Ragala ucuquirana-branca (Fig. 1) is a canopy tree native to the Amazon Basin, typically occurring in lowland terra-firme forests. The species is locally known by several vernacular names, such as “ucuquirana-brava-da-folha-vermelha” (Rodrigues 1974), “balata-brava”, “coquirana-branca”, and “coquirana-rocha” (Pennington 1990; van Roosmalen and Garcia 2000). Its distribution is primarily documented in central and northern Amazonia, extending into southern Venezuela (Pennington 1990). Because many Amazonian Sapotaceae exhibit varying degrees of habitat specialization, Ragala ucuquirana-branca provides a useful case study for investigating how environmental filtering shapes the distribution of ecologically specialized Amazonian tree lineages.

We compiled all available occurrence (presence) records from the Global Biodiversity Information Facility (GBIF) (https://www.gbif.org; GBIF.org 2024) and the speciesLink network (https://specieslink.net; Canhos et al. 2022). In addition, we incorporated records from physical reference collections associated with long-term forest monitoring plots in Central Amazonia, including the Projeto Dinâmica Biológica de Fragmentos Florestais (PDBFF) and Programa de Pesquisa em Biodiversidade (PPBio). In all cases, only records of preserved specimens were retained. During an initial data screening, we excluded specimens from other species through taxonomic verification of digital and/or physical specimens, as well as duplicate records (i.e. specimens with the same collector and collection number, even when deposited in different herbaria).

To minimize georeferencing errors, original coordinates were manually verified to ensure consistency with the corresponding locality descriptions. When coordinates were missing, they were estimated based on the locality descriptions, allowing for reliable georeferencing using Google Earth (https://earth.google.com/web). We also flagged and removed potentially erroneous records by checking proximity to well-known geographic features (e.g. centroids for administrative regions and institutions), identifying spatial outliers, and finding aberrant values or formatting problems using the R package CoordinateCleaner v.3.0.1 (Zizka et al. 2019).

In addition to the occurrence dataset for the ensemble SDM workflow, georeferenced records were analysed in R v.4.4.2 (R Core Team 2024) using the package ConR v.2.1 (Dauby et al. 2017) to estimate Extent of Occurrence (EOO), Area of Occupancy (AOO; 2 × 2 km grid), number of locations (10 km grid), dispersal radius used to delimit subpopulations as a proxy for species dispersal ability (Rivers et al. 2010), number of subpopulations, and degree of severe fragmentation among subpopulations. These spatial parameters were used to perform a preliminary conservation assessment of R. ucuquirana-branca under the IUCN Red List Categories and Criteria (IUCN 2012; IUCN Standards and Petitions Committee 2024).

The geographic extent used for model calibration corresponded to the Amazon Basin (sensu Eva et al. 2005), including adjacent Cerrado and montane regions draining into the Amazon River. This delimitation of the Amazon Basin (Amazonia s.l.) covers approximately 7,595,000 km² and includes areas of Brazil, Bolivia, Peru, Ecuador, Colombia, Venezuela, Guyana, Suriname, and French Guiana. All subsequent modelling steps (background selection, pseudo-absence generation, variable extraction, and ensemble forecasting) were performed within this geographic extent to ensure ecological coherence and minimize spatial bias in model performance and properly encompass the full known distributional range of Ragala ucuquirana-branca.

Environmental variables were selected to capture the environmental gradients potentially driving the distribution of Ragala ucuquirana-branca in the Amazon Basin. The dataset included 37 predictors related to climate, topography, soil, land cover, and habitat heterogeneity (Table 1) that are expected to influence plant physiology, morphology, distribution, and ecology (Toledo et al. 2012; Stein et al. 2014; Lambers and Oliveira 2019).

Standardized environmental layers were downloaded as GeoTIFF raster files from WorldClim 2.1 (https://www.worldclim.org; Fick and Hijmans 2017), SoilGrids (https://isric.org/explore/soilgrids; Poggio et al. 2021), and EarthEnv (https://www.earthenv.org; Tuanmu and Jetz 2014, 2015) at 30 seconds (~1 km) spatial resolution and under historical (near current) environmental conditions (Table 1). Soil data were downloaded using the R package geodata v.0.6-2 (Hijmans et al. 2024a).

To reduce multicollinearity and prevent model overfitting, we selected a subset of weakly correlated bioclimatic predictors (Dormann et al. 2013). Automated variable selection was performed using the select07 function implemented in the R package mecofun v.0.7.1 (Zurell 2024), which identifies correlated pairs (Spearman’s rank correlation coefficient |r| > 0.7) based on environmental values (i.e. the full set of 37 variables) extracted at species occurrence points. Within each correlated pair, we retained the variable with the highest explanatory power, assessed by its univariate importance in a quadratic generalized linear model (GLM) ranked by Akaike Information Criterion (AIC). This procedure resulted in a final set of 21 variables, which were used in all subsequent analyses (Fig. 2), including 11 climatic, 7 soil, 1 topographic, and 2 land-cover variables.

We modelled species’ potential distribution using five algorithms: (1) Boosted Regression Trees (Elith et al. 2008); (2) Classification and Regression Trees (Breiman et al. 1984; De’ath and Fabricius 2000); (3) Generalized Additive Model (Hastie and Tibshirani 1986; Yee and Mitchell 1991); (4) Maximum Entropy (Phillips et al. 2006); and (5) Random Forest (Breiman 2001) (Table 2). We used the R package dismo v.1.3-16 (Hijmans et al. 2024b) to perform Boosted Regression Trees and Maximum Entropy; rpart v.4.1.23 (Therneau and Atkinson 2023) to perform Classification and Regression Trees; mgcv v.1.9-4 (Wood 2011) to perform Generalized Additive Model; and randomForest v.4.1.23 (Liaw and Wiener 2002) for the Random Forest models. We used an ensemble forecasting approach (i.e. combining multiple models for better predictions) to minimize errors (Araújo and New 2007).

We sampled pseudo-absence points across the Amazon Basin. We matched and randomly generated pseudo-absence points, 10 times the number of occurrence points, for all algorithms (Barbet-Massin et al. 2012). Moreover, we randomly split the combined dataset (comprising occurrence records and pseudo-absence data) so that 70% was used for model training and the remaining 30% for testing.

We evaluated the model’s performance by calculating the True Skill Statistic (TSS; [(sensitivity + specificity) - 1]) (Allouche et al. 2006), and the Area Under the Curve (AUC) of receiver operator characteristics (ROC; Fielding and Bell 1997). We generated 10 replications for each algorithm and selected the replicates presenting accuracy values of AUC ≥ 0.75 and TSS ≥ 0.55 to construct the ensemble rasters. Then, we ensembled the models based on the average of the selected rasters weighted by AUC values ≥ 0.75 (Hao et al. 2020).

To create binary distribution (presence-absence) of the predicted habitat suitability, we applied two threshold-selection methods: minimum (or lowest) training presence (MPT), which consists of the lowest predicted suitability value for an occurrence point; and 10th percentile training presence (P10), which omits all regions with habitat suitability lower than the suitability values for the lowest 10% of occurrence records (Morrow 2019).

The ensemble SDM workflow was performed in R v.4.4.2 (R Core Team 2024) using custom-made R scripts and add-on libraries. Final maps were prepared using QGIS v.3.28.1 (QGIS Development Team 2022).

The compiled records from preserved herbarium specimens and long-term forest monitoring plots substantially expand the empirical basis previously available for the species, whose distribution had been documented primarily through sparse records in the last taxonomic treatments for the genus (e.g. Pennington 1990, 2006). Large portions of Eastern Amazonia, Southwestern Amazonia, and especially Southeastern Amazonia (SEA) lack confirmed occurrences despite historical botanical inventories (Fig. 3A) (Hopkins 2007; Stropp et al. 2020; Carvalho et al. 2023), suggesting that the absence of records is unlikely to be explained solely by sampling gaps. Instead, this pattern points to underlying ecological constraints that limit the establishment of the species across large areas of apparently continuous forest cover.

The pronounced discrepancy between EOO and AOO highlights a strongly aggregated spatial pattern, with occurrences restricted to discrete localities (Gaston 1996; Clavel et al. 2011) rather than being evenly distributed across Amazonian forest cover. This indicates that the species occupies a limited fraction of its overall geographic range.

Taken together, the occurrence data indicate that R. ucuquirana-branca occupies only a limited subset of environmentally suitable habitats within its broad geographic extent, with confirmed records forming distinct clusters across the basin. Similar patterns of broad geographic extent combined with highly aggregated occurrence records have been reported for other Amazonian Sapotaceae, including Ecclinusa guianensis Eyma (Vasconcelos et al. 2020), Pradosia ptychandra (Eyma) T.D.Penn. (Vasconcelos et al. 2024), and Pouteria decorticans T.D.Penn. (Gama et al. 2025). This pattern provides an essential empirical baseline for evaluating the environmental drivers and modelled habitat suitability explored in the subsequent sections.

AUC and TSS metrics indicate superior predictive capability (Fig. 4), effectively distinguishing suitable from unsuitable habitats with minimal omission errors (Allouche et al. 2006). The consistently high performance observed across algorithms agrees with previous studies showing that machine learning methods better captures complex, non-linear interactions between abiotic variables and species presence than traditional frequentist models (Elith et al. 2008; Elith and Leathwick 2009).

The strength of the ensemble approach lies in its ability to mitigate the biases of individual algorithms, yielding a consensus projection that is generally more robust and biologically realistic than single-model outputs (Araújo and New 2007; Marmion et al. 2009). By combining predictions across algorithms, ensemble models tend to reduce overfitting commonly observed in high-performance methods such as MaxEnt or Random Forest, resulting in more conservative and reliable suitability estimates. This methodological rigor increases confidence that the identified suitability hotspots are not merely artifacts of sampling bias but instead reflect areas where the environmental conditions fall within the core ecological envelope of Ragala ucuquirana-branca (Phillips et al. 2009).

In the Amazon, soil chemistry acts as a powerful filter; species adapted to specific pH levels are often associated with physiological trade-offs that prevent them from competing in soils with different nutrient availabilities or aluminium toxicity levels (Phillips et al. 2003; John et al. 2007; Quesada et al. 2012; Condit et al. 2013; Figueiredo et al. 2018).

Furthermore, the influence of vapor pressure and minimum temperature of coldest month (Bio6) suggests a narrow physiological tolerance for atmospheric water demand and thermal stress. Vapor pressure, closely related to atmospheric water demand and vapor pressure deficit, influences stomatal regulation in canopy trees, limiting carbon assimilation and growth (Esquivel-Muelbert et al. 2017; Grossiord et al. 2020). The reliance on stable, humid conditions explains the species’ absence from the more seasonal fringes of Eastern Amazonia and particularly Southeastern Amazonia. This suggests that the species’ distribution is governed by a dual-filter system: a chemical filter at the rhizosphere (pH) (e.g. Dubuis et al. 2013) and a physical filter at the leaf-atmosphere interface (vapor pressure) (e.g. Salm et al. 2007), both of which must be met simultaneously for the species to persist.

Cation exchange capacity of the soil (Cec) is a useful indicator of soil fertility because it reflects the soil’s ability to retain and exchange essential nutrients and is closely associated with clay mineralogy and organic matter content. Its relevance suggests that R. ucuquirana-branca may preferentially occur in soils with greater nutrient-retention capacity, which can buffer nutrient limitation in highly weathered Amazonian substrates. Similar associations between Cec and tree species distributions have been documented for several Amazonian forest communities (Phillips et al. 2003; Quesada et al. 2012; Toledo et al. 2017). Together, these results indicate that the realized niche of R. ucuquirana-branca is shaped by a tight coupling between belowground chemical constraints and aboveground atmospheric conditions, reinforcing the view that multiple, interacting environmental filters determine tree species distributions in the Amazon.

The fact that the ensemble model predicts low suitability in regions with contrasting soil pH and atmospheric moisture, including areas historically subject to botanical surveys (Nelson et al. 1990; Hopkins 2007), provides support for the hypothesis of edaphic specialization. This suggests that R. ucuquirana-branca is a true habitat specialist constrained by the humid and less seasonal environmental conditions of Central Amazonia and neighbouring portions of Northwestern Amazonia and Southwestern Amazonia.

Areas predicted as suitable but lacking confirmed occurrence records should not be interpreted exclusively as evidence of undersampling. Species distribution models estimate potential environmental suitability rather than realized occupancy, and discrepancies between predicted and observed distributions may emerge through multiple, non-mutually exclusive mechanisms (Soberon and Peterson 2005). Unmeasured environmental constraints, including soil properties not represented in the predictor set or biotic dependencies such as pollinators and dispersers, may restrict establishment despite otherwise favourable conditions (Araújo and Guisan 2006). In addition, dispersal limitation and historical biogeographic processes may prevent populations from reaching areas with favourable environmental conditions (Svenning and Skov 2004). Such constraints may be especially relevant for ecologically specialized species, whose realized distributions may occupy only a subset of their environmentally suitable range (Pulliam 2000). Nevertheless, given the persistent geographic biases and collection gaps documented across Amazonian biodiversity data (Feeley 2015; Carvalho et al. 2023), incomplete sampling remains a plausible explanation for at least part of the observed discrepancy.

The concentration of highly suitable habitats within Central Amazonia and neighbouring portions of Northwestern and Southwestern Amazonia suggests the existence of a core ecological region for the species, where environmental conditions closely match its physiological and edaphic requirements. In contrast, habitats within Eastern Amazonia and particularly Southeastern Amazonia appear to represent environmentally marginal regions approaching the limits of the species’ ecological tolerance. This pattern of fragmented and spatially isolated patches of predicted habitat observed under both thresholding approaches (Fig. 7) suggests that substantial portions of Eastern Amazonia represent environmentally marginal or unsuitable environments, likely due to the combined effects of lower soil fertility and increased climatic seasonality (Toledo et al. 2017).

This projected fragmentation has deep evolutionary and ecological implications. Small, isolated patches of suitable habitat can lead to reduced gene flow and increased vulnerability to stochastic events (Rosas et al. 2011). As climate change alters rainfall patterns and vapor pressure across the basin (Esquivel-Muelbert et al. 2019; Flores et al. 2024; Lima et al. 2026), these core areas may shrink further under projected increases in climatic seasonality and atmospheric water demand. The broad region of low habitat suitability predicted in Southeastern Amazonia may act as a contemporary environmental barrier, suggesting that the species may be unable to migrate southward in response to changing climates, potentially constraining future range shifts toward southern Amazonia (Araújo and Pearson 2005; Bush et al. 2011).

Our results indicate that Ragala ucuquirana-branca is characterized by ecological specialization associated with soil chemical properties and atmospheric stability, rather than broad environmental tolerance. This finding has important implications for taxonomy, as environmentally structured distributions may contribute to morphological and genetic differentiation across the species’ range, a pattern frequently observed in Amazonian lineages (Gentry 1981; Tuomisto et al. 2003; Fine et al. 2005, 2014). The environmentally structured distribution revealed by the ensemble SDM provides testable hypotheses for future taxonomic and phylogeographic studies. If populations occupying distinct soil or climatic domains exhibit consistent morphological or genetic differentiation, this would support the role of ecological specialization as a potential driver of ecological and evolutionary differentiation within the lineage.

The environmental predictors identified by the ensemble SDM are also consistent with ecological and reproductive patterns reported for R. ucuquirana-branca and related Amazonian Sapotaceae. The influence of soil chemical variables, particularly soil pH and cation exchange capacity, reinforces the interpretation that the species exhibits marked edaphic specialization associated with terra-firme forests, including sandy soils (van Roosmalen and Garcia 2000). In addition, the importance of atmospheric water vapor pressure and low climatic seasonality is congruent with phenological observations reported by Alencar (1994), who demonstrated that reproductive and vegetative phases in R. ucuquirana-branca are associated with seasonal variation in precipitation, insolation, and evaporation in Central Amazonia.

Phenological patterns linked to climatic stability may help explain the concentration of highly suitable habitats in humid and weakly seasonal regions of the Amazon Basin. Alencar (1994) further suggested that reproductive dynamics in Amazonian Sapotaceae are influenced not only by climatic factors but also by ecological interactions involving pollinators, seed predators, and dispersers. Although direct information on pollination biology remains unavailable for R. ucuquirana-branca, observations from Central Amazonia indicate that fruits and seeds interact with a diverse assemblage of vertebrates, including sakis, capuchin monkeys, woolly monkeys, and parrots (van Roosmalen and Garcia 2000). These interactions suggest that both seed predation and vertebrate-mediated dispersal may contribute to the aggregated and spatially discontinuous distribution pattern observed across the species’ range.

Recent macroecological studies further indicate that zoochorous tree species in Amazonia are strongly associated with humid environments and precipitation-related climatic stability, while disruptions in frugivore communities may negatively affect dispersal and recruitment dynamics (Correa et al. 2023). In this context, the fragmented distribution of environmentally suitable habitats predicted by the ensemble SDM may reflect not only environmental filtering, but also dispersal limitation, reduced population connectivity, and the ecological dependence of vertebrate-mediated dispersal processes across the Amazon Basin.

From a conservation standpoint, the pronounced discrepancy between the broad EOO and the comparatively restricted AOO indicates that large portions of the Amazon Basin are likely unsuitable habitats for R. ucuquirana-branca. The concentration of occurrence records within fragmented and environmentally restricted habitat patches reinforces the interpretation that the species is ecologically specialized and potentially vulnerable to ongoing environmental change. Under IUCN Criterion B, species exhibiting restricted occupancy, severe fragmentation, and inferred continuing decline in habitat quality may qualify for threatened categories even when geographic extent remains relatively large. In this context, the ensemble species distribution models provide important ecological support for the preliminary classification of R. ucuquirana-branca as Endangered [EN B2ab(iii)].