We targeted the native non-endemic plant species recorded for Socotra and characterised their distributions on a 1° × 1° global grid. Because chorotypes are defined as recurrent patterns of geographic co-distribution, strictly endemic species were excluded, as their ranges are restricted to the archipelago and therefore do not inform broader regional or global co-distribution patterns. Occurrence records for the 477 target species were downloaded from (GBIF 2025), retaining only those with valid geographic coordinates and excluding fossils, living specimens from cultivation (e.g. botanical gardens), records flagged as non-native, and observations with spatial uncertainty > 10 km or with missing uncertainty metadata. Non-native species were excluded from all analyses because their distributions reflect recent anthropogenic dispersal rather than historical biogeographical processes.

Records were then overlaid on the 1° grid following the approach of (Kreft and Jetz 2010), retaining at most one presence per species per cell (coded 1; absence coded 0). To obtain the final dataset used for species distribution modelling and subsequent chorotype analyses, we required a minimum of 15 occurrences per species as a threshold for model calibration (van Proosdij et al. 2016). All the analyses in this paper were performed in R (R Core Team 2025).

Climatic predictors were sourced from CHELSA (Karger et al. 2023) and resampled to the 1° analysis grid. Starting from an initial stack of 36 variables, we reduced multicollinearity using variance inflation factor (VIF) screening with a threshold of 5 using the R package ‘usdm’ (Naimi et al. 2014) and inspected pairwise correlations with a correlation plot. The final predictor set comprised Bio03 (isothermality), Bio07 (temperature annual range), Bio08 (mean temperature of wettest quarter), Bio15 (precipitation seasonality), Bio19 (precipitation of coldest quarter), humidity range (Hursrange), net primary productivity (NPP, as a coarse integrative proxy of ecosystem productivity), potential evapotranspiration range (PET range), and mean wind speed (Windmean), plus a longitude layer to help separate western and eastern hemispheres (acknowledging the circular nature of longitude).

We used species distribution models (SDMs) purely as a step to regularise GBIF occurrences and derive coarse-grained (1°) potential species ranges for subsequent chorotype analyses, not for ecological inference on predictors (Beck et al. 2014). Accordingly, standard model configuration was chosen to ensure consistency and scalability across a large number of species, rather than to maximise predictive accuracy for individual taxa, which would require species-specific tuning and extensive optimisation (Valavi et al. 2022). SDMs were fitted in ‘biomod2’ (Guéguen et al. 2025) using Gradient Boosting Machines (GBM), Random Forest (RF), and MaxEnt. For each species, we set a fixed seed (123), generated pseudo-absences with a disk strategy (3 replicates; 1,000 PAs per replicate; minimum/maximum distances 1,000–10,000 km), and split the data 80/20 for training/testing with three random cross-validation replicates. Single models were evaluated with ROC (AUC) and TSS, and ensembles retained only single models exceeding ROC ≥ 0.7 and TSS ≥ 0.6. Because we aimed to delineate broad ranges, we converted continuous suitability to binary presence/absence using the TSS-maximising threshold, prioritising sensitivity to limit commission errors and spurious spillover at 1° (Hellegers et al. 2025). Although threshold selection and binarisation may introduce uncertainty and information loss (Liu et al. 2013; Guillera-Arroita et al. 2015), especially for presence-only data, this step is a pragmatic compromise. Ensemble predictions were then projected globally and binarised per 1° cell to reconstruct the species × grid matrix of modelled presences. After removing cosmopolitan and very narrowly distributed species, this matrix formed the input for chorotype calculations.

We aggregated species’ binary ranges into a presence-absence matrix by grid cell. To avoid artefacts in between-cell compositional dissimilarity at coarse-grained scales, we filtered the species set by excluding cosmopolitan taxa known to inflate similarity across distant regions and removing very rare species occurring in ≤ 3 grid cells (to limit singletons).

Between-cell ecological dissimilarity was quantified with the Sørensen family of β-diversity using the turnover component (β_sim) from ‘betapart’ (Baselga and Orme 2012), following the guidelines from Kreft and Jetz (2010); we denote the resulting dissimilarity matrix as D_ecol. Geographical separation was computed as the great-circle (haversine) distance between cell centroids and min–max normalised to [0,1], yielding D_geo. We combined the two sources of dissimilarity with the convex mixing of ClustGeo (Chavent et al. 2018):

D α = ( 1 − α ) D ecol + α D geo

We scanned α in {0.0, 0.1, …, 0.9} and the number of clusters k in {5, …, 15}. For each (α, k), we partitioned cells by Partitioning Around Medoids (PAM) on D_α and computed within-cluster inertia curves as a diagnostic tool. Inspection of inertia curves across (α, k) and average silhouette width identified a parsimonious solution balancing ecological turnover and spatial cohesion; we retained α = 0.5 and k = 10 and assigned final cluster memberships via PAM on D_0.5.

To test whether a species is present in some clusters over others, we adopted a use-availability framework originally developed for habitat-selection studies in animals (Neu et al. 1974; Manly et al. 2004), transferring it to plant distributions across clusters. Although conceptually related to indicator-species approaches such as the framework of Dufrêne and Legendre (1997), our objective was not to identify species characteristic of single clusters, but to quantify their associations with one or multiple clusters and retain full compositional signatures across space. In our setting, “availability” is the spatial support of each cluster and “use” is the realised set of grid cells occupied by a species.

First, for each cluster k we computed its availability as the fraction of equal-area cells it covers, p_k = N_k / N (with N_k cells in cluster k and N the total). For a given species, we then counted how many template cells it occupies in total (n) and how many of those fall in cluster k (x_k), which gives the species’ observed proportion in that cluster, p^kobs=xk/n. We measured the deviation from availability-based expectations using an availability-corrected enrichment, defined as a selection ratio (use/availability), conceptually equivalent to species fidelity indices used in vegetation ecology (Cáceres and Legendre 2009). For a given species and cluster k, enrichment was calculated as:

Ek=p^kobs pk, log2Ek=log2(p^kobs pk)

Where values E_k > 1 mean the species appears in cluster k more often than expected from availability alone. For ease of interpretation, the ratio was log₂-transformed, so that values of 0 indicate no deviation from expectation, positive values indicate higher-than-expected occupancy, and negative values indicate lower-than-expected occupancy. Statistically, we asked whether the count x_k could plausibly arise if occurrences were placed at random according to availability, by testing H₀:x_k ~ Binomial(n, p_k) against the one-sided alternative x_k > np_k. We adjusted the resulting p-values across clusters within each species using Benjamini-Hochberg FDR. A species was deemed enriched in cluster k only if three conditions were all met: FDR-adjusted q_k < 0.05, effect size log₂E_k ≥ 0.8 (i.e., at least 1.74× the null expectation), and a minimum support of x_k ≥ 5 cells. This yields a binary species-by-cluster table flagging significant enrichments and for each species, a compositional signature p_k^obs describing how its occurrences are distributed across clusters.

Our goal was not to redraw geographic regions, but to identify chorotypes, i.e., groups of species that share similar geographic distribution. To achieve this, we adopted a two-step procedure: first, grid cells were clustered into spatially coherent units based on species turnover and geographic distance; second, species were grouped into chorotypes according to their proportional occurrences across these spatial clusters. In this framework, a chorotype may occupy several geographic clusters identified through the cluster analysis, and in fact, it is defined by species composition, not by a single spatial cluster.

First, using the enrichment results, we treated each species as a set of geographic clusters in which it was enriched. To define chorotypes quantitatively, we compared species by their full compositional signatures across clusters. We built a species × cluster matrix of observed proportions (p_k^obs) and computed Jensen-Shannon distances between species. This metric quantifying the dissimilarity between probability distributions is symmetric and bounded, and is therefore appropriate for comparing compositional signatures of species across clusters, where values represent proportions rather than absolute counts (Lin 1991). Species were then clustered with Partitioning Around Medoids (PAM) on this distance matrix. We explored k = 2–40 and chose k by maximising mean silhouette width; for reporting and comparability with the coarse analysis, we retained a parsimonious k = 10 solution and used medoids as examples of each chorotype.

To resolve the within-group structure that is blurred at the global extent and coarse-grained, we performed a restricted chorotype analysis centred on the region of interest (Socotra and surroundings). This step refines a subset of global macro-chorotypes into finer, regionally coherent groupings, which hereafter we call “meso-chorotypes”, and reduces confounding from long-distance biogeographic breaks. These terms are purely operational labels to distinguish patterns at different spatial scales, and both correspond to global chorotypes sensuFattorini (2015). In practice, they correspond to species with a broad and a narrow distribution range.

We first selected a set of global macro-chorotypes which exclude North and South America and Oceania. On the restricted grid, we recomputed between-cell ecological dissimilarity with the same distances as before. We scanned α ∈ {0.0,0.1,…,0.9} and k ∈ {5,…,30}, partitioned with PAM on D_α, and used within-cluster inertia to select a parsimonious solution. This yielded a locally finer partition (e.g., α = 0.5, k = 25) used as the cluster raster for subsequent steps.

We repeated the availability-corrected enrichment test exactly as above to obtain a binary species × cluster matrix and per-species proportional signatures. We computed Jensen-Shannon distances among species’ proportional signatures (ε-smoothing = 1×10⁻⁹) and clustered species with PAM on this distance matrix. We explored k = 2–40 and chose k by maximum mean silhouette; the retained solution was k = 12, reflecting greater heterogeneity at the restricted extent.

Chorotype names were assigned after the analyses to aid interpretation and mapping. For each chorotype, we inspected its mapped extent and chose a geographic anchor based on the major area covered and the medoid cell, added habitat/biome qualifiers reflecting the prevailing environments within the group’s footprint at 1°, and noted biogeographic affinities using standard terminology from regional floristic schemes.

Species that were not classified chorotypically by the main pipeline (e.g., excluded by data or occurrence thresholds) were not used in any statistics. For completeness, we provide a qualitative crosswalk in which each such species’ known distribution was visually compared with the mapped chorotypes and annotated with the most compatible chorotype(s), following the common practice when chorotypes are already defined in the literature.

To link chorological patterns with vegetation units, synoptic tables of vegetation types according to De Sanctis et al. (2013) were combined with species-level chorotype assignments. For each vegetation type, species frequencies were aggregated by chorotype and normalised to obtain relative chorotype contributions. The chorological signature of each vegetation type was summarised by identifying the dominant chorotype and the three most represented chorotypes. To quantify the degree of chorological specialisation versus mixing, we calculated Shannon diversity of chorotypes as a measure of chorological evenness.

We retrieved GBIF occurrence records for Socotra’s species (4,493,566 records, Suppl. material 1). Of the 477 target taxa, 417 had usable GBIF occurrences and were retained for analysis. Then we applied an ≥ 15-occurrence filter, which yielded 389 species to be modelled and excluded 28 species for insufficient data.

From the initial set of predictors, we retained ten variables for SDMs: Bio03 (isothermality), Bio07 (annual range of air temperature), Bio08 (mean daily mean air temperatures of the wettest quarter), Bio15 (precipitation seasonality), Bio19 (mean monthly precipitation amount of the coldest quarter), Hursrange (annual range of monthly near-surface relative humidity), NPP (net primary productivity), PET range (annual range of monthly potential evapotranspiration), Windmean (mean monthly near-surface wind speed), and longitude. 26 variables were removed for collinearity issues. We successfully fitted SDMs for 389 species. Average test performances across species are summarised in Table 1.

Looking at the model performance indicators, for the AUC on average, the models can be classified as excellent for all three algorithms. The TSS also offers good results for model performance, with values exceeding 0.8. The TSS had an average value of 0.826 for the ensemble models. This was the only performance metric we used to merge single models. From the binary maps of species resulting from the SDMs and the 28 species not used for modelling but useful for the next part of the analysis, we achieved 114,219 occurrences in the final matrix, covering 7,969 cells and a total of 417 species.

Using ClustGeo mixing, scanning α ∈ [0, 0.9] and k = 5–15, the inertia surface identified a parsimonious solution at α = 0.5, k = 10 (Suppl. material 1). The average silhouette of the retained PAM partition was 0.325.

The ten clusters form continent-scale, spatially coherent groups at 1°. In the Western Hemisphere, C5 and C10 cover North/Central America and South America, respectively. In the Eastern Hemisphere, Africa is split into two clusters, the Arabian Peninsula appears with extensions into the Horn of Africa and the Indian region, Australia forms a distinct cluster, and C2 spans the Afro-Mediterranean belt and much of the Levant region (Figure 1).

Availability-corrected tests (binomial one-sided with BH FDR) identified significant species-to-cluster enrichments (Suppl. material 2). Per cluster, enriched species counts were: C1 [45], C2 [85], C3 [65], C4 [4], C5 [8], C6 [230], C7 [195], C8 [179], C9 [19], C10 [8]. Since the total enriched counts are 838, it is clear that more than one species shares the same cluster.

In fact, Table 2 shows the combinations of clusters associated with each chorotype identified by the PAM clustering on compositional signatures, together with the number of species assigned. We ultimately retained a ten-group solution based on the highest mean silhouette (0.381), which forms the distinction of ten different chorotypes based on different cluster combinations. The cluster combination C6-C7-C8 includes the largest number of species, chiefly covering Africa and the Arabian Peninsula, whereas the least represented combination is C3-C6-C9, with only 16 species, spanning the Arabian Peninsula, India, Indochina, and Australia.

We retained global macro-chorotypes 1, 2, 4, 7, and 9 from the previous cluster analysis (which correspond to cluster 2, 3, 6, 7, and 8) to build a restricted dataset more focused on the surroundings of Socotra Island. After filtering, the analysis comprised 4,049 cells and 242 species.

Scanning α ∈ [0, 0.9] and k = 5–30, the selected partition was α = 0.5, k = 25, with a mean silhouette of 0.3438806 (Suppl. material 1). Here, the partition is finer, with many clusters that nevertheless delineate geographic areas clearly (Figure 2). For example, clusters 13 and 16 occupy the central-eastern belt of the Arabian Peninsula and its southern sector with extensions into the Horn of Africa, respectively. Across Africa, there is also a clear regionalisation, with several groups separating west from east and likewise north from south. Madagascar is, moreover, treated as a distinct cluster.

Availability-corrected tests (one-sided binomial with Benjamini-Hochberg FDR) showed a wide disparity in species enrichment among clusters: several have low or even zero counts (e.g. clusters 1, 2, 3, 5, and 10), whereas others are strongly enriched, with 174 species in cluster 16, 121 in cluster 19, and 119 in cluster 13. Nonetheless, the overall number of enriched species is again high (1,062), indicating that many species share the same clusters and thereby providing a robust basis for chorotype delineation (Suppl. material 3).

Proceeding to the next step, we examined combinations of clusters together with species’ compositional signatures and obtained a 12-group solution, selected by a mean silhouette of 0.374. Given the larger number of groups arising from the initial species-level clustering, the combinations available to define chorotypes are correspondingly richer and span many geographic areas (see Table 3). Even so, the number of species per chorotype is broadly stable across groups. Moreover, the clusters with low counts of species from the enrichment analysis are missing from the results and from the final cluster analysis.

Having established these results, we mapped both macro- and meso-chorotypes to provide a spatial depiction of their distributions, together with a detailed description (Suppl. material 4). By adding previously excluded species (e.g. cosmopolitans or taxa with too few usable occurrences) and assigning them to the most compatible chorotypes, we compiled a chorological spectrum for Socotra’s non-endemic flora (Suppl. material 1). This spectrum is dominated by the Xeric Tropical-Subtropical chorotype (13.8%), followed by the Paleoarabian-Somali-Masai chorotype (10.1%) and the Temperate-Subtropical Belt (9.2%). The least represented chorotype is the Western Indian Ocean Humid Arc (0.8%) (Figure 3).

Across vegetation types, chorological signatures showed marked differences in both dominant biogeographic affinities and degree of mixing (Figure 4; Suppl. material 5). Most woodland and shrubland types were strongly dominated by endemic elements, with dominance values often exceeding 55–70% (e.g. BaW, ChHsS, CoW), suggesting high chorological specialisation. In contrast, several halophytic and grassland types showed lower dominance and higher Shannon diversity, reflecting a more mixed chorological composition (e.g. PaG, AaTaG). Shannon entropy revealed a gradient from highly specialised vegetation types, characterised by strong prevalence of a single chorotype, to more heterogeneous assemblages in which multiple chorotypes contribute substantially to community composition.

The chorotype analysis for Socotra suggests a split into two broad categories: macro-chorotypes with wide distributions, and meso-chorotypes with more restricted extents. Macro-chorotypes represent distributional patterns that span entire floristic realms and may reflect wide ecological tolerances or effective dispersal. Meso-chorotypes, by contrast, have narrower ranges and are typically tightly connected to the ecology and palaeoecological history of Africa and Eurasia. Socotra’s meso-chorotypes often display relictual or disjunct characters. The relatively coarse 1° grid is nevertheless sufficient for the spatial scale presented here.

It is important not to confuse our analysis, based on present-day geographical patterns, with a phylogeographic analysis, which would consider the clade rather than the species as the unit of study and would therefore allow the placement of endemics within broader chorotypes. Unfortunately, this is not currently feasible because genetic knowledge of Socotran, and indeed African, endemic species remains too limited and patchy. While some emblematic lineages have been studied, multi-taxon phylogeographic datasets are still scarce (Miller and Morris 2004; Linder 2014). It should also be noted that the phylogeographic analyses carried out so far, with some exceptions, have mostly investigated colonisation processes rather than characterising chorotypes (Sanbonmatsu and Spalink 2022; Španiel and Rešetnik 2022; Dantas-Queiroz et al. 2023). As our study shows, a chorotype framework allows one to test hypotheses not only about migration and dispersal, but also about how ranges contract and expand in response to ecological change through time.

The units we retrieved often align with classic qualitative frameworks. In particular, the clusters encompassing the Horn of Africa and southern Arabia echo Somalia-Masai and Eritreo-Arabian concepts, whereas arid belts across northern Arabia and the Sahara match Saharo-Arabian/Saharo-Sindian delimitations. Socotra’s placement within an Indian Ocean-facing tropical unit, yet with strong links to the southern Arabian escarpments, is consistent with its historical treatment as part of the Eritreo-Arabian complex and, in some schemes, as a distinct Socotra Province. Where our units diverge from classic outlines, three design choices plausibly contribute: our global step uses turnover-only dissimilarity (emphasising replacements over richness gradients), we analyse non-endemic taxa at coarse grain, and we reflect contemporary ranges that integrate recent dynamics. In these places, the ‘lines’ drawn historically may better represent broad transition zones than clear-cut borders. Our results should therefore be seen as complementary to those frameworks, highlighting patterns of regional connectivity and a shared species pool rather than centres of endemism. Miller and Morris (2004) do not formally define chorotypes in a quantitative sense but remain the most comprehensive qualitative synthesis for interpreting the chorology of the Socotran flora. They identify strong affinities with southern Arabia and the Horn of Africa, connections with the Somali-Masai region, and several distant disjunctions. These patterns are broadly consistent with those retrieved by our analysis. The dominance of xeric tropical-subtropical and Paleoarabian-Somali-Masai chorotypes in our results mirrors the emphasis placed by Miller and Morris (2004) on Eritreo-Arabian and Somalia-Masai elements as the primary non-endemic components of the flora.

Our study differs in approach rather than in conclusions. Indeed, Miller and Morris (2004) rely on expert floristic knowledge and historical-ecological reasoning, and we recover similar biogeographic signals through a quantitative analysis of species co-distribution at global scale, formalising chorological relationships long recognised in the literature.

The flora shows well-known disjunctions (Tethyan, African dry-corridor, and boreotropical) and mixed affinities with the Old-World tropics, northeast tropical Africa/Arabia, and the Indian subcontinent (Mies 1996; Banfield et al. 2011). The spatial coherence of macro-chorotypes can be interpreted as consistent with combined effects of environmental filtering (aridity and seasonal monsoon regimes, persistent coastal fog, elevation-driven moisture) and historical legacies that limit exchange across barriers (e.g., sea straits, desert interiors) while fostering diversification in refugial uplands. Within the restricted domain, it highlights nested assemblages along escarpments and montane belts, consistent with upslope range stacking and asymmetric source-sink dynamics from mesic refugia to surrounding xeric lowlands.

The chorological spectrum of Socotra’s non-endemic flora is dominated by xeric, tropical-subtropical Eastern Hemisphere chorotypes, suggesting that the island sits within a thermo-xeric macro-network linking Northeast Africa, southern Arabia, and adjoining arid belts. This pattern is consistent with classical placements of southern Arabia within Somalo-Arabian/Eritreo-Arabian phytochoria and with a long history of aridity-adapted lineages moving along desert and semi-desert corridors (Zohary 1973; White 1983; Takhtajan 1986).

Several lineages suggest deeper antecedents to these xeric affinities. Parts of the Somali-Masai complex may retain a signature of Miocene expansion and fragmentation of seasonally dry tropical formations across northeast Africa and Arabia (Trauth et al. 2005, 2009), before their subsequent contraction during late Miocene to Pliocene aridification (deMenocal 2004). Under this view, high counts in the Paleoarabian-Somali-Masai chorotype (Figure 3) could represent composite outcomes of the Neogene assembly of dry biomes under progressive aridity and tectonic reorganisation (Linder 2014), including the uplift of East African and Arabian topography and the strengthening of rain-shadow and monsoon seasonality (Sepulchre et al. 2006). These features were later modulated by Quaternary sea-level oscillations that periodically opened and closed coastal corridors around the Red Sea and Gulf of Aden (Siddall 2004; Lambeck et al. 2014), with less distance between the Horn of Africa and Socotra Archipelago during lowered sea levels. While our data cannot date chorotypes, the geographic concordance, with strong occupancy along rift-parallel uplands and the Red Sea margins, fits expectations from this two-stage scenario. Although certain Afro-Arabian lineages are very old (e.g. Dirachmaceae), Neogene aridification profoundly reshaped the flora, obscuring much of its original signature (deMenocal 2004; Linder 2014). We therefore interpret the corridor signal not as exclusively late-Quaternary, but as a plausible scenario involving recurrent re-activation of older xeric linkages whose footprints were established during the Miocene and later modulated by glacial-interglacial sea-level lows.

A smaller but coherent contribution from Afrotropical and Indo-Pacific-aligned chorotypes accords with monsoon and fog influences on the south-facing escarpments and highlands. Conversely, occasional ties to intercontinental seasonally dry belts reflect long-range xeric linkages rather than recent exchange across humid barriers and appear as higher-order units that pair coastal-desert clusters with steppe/plateau units. This mirrors what was already found by Banfield et al. (2011).

Interpreted through use/availability-corrected enrichments, Socotra appears as a gateway “sink/waystation” for thermo-xerophilous dispersers. Non-endemic species associated with coastal-desert chorotypes accumulate on the archipelago, whereas strictly mesic elements show weaker or patchier signals. As such, this conclusion reflects study design as much as biological reality; deeper insights are possible only with phylogeographic analysis. This asymmetry is consistent with the selective permeability of the corridor: efficient for dry-adapted lineages moving along coastal and plateau routes, restrictive for humid-adapted taxa constrained to fog-fed enclaves.

These interpretations are, for now, partly speculative because it is not possible to date chorotypes. However, this may become feasible in future by studying recurrent speciation patterns using molecular phylogenies. Although currently constrained by data scarcity and by the lack of statistical methods for comparing clades across different genera, such an approach would open a new phase in phylogeography, one in which the traditional chorotype framework is combined with molecular phylogenetics.

Taken together, these patterns point to a small number of dominant floristic assembly pathways shaping Socotran flora: a primary Afro-Arabian xeric corridor linking southern Arabia and the Horn of Africa, and a secondary, moisture-mediated pathway associated with monsoon-influenced escarpments and refugial uplands. As such, Socotra behaves as a continental-fragment island whose non-endemic flora largely reflects structured regional connectivity. Compared to the Arabian Peninsula and the Horn of Africa, Socotra acts less as a source area and more as a selective sink, where regional floristic signals are filtered, compressed, and preserved under insular conditions.

Availability-corrected enrichments identify species that occur in particular chorotypes more often than expected given each chorotype’s spatial availability. This yields per-chorotype sets of indicator species that remain comparable even when chorotypes differ greatly in spatial extent. To compare species consistently across the full set of chorotypes, we then summarised for each species a compositional signature (the proportions of its occurrences in each chorotype) and measured between-species similarity using Jensen-Shannon distance, which is appropriate for compositional data and relatively insensitive to sparse zeros. Clustering species with PAM on this metric returns quantitative chorotypes with visually interpretable medoids and standard diagnostics of within-group cohesion (silhouettes). Taken together, availability-corrected enrichment and clustering of compositional signatures delineate reproducible co-distribution units, separating thermo-xeric coastal belts from more mesic montane complexes, and revealing where species share multi-chorotype affinities (e.g. coastal and plateau combinations around the Red Sea and Gulf of Aden). We emphasise that these are spatial co-occurrence patterns, not evidence of interaction; their value lies in defining operational units and testable hypotheses about environmental constraints and historical connectivity.

The chorological spectra observed across vegetation types provide a quantitative link between species biogeography and community composition. By linking species-level chorotypes to synoptic vegetation tables, our results suggest that chorological structure is not randomly distributed across vegetation types but instead reflects coherent biogeographic signatures of plant communities and an ecological gradient across vegetation types. Woodland and shrubland vegetation units are characterised by a strong dominance of endemic chorotypes, highlighting the importance of local evolutionary history in structuring these assemblages. Conversely, grasslands and halophytic communities show more heterogeneous chorological spectra, with a stronger contribution of Afro-Arabian and Temperate-Subtropical chorotypes, suggesting higher connectivity with regional species pools and higher turnover along environmental gradients and, in many cases, recent arrival from the continent. Overall, the chorotype composition reflects a gradient from locally specialised, refuge-like communities to more open, climatically driven assemblages with broader biogeographic affinities, a pattern also observed in other studies where endemic-rich communities correspond to more centralised and ecologically specialised vegetation types (Pignatti et al. 1991). For example, Croton sarcocarpus shrubland (CsarS), Boswellia ameero woodland (BaW), and Buxanthus pedicellatus shrubland (BpS) (see Suppl. material 5) show a higher proportion of the Paleoarabian-Somali-Masai chorotype than other woody communities, indicating a connection with ancient, relict vegetation types. Conversely, Eragrostis papposa-Arthraxon micans grassland (EpAmG) and Arthrocnemum macrostachyum halophytic communities (ArmaH) are dominated by the Temperate-Subtropical Belt and other chorotypes that suggest continued immigration from the mainland, despite the presence of a substantial endemic component.

This quantitative integration of chorotypes with vegetation types supports earlier qualitative interpretations of Socotra’s vegetation mosaics (De Sanctis et al. 2013). Similar approaches have been applied in European vegetation studies where chorological spectra are used to characterise plant communities and interpret their ecological and biogeographic affinities (Nimis and Bolognini 1993; Berg et al. 2017).

Our workflow was developed specifically to derive chorotypes for Socotra’s non-endemic species from the data currently available. It is still experimental and has several caveats that deserve explicit mention. First, GBIF sampling is uneven; we therefore have not used SDMs for ecological inference, and we binarised predictions conservatively to limit commission at the 1° grain. Because we relied on random (not spatial) cross-validation and pseudo-absences, absolute performance metrics may be optimistic. The downstream analyses are based on binary ranges and between-cell dissimilarity at a coarse spatial resolution and should therefore be interpreted primarily as broad-scale chorological patterns rather than precise distributional estimates. Second, the clustering solution depends on the choice of α (the mixing parameter) and k (the number of groups). Our selected values are supported by inertia and silhouette diagnostics, but they are not unique; importantly, the qualitative structure appeared stable under small changes in α, k, and in the β-diversity component used, although a formal sensitivity analysis was beyond the scope of this study. Third, for species-cluster enrichment, we controlled multiplicity within species using Benjamini-Hochberg FDR, and we report effect sizes (log₂E) to summarise the strength of associations across species rather than inflating claims with additional multiple testing. Finally, the scope is deliberately narrow: analyses are single-taxon and presence-based; we did not model phylogenetic relatedness, functional traits, or temporal dynamics, which remain important avenues for future work.

Our analysis differs from what is highlighted in studies of European chorotypes, where chorotypes are closely tied to macroclimatic patterns, as strongly argued by Meusel et al. (1965) and Berg et al. (2017). By contrast, the chorotypes we identify cannot be fully explained by ecology alone; they require a historical interpretation. Socotra’s non-endemic species appear to retain signatures consistent with geological and climatic processes operating since the Neogene and Quaternary periods.

Chorotypes are routinely used in Europe to characterise plant communities ecologically and biogeographically (Pignatti et al. 1991; Dzwonko and Loster 2000; Košir et al. 2013; Mahmutaj et al. 2015). The gradient identified above allows the different vegetation communities of Socotra to be distinguished not only floristically, but also based on their chorological spectra, thus supporting the classification of De Sanctis et al. (2013). An interesting feature of our study is that, in the case of Socotra, it becomes possible not only to interpret the ecology of species and communities, but also their evolutionary history. This will not only refine their ecological characterisation but also help interpret community evolution. Evolutionary theory is strongly centred on population genetics; introducing chorotypes may shift the emphasis to the broader community context in which populations evolve through species interactions.