2.1 Taxa, ecological niche binnings and ecomorph syndromes

Our sampling includes at least one species from each accessible ant subfamily; we sampled all subfamilies except Apomyrminae and Martialinae, which are extremely rare in museum collections. In total, we sampled 15 subfamilies, 113 genera and 167 species, measuring 320 specimens in total (see Data Availability Statement section, Dryad Digital Repository). When possible, we sampled three specimens per species. When insufficient specimens were available, we measured as many conspecific specimens as were present in collections. Many species are polymorphic or have specialized castes for specific ecological functions: in these situations, we either sampled the media caste for polymorphic species or sampled the nonspecialized workers for species with specialized castes. We included specimens from the New Jersey Institute of Technology (NJIT), the American Museum of Natural History (AMNH), the Smithsonian National Museum of Natural History (NMNH) and the Harvard Museum of Comparative Zoology (MCZ).

All specimens were assigned a binning from each of three ecological niche aspect categories: functional role (six binnings), nesting niche (five binnings) and foraging niche (four binnings) (Figure 2; Table 1). We selected these niche aspects as they are broadly applicable: while other key abiotic or biotic factors such as climate or community composition may also influence morphology, our selected aspects are generalizable across widely disparate communities of species, allowing for global ecomorph definitions. Moreover, to achieve wide taxon sampling, our morphometric dataset was confined to historical museum collections, which frequently do not retain specimen-level information relating to community structures or precise localities. Binning determinations were based on surveys of the literature (Appendix S1, AntWiki). A list of natural history binning data sources used in the study are provided in the Supporting Information section. If an aspect of a species' niche was unclear, the species was assigned an ‘unknown’ binning and excluded from all further analyses for that niche aspect. We attempted to sample representatives from every known ecological niche that ants inhabit; however, due to the extremely diverse nature of ants, the sampling is broadly comprehensive but likely not universal. In particular, given our reliance on museum specimens, we may be missing diversity from geographic areas that have been historically neglected as well as ecologies that are difficult to sample (e.g. subterranean taxa).

Additionally, we collapsed the 35 observed combinations of niche binnings across all niche aspects into 11 ecomorph syndromes (Figure 2; Table S1). These simplified syndromes were defined after preliminary agnostic analyses suggested extensive morphological overlap between some niche combinations. While we defined 11 ecomorph syndromes, one syndrome (subterranean omnivore) was represented by only one species and three specimens. This species, Acropyga oreithauma, represents a clearly defined ecomorph syndrome, but we did not have enough specimens for adequate statistical comparisons and so removed it from subsequent analyses.

Our analyses are constrained by the accuracy of natural history information; in some cases, species may be plastic in their behaviour, or their ecological niche may not be fully known. While our models do not explicitly incorporate this uncertainty, this may be explored in the future through sensitivity analyses with alternate niche assignments for ambiguous taxa. We do, however, incorporate some behavioural plasticity in our collapsed ecomorph syndromes; such as ‘ground or leaf litter-nesting epigaeic omnivores’, or the synonymization of specialist and generalist predators into a general predator class (Table S1).

2.2 Measurements

Morphometric sampling included linear measurements of 12 cephalic traits and five post-cephalic traits (Table S2). Fifteen selected traits have been previously correlated with ecology. These traits include metafemur length (Yates et al., 2014), eye size (Gibb et al., 2015; Weiser & Kaspari, 2006), body size (Gibb et al., 2015; Salas-López, 2017) and mandible length (Weiser & Kaspari, 2006). Our initial sampling included five additional metrics (tarsal claw length, tarsal claw width, mandible articulation angle, ventral head length and ventral pronotal length); however, these were excluded from data collection after preliminary analyses indicated overlap with other metrics, variation due to measurement artefacts or lack of significance across all niche aspects.

All measurements were conducted on point-mounted specimens under stereo microscopy. Including raw measurements in dimension reduction techniques such as principal component analysis (PCA) can result in body size driving the overwhelming majority of variation in a dataset, masking other potentially important contributors. To explore the impact of body size, we created two datasets for analyses: a dataset comprising raw measurements and a size-corrected dataset using only ratios (Table S3).

2.3 Data analysis

Our analyses spanned three broad categories: assessments of relationships between individual traits and ecologies; multivariate analyses of traits in the context of ecology; and multivariate predictive modelling of ecology from trait data. To agnostically determine whether or not individual traits were correlated with ecology, we performed analyses of variance (ANOVAs) on trait data and functional role, nesting niche and foraging niche binnings. We accounted for the potential effect of phylogenetic relatedness on ANOVAs by performing phylogenetic generalized least squares (PGLS) on our trait data using a published phylogeny of ant genera (Blanchard & Moreau, 2017). To evaluate which traits drove maximum variation and to visualize ecomorphospaces (representations of taxa and their ecologies based on morphometric data) among groups, we ran principal component (PCAs) and principal coordinate analyses (PCoAs). While both PCAs and PCoAs are dimension reduction techniques, PCoA is unique in that it is derived from distance matrices and so can include inapplicable or missing data. To determine which traits drive maximum separation between groups, and whether traits can delineate distinct ecomorphospaces, we conducted linear discriminant analysis (LDAs), which reduces dimensions in the context of predefined binnings. We further tested the accuracy of LDA group separation by comparing posterior probabilities of classification. Finally, we evaluated the predictability of ecological niche from morphometric trait data using a supervised machine learning algorithm, Random Forest (RF) analysis. All analyses were performed in r v.3.6.3 (R Core Team, 2020).

3.1 Assessing measured trait relevance (ANOVAs and PGLS)

While nearly all raw measured traits were found to be significantly different across foraging niche, nesting niche and ecomorph syndrome binnings, few traits appear to be implicated in functional role binnings (Figure 3a–c). When accounting for phylogeny, we find most measured traits remain implicated in niche aspects (Table S4). Comparatively, size-corrected ratios were less variable between binnings. Most ratios were still different between binnings before accounting for phylogeny, but the number of significantly different ratios was reduced afterwards (Table S5). Typically, procoxal-to-body-size proportion, eye-to-head-length proportion, pronotal expansion, pronotal flattening, scape-to-body-size proportion and metafemur-to-body-size proportion were all significantly different between binnings even after accounting for phylogeny, regardless of the niche aspect examined (Figure S9).

Considering functional role, specialist predators typically exhibit smaller procoxae and metafemura than omnivores, reflecting overall shorter leg length in comparison with the body (Figures S2a and S6). Predators also tend towards longer and straighter mandibles than phytophagivores and omnivores. Specialist predators typically have shorter antennae in comparison with the rest of the head, while omnivores, fungivores and generalist predators have longer antennae; phytophagivores typically have extremely short antennae. Eye position on the anteroposterior axis of the head capsule is variable, though typically predators have their eyes positioned somewhat more anteriorly than omnivores. Looking at nesting niche, lignicolous and subterranean nesters all have much shorter legs, both procoxa and metafemur, than ground-nesting ants, with leaf litter ants more variable overall (Figures S3 and S7). Arboreal ants, both lignicolous and carton-nesting, have much larger eyes than subterranean, leaf litter and ground-nesting ants. Lignicolous ants are also often distinguished by a flatter, broader mesosoma in comparison with ground-nesting ants, though this is not universal; and lignicolous ants have much shorter scapes than virtually all other groups, although subterranean ants typically possess shorter scapes as well. With respect to foraging niche, epigaeic ants overall have much longer legs, both procoxa and metafemur, than do leaf litter, arboreal or column-raiding ants (Figures S1 and S5). Scape length is also greater on average in epigaeic ants. Eye size is highly variable within groups but is greater among arboreal and epigaeic ants—arboreal ant eyes are the largest on average; column-raiding and leaf litter ants tend towards smaller compound eyes or eyelessness. Illustrating the overlap between lignicolous-nesting and arboreal-foraging species, arboreal-foraging ants are often distinguished by the broader, flatter mesosoma that typifies lignicolous ants.

3.2 Dimension reduction analyses

3.2.1 Agnostic visualization of niche structure

Both PCA and its corollary PCoA produced ecomorphospace with a high degree of overlap, especially when using raw measurements alone (Figure 4; Figures S10 and S11). The typical pattern produced by raw measurements was an area of high density where most of the specimens plotted, with some ecomorphs radiating into their own unique ecomorphospace. When using only raw measurements, no ecomorph was completely distinct, though typically specialist predators and omnivores plotted further away from one another.

For our raw measurement dataset, principal component 1 was body size, comprising 87.2% of the variance; principal component 2 was the overall size of the mandibles, comprising 5.2% of the variance. The first five principal components explain over 97% of the variance, each explaining >1% of the variance (Table S6). The pattern was similar with PCoA: the first five principal coordinates explain over 95% of the variance, with the first principal coordinate comprising 83.8% of the variance and the second principal coordinate comprising 5.0% (Table S8).

Using only size-corrected ratio measurements produced a slightly different pattern. There was no central area where most specimens plotted, but a greater degree of disparity throughout the ecomorphospace (Figures S12 and S13). While again, no ecomorphospaces separated clearly and there was a high degree of overlap, specialist predators tended to separate more distinctly from omnivores. Using ratio measurements, individual principal components accounted for much less of the overall variance; principal component 1 comprised only 33% of the variance, while principal component 2 comprised 24.3%. Over 90% of the variance is explained by the first seven principal components (Table S7). Most of the variance is explained by ratios concerning the scape, metafemur and mandibles; principal component 1 was driven mostly by scape and metafemur proportion, while principal component 2 was mostly mandibular proportions. Again, the pattern was similar using PCoA with size-corrected ratio measurements: the first principal coordinate comprised 34.1% of the variance and the second comprised 23.1% (Table S9).

3.2.2 Visualization of discrete niche structure

Our LDA produced ecomorph clusters with a variable degree of overlap, contingent on the ecological niche aspect being plotted and raw versus size-corrected ratio measurements (Figures 4 and 5; Figure S14). When analysing functional role, we found that again specialist predators and omnivores plotted separately, with phytophagivores also plotting out distinctly; while lignicolous and ground-nesting ants were most distinct; and arboreal, epigaeic and leaf litter-foraging ants all partially plotted in distinct ecomorphospaces. Similar to the PCAs and PCoAs, we recover a high-density area of ecomorphospace overlap, with some ecomorphospaces radiating out into unique topography.

Across ecological niche aspects, the most predictive traits for ecological niches are scape, metafemur and mandible proportions, eye position and raw trait measurements. Traits that did not strongly contribute to the major linear discriminants include overall head shape and mesosomal measurements, suggesting that these have a lower predictive value for discerning ecological niche categories. Generally, the first linear discriminant comprised 40%–55% of the total variance, with the second linear discriminant comprising 20%–35% of the variance, regardless of whether raw or size-corrected ratio measurements were considered (Tables S10–S17).

We used the posterior probabilities of the linear discriminant analyses for each ecological niche aspect to evaluate how accurately LDA models separated specimens into their known binnings. The results were variable (Table S18): overall the linear discriminant analyses predicted foraging niche and nesting niche from morphology with moderate accuracy when using the raw and size-corrected ratio measurement datasets. Functional role was poorly predicted, typically with only ~50% accuracy. Ecomorph syndrome predictions were variable depending on whether we used raw or size-corrected measurements: ecomorphs were predicted with moderate accuracy (70.83%) using raw measurements, but poorly (56.25%) using size-corrected ratios. When using datasets comprised of only PGLS-selected traits (traits shown to be significantly different between binnings even after accounting for shared ancestry), accuracy typically decreased, though this was contingent on niche aspect; for example, all traits were PGLS-selected for foraging and nesting niche.

3.3 Quantitatively predicting ecological niche from morphology

Random Forest analyses produced a range of accuracies, though consistently much higher than linear discriminant analyses. Both raw and size-corrected measurements predicted niche aspects and ecomorph syndrome with moderate to good accuracy, with out-of-bag (OOB) accuracy estimates from 77% to 85% (Table 2). Nesting niche and foraging niche were consistently predicted with greater accuracy than functional role or ecomorph syndrome, though these were not poorly predicted as with the LDAs. Again, functional role was most poorly predicted out of all niche aspects. There was no clear difference between the raw and size-corrected measurement datasets in terms of which dataset predicted niche aspects most accurately; this was variable from niche aspect to niche aspect, and differed in only a few percentage points. However, when using the PGLS-selected dataset—our most conservative set of traits—prediction accuracy rates typically decreased slightly (Table S19).

When visualizing the posterior probabilities of model classification using our t-SNE plots, we found that most binnings and ecomorph syndromes were well-separated, indicating that our Random Forest models performed well in discretely separating our binnings (Figures 4 and 5; Figure S15). In general, the greatest degree of overlap was typically seen in ecomorph syndromes, where leaf litter-nesting and foraging omnivores would cluster near ground/leaf litter-nesting epigaeic omnivores, and leaf litter-nesting epigaeic predators would cluster with ground-nesting epigaeic predators. This may reflect similar morphology adapted to overlapping nesting and foraging strata, with similar feeding mechanisms.

When using raw measurements, overall, the niche aspect models most frequently confused carton-nesting species with other binnings, particularly with ground-nesting species; fungivores with other binnings; and arboreal foragers with epigaeic foragers. The ecomorph syndrome model most frequently confused subterranean predators with column-raiding predators, leaf-litter omnivores with epigaeic ground-nesting omnivores; and carton-nesting arboreal omnivores with epigaeic ground-nesting omnivores. When using size-corrected ratio measurements, the niche aspect models most often confused leaf litter-nesters for ground-nesters; leaf litter foragers for epigaeic foragers; and granivores with other binnings. In terms of the ecomorph syndromes, model confusion was similar to models trained on raw measurements. These errors likely reflect similar morphologies for similar life habits, a flexibility in some life habits or minimal data for some groups—in binnings with fewer representative species, it was more difficult for the model to be trained sufficiently.

Restricting the number of terminal nodes indicated that the model was not highly overfitted (Figure S16). We tested this using the raw morphological trait measurement dataset predicting ecomorph syndromes. When restricted to only 20 terminal nodes, the model's accuracy was 64.08%, which is approximately 15% below maximum accuracy. The model approached maximum predictive accuracy extremely rapidly; at 50 terminal nodes it was only slightly below maximum accuracy, and it converged on maximum accuracy around 100 nodes. This prediction accuracy despite a highly constrained number of predictive pathways indicates a strong correlation between our sampled morphological traits and ecomorph syndromes.

Our Random Forest analysis identified several key traits as the strongest predictors of ecological niche; fundamental divisions in the classification system were based on these traits (Figure 3d–g; Figures S17 and S18). For nesting niche, the strongest predictors were eye, scape, procoxa and metafemur morphology both for raw measurements and size-corrected ratio data. Traits that were the strongest predictors of foraging niche were eyes, scapes, metafemura (again raw length and proportionate length) and thoracic shape. The strongest predictors of functional role were eyes, scapes, mandibles and metafemura (again raw length and proportionate length). In congruence with some LDA results, we found that traits such as mesosomal shape, pronotal shape and head shape typically had little contribution to the model for discerning ecological binnings, and other traits, such as eye position and body size had variable contributions depending on the ecological niche aspect being examined.

4.1 Functional morphology and ant ecomorphs

There are several morphological traits that we find to be indisputably linked to ecology. Many of these have been identified through other comparative approaches, which informed our own sampling (Gibb et al., 2015; Guilherme et al., 2019; Kaspari & Weiser, 1999; Parr et al., 2003; Weiser & Kaspari, 2006; Yates & Andrew, 2011; Yates et al., 2014). We build upon these previous key findings through multivariate analysis, which explicitly incorporates combinations of these traits to provide an integrative view of overall morphology as it relates to ecology. Collectively, these results quantitatively describe patterns in ant evolution that have been described by taxonomists and ecologists for well over a century (e.g. Wheeler, 1910). Our results recover suites of traits that are linked to ecology, define simplified ecomorph syndromes and provide relative assessment of individual trait relevance.

Arboreal ants, particularly lignicolous nesters, are typified by larger eyes, shorter and stubbier appendages (procoxae, scapes, etc.) and flatter, broader mesosomas. This is likely reflective of nesting in small cavities within trees and shrubs, and the advantage of being more appressed to foraging substrate. An exemplar is Melissotarsus, which cannot walk on flat surfaces outside of the wooden tunnels in which it lives (Khalife et al., 2018). Leaf litter ants are similar in these traits—their scapes and legs are shorter and their mesosomas more compact—but their eyes are usually small, given that their habitat is dark interstitial spaces on the forest floor.

Conversely, epigaeic ground-nesting ants tend to be greater in size and longer-legged, with larger eyes. With less complex foraging surfaces (Kaspari & Weiser, 1999), there is likely a selective pressure for faster-moving ants to improve rates of resource acquisition and predator avoidance. Here, Cataglyphis is an archetype with elongated legs and eyes in the context of its niche as an obligate epigaeic forager in the Saharan desert (Wehner, 1983). Some ecomorphs are an amalgamation of conflicting traits: column-raiding predators may occupy both subterranean as well as epigaeic habitats and so can be long-legged and elongated relative to ants that nest in complex interstitial spaces. However, the reduction in eyes and typically shortened scapes are consistent with subterranean nesting. We find that predators trend towards longer and flatter mandibles, and omnivores towards shorter, curved mandibles. Prey capture may be easier with larger stouter mandibles, while shorter curved mandibles may be adapted to the acquisition and manipulation of diverse food items, from seeds to honeydew.

When considering global ant morphology in the context of ecology, we are able to make relative comparisons of ecomorphological proximity. Proximate ecomorphospace indicates similarity in morphology, while overlap reveals trait morphology that is applicable to multiple ecologies. Across our analyses, we find that ecomorphospaces overlap in a core region, where a large proportion of species are clustered. These ants reflect a generalized platonic ant morphology, where they are not strongly morphologically specialized relative to other taxa. The species occupying this common space are often omnivorous, ground-nesting and epigaeic, suggesting that these niche binnings are associated with more generalized morphology.

Some ecologies straddle both this common space as well as unique topography, while others occupy primarily unique morphospace. Species that occupy distinct morphospace display a suite of traits that are strongly correlated with niche binning, such as Pseudomyrmex (lignicolous nesting, arboreal foraging omnivore), Liometopum (carton-nesting arboreal foraging omnivore) and Strumigenys (leaf litter nesting and foraging predator) species.

We find that multidimensionality greatly aids in the characterization of ecomorphs. Species displaying shorter and stubbier appendages could be arboreal, leaf litter or subterranean dwelling. It is only after we incorporate other traits, such as eye size and mesosomal shape, that we can more accurately predict its likely ecological niche. Collapsed or summarized ecomorph syndromes improve prediction by simultaneously considering multiple ecological niche aspects and accommodating some plasticity. Ecomorphs are often defined by integration of traits moreso than individual measurements; for example, when using random forest analysis to predict only nesting niche, carton-nesters are often inaccurately classified as other niche binnings, and in dimension reduction analyses carton-nesters typically plot in a generalized platonic ecomorphospace. However, when the integrated ecomorph of the carton-nesting arboreal foraging omnivore is considered, species are consistently predicted accurately.

Our collapsed ecomorphs are convergently evolved in ants. Each of our 10 sampled ecomorph syndromes span multiple phylogenetic origins (Figure 6). Ecomorphs are typified by whole-body convergent morphology: arboreal dorylines strongly resemble arboreal pseudomyrmicines despite being separated by ~100 million years (Borowiec et al., 2019). Morphological convergence on shared ecomorphs between phylogenetically disparate taxa suggests predictability in the evolution of these ecomorphs and convergent selective pressures (Mahler et al., 2013). This predictability reinforces findings of repeated evolution of ecomorphs across lineages, where taxa evolve a suite of specific traits to occupy available niches (Cooper & Westneat, 2009; Gillespie et al., 2018). Moreover, our analyses and previous assessments of trait variance in a phylogenetic context indicate that morphological similarity within niche binnings is not the result of shared ancestry (Figure 3).

4.2 Predicting ecological niche from morphology

We find that ecological niche aspects or ecomorphs can be predicted from morphology with 77%–85% accuracy, contingent on raw measurement or size-corrected ratio input data. This accuracy is comparable to other studies incorporating RF analysis (Pigot et al., 2020; Rojas et al., 2012; Rossel & Arbizu, 2018). Even as our number of binnings ranges between four (foraging niche) and ten (ecomorph), we recover high predictive accuracy for all our models, which suggests a high degree of trait similarity within ecological binnings. Importantly, RF prediction incorporates nonlinear relationships of traits, which improves accuracy relative to techniques limited to linear relationships such as LDA (Breiman, 2001; Pigot et al., 2020). The predictive power of RF illustrates the strong correlation between organism-wide morphology and ecology.

While there is a core group of morphological traits that our models found were most important in classifying various binnings and syndromes, there were some distinct differences between niche aspects. Consistently, raw eye size and eye size in relation to body size were most important across all niche aspects and ecomorph syndromes: this reflects the extreme consistency of eye size within binnings, with subterranean ants bearing minute eyes or entirely eyeless; leaf litter ants with smaller eyes; epigaeic ants with moderately sized to large eyes, and arboreal ants with extremely large eyes. The importance of eye size is reflected even in functional role classification, perhaps reflecting correlation between certain feeding habits and foraging or nesting microhabitats. Models ranked mandible size in relation to body size and raw mandible size as a highly important variable in functional role binning predictions, possibly reflecting the importance of mandibles in feeding habits. Leg (procoxa and metafemur) length and scape length were consistently important in all niche aspect binning classification, reflecting the importance of appendage size in nesting microhabitat and foraging in interstitial spaces especially.

Model misclassification is also informative; our models typically misclassify niches that exhibit some degree of morphological overlap or ecological plasticity. One of the most frequently misclassified ecomorphs is the subterranean predator ecomorph, often misclassified as a column-raiding predator. Column-raiding predators have a mixed nesting niche where they will occasionally make temporary subterranean nests, in between periods of nomadism where they form bivouacs for nesting (Borowiec, 2016). Morphological adaptations that are implicated in column-raiding could also relate to subterranean-nesting—column-raiders are typically highly dependent on chemical foraging trails and are often eyeless or bear minute eyes, another feature of subterranean dwellers.

Our models also generate misclassifications between species that forage and nest variably in epigaeic surficial and leaf litter habitats. This may be a result of plasticity in nesting strata: some species may nest in leaf litter structures, such as logs, and also construct ground nests. Indeed, these plastic ecologies may reflect the same pattern as overlapping ecomorphospaces plotted by our LDAs. A ‘platonic ecomorph’ will bear more generalized traits, and therefore is less easily classified using machine learning.

4.3 A pipeline for predictive ecomorphological modelling

We include in our Supporting Information both an expandable morphometric dataset and template code to provide a pipeline for future ecomorphological modelling. For any specimen databank, workers can be measured according to traits sampled here. The predictive ecomorphological model is then trained on this expanded dataset, and can be used to evaluate hypotheses related to ecological diversity. We include both LDA model code and Random Forest model code, though Random Forest typically generates a higher predictive accuracy.

Increased taxonomic sampling in future datasets will further illuminate links between form and function in ants. Moreover, future intraspecific sampling of distinct castes may reveal more complex trends related to the capacity for lineages to become ecologically specialized. Here, we provide two options for predicting the ecomorph of a worker: a simplified set of ecomorph syndromes, or for more granular analysis, classification of worker functional role, foraging niche and nesting niche. Additionally, new ecological dimensions may be incorporated including alternative ecomorph syndromes, abiotic features or ethology.