxa and their distribution range, but also on the weight. A high weight indicates a high conservation value of cells where these taxa are distributed. As there is no rule for weight setting, we assigned values between 1 and 0 regardless of taxa distribution ranges, which is automatically considered in the Zonation algorithm to guarantee the representation of locations where limited-range distributed taxa occur within the most valuable conservation area. Thus, weights were assigned as follows: Critically endangered, CR: 1; Endangered, EN: 1; Vulnerable, VU: 0.8; Near threatened, NT: 0.5; Data deficient, DD: 0.3; Least concern, LC: 0.2 Not evaluated, NE: 0.1. SSI taxa were all weighted similarly with 1 in order to represent the 98 SSI taxa and their occurrences in the top fraction of the most valuable conservation area, as these areas could be considered as ‘irreplaceable’ in terms of conservation. The conservation of these taxa that are only known in a few locations is crucial to maintain their populations. Information on weights for taxa with and without SDM is included in the file that lists the 5004 conservation features and the SSI file, respectively.To include the information on habitat, we included 116 habitat maps which guide the selection of cells to areas where its presence is more probable (see the Methods section: “Habitat preference”). This option can only be used for taxa represented by a raster layer, and is not available for SSI taxa included via occurrence records. By enabling the “landscape condition” option of Zonation, each habitat map is linked to a specific conservation feature layer. Areas with unfavorable habitats will quickly be masked out during the selection of cells in order to obtain a solution that favors conservation areas within areas of preferred habitat.We generated three final scenarios to identify conservation areas for (a) all taxa, (b) taxa exclusively distributing in natural vegetation, and (c) taxa associated with a wider range of habitats such as natural vegetation, agricultural and urban areas. The Zonation configuration remained similar among the three scenarios. When taxa were not included in a given scenario, we assigned a value weight of 0. This excluded the feature to be considered for the hierarchical prioritization of the landscape, but still allowed to evaluate the taxa during post-processing.To evaluate the spatial results (Supplementary Fig. 11), we analyzed performance curves to represent proxies of genetic differentiation within each taxon range (Supplementary Fig. 12). Also, we considered the most valuable 20% area of Mexico to calculate the coincidence of the three scenarios (Supplementary Fig. 13), and the overlap with federal protected areas125 and indigenous regions126,127 (Supplementary Fig. 14), and land cover data used in the analyses (Supplementary Figs. 9, 15).We discussed the proposed methodological framework, input layer and criteria during a fourth workshop in Mexico. It is worth mentioning that we ran several analyses including additional layers, such as areas where indigenous communities live that promote the presence of CWR in the landscape6. However, as the output indicated no evident difference by including this information, final analyses did not consider these data. We neither included protected areas nor tried to expand on the current 12% protected area system, because most management plans do not specifically address CWR management (but see the management program of the Protected Area of ‘Sierra de Manantlán’128), and thus generally do not adequately plan for wild and native genetic resources129. We also discussed different approaches to consider connectivity for taxa, habitats and proxies of genetic differentiation in the Zonation processing. Still, we finally decided to run the analysis without particularly accounting for connectivity as we had no taxa-specific information on dispersal abilities or possible effects of fragmentation, and we did not want to lose efficiency of the solution to represent taxa by or include lower-quality habitats by forcing the solution to an aggregation of pixels.Reporting summaryFurther information on research design is available in the Nature Research Reporting Summary linked to this article. More
