Ant-gardens (AGs) are ant/plant mutualisms in which ants farm epiphytes in return for nest space and food rewards. They occur in the Neotropics and Australasia, but not in Africa, and their evolutionary assembly remains unclear. We here use phylogenetic frameworks for important AG lineages in Australasia, namely the ant genus Philidris and domatium-bearing ferns (Lecanopteris) and flowering plants in the Apocynaceae (Hoya and Dischidia) and Rubiaceae (Myrmecodia, Hydnophytum, Anthorrhiza, Myrmephytum and Squamellaria). Our analyses revealed that in these clades, diaspore dispersal by ants evolved at least 13 times, five times in the Late Miocene and Pliocene in Australasia and seven times during the Pliocene in Southeast Asia, after Philidris ants had arrived there, with subsequent dispersal between these two areas. A uniquely specialized AG system evolved in Fiji at the onset of the Quaternary. The farming in the same AG of epiphytes that do not offer nest spaces suggests that a broadening of the ants' plant host spectrum drove the evolution of additional domatium-bearing AG-epiphytes by selecting on pre-adapted morphological traits. Consistent with this, we found a statistical correlation between the evolution of diaspore dispersal by ants and domatia in all three lineages. Our study highlights how host broadening by a symbiont has led to new farming mutualisms.
Farming mutualisms, wherein an organism promotes the growth of another on which it depends for food, have evolved in many lineages of the tree of life [1,2]. Examples include bacteria farmed by amoebae , fungi  or deep-sea crabs , and algae farmed by sloths or damselfish [6–8]. More complex forms of farming with active control of the partner's growth, reproduction and dispersal have evolved in fungus-farming ants [1,9–11], beetles [12,13] and termites [14,15]. Beyond fungi, ants also farm a range of other organisms (see  for a review). Control over the dispersal of the ‘crop’ by the ‘farmer’ is a first step required in the evolution of farming mutualisms. In ants, mutualisms between seed-dispersing workers and plants with suitable propagules have been documented or inferred for thousands of species in 77 families of flowering plants [16,17]. Typically, these interactions involve ants gathering seeds to feed on lipid- and protein-rich appendages (elaiosomes) and then abandoning the seeds somewhere nearby . In the Neotropics, ants also disperse certain epiphytes by placing their seeds inside carton nests, where they germinate and eventually form the so-called ant-gardens (AGs) [19–21]. The epiphytes benefit from being dispersed to sites far above the ground and planted in the nutrient-rich carton material (often enriched with vertebrate faeces), and the ants benefit through the scaffold formed by epiphyte roots that stabilizes their nests and by nutritional rewards, such as extrafloral nectar .
The most specialized AGs are farming mutualisms because seed planting, cultivation, fertilization or defence against pathogens or insects , as well as harvesting of the crop, all are present. Farming mutualisms require seed recognition, here mediated by chemical cues, and convergent odours have apparently driven the assembly of Neotropical AGs [23–27]. Surprisingly, Neotropical AGs do not involve domatium-bearing plants, while Australasian and Southeast Asian AGs are dominated by such species; in the Neotropics, the ants therefore only nest between roots while in Asia, they are provided nesting space within their cultivated plants. This raises the question of how the assembly of Australasian and Southeast Asian AGs began. Did the assembly start with the cultivation of plants for food (extrafloral nectar or fruits but without domatia) or from domatium-based symbioses?
AGs in Southeast Asia and Australasia involve at least five genera of ants that build carton nests (Philidris, Crematogaster, Pheidole, Camponotus and Diacamma) and 17 genera of plants with together about 78 species (, this study), while Fijian AGs consist only of six plant species . About 38 of these species belong to the Rubiaceae (e.g. Hydnophytum, Myrmecodia and other genera in subtribe Hydnophytinae), the Apocynaceae (Hoya, Dischidia) and the fern genus Lecanopteris (figure 1). The Hydnophytinae are also the largest clade of ant–plants worldwide . To trace the evolutionary history of Australasian and Southeast Asian AGs, we generated phylogenetic frameworks for Lecanopteris, Hoya, Dischidia and the Hydnophytinae, as well as their most commonly associated ant genus, Philidris. Based on clock-dated phylogenies for the ants and the plants, we aimed to answer the following questions: (i) when, where, and in which sequence did AG epiphytes and AG-forming Philidris ants originate, and (ii) which trait combinations favored the evolution of AGs in Australasia and Southeast Asia?
2. Material and methods
Taxon sampling details are provided in the electronic supplementary material, Materials and Methods.
Total genomic DNA was extracted from around 20 mg of leaf tissue, using a commercial plant DNA extraction kit (NucleoSpin; Macherey–Nagel, Düren, Germany) according to manufacturer protocols. Polymerase chain reaction (PCR) was performed using Taq DNA polymerase (New England Biolabs, Cambridge, MA, USA) and a standard protocol (39 cycles, annealing temperature 56°C). PCR products were purified using the ExoSap clean-up kit (Fermentas, St Leon-Rot, Germany), and sequencing relied on Big Dye Terminator kits (Applied Biosystems, Foster City, CA, USA) on an ABI 3130 automated sequencer (Applied Biosystems, Perkin-Elmer). Sequences were edited in Sequencher v. 5.1 (Gene Codes, Ann Arbor, MI, USA). All new sequences were BLAST-searched in GenBank. Sequence alignment was performed in MAFFT v. 7 in the online server (http://mafft.cbrc.jp/alignment/server)  under default parameters except for the internal transcribed spacer region, which was aligned under Q-INS-i optimization, which takes rRNA secondary structure into consideration. Minor alignment errors were corrected manually in Mesquite v. 2.75 . Maximum-likelihood (ML) inference relied on RAxML v. 8.0  and the GTR + Γ substitution model, with empirical nucleotide frequencies and 25 gamma rate categories; bootstrap support was assessed from 100 replicates under the same model. We also conducted Bayesian inference in MrBayes v. 3.2  under the substitution model selected by jModeltest2 for each marker  and using the program default of two runs and four chains (one cold and three heated), with the uniform default priors. We set a 10 × 106 Markov chain Monte Carlo (MCMC) chain, sampling trees every 1000th generation. Split frequencies approaching zero indicated convergence. We used the 50% consensus tree to assess posterior probabilities for nodes of interest.
(a) Molecular-clock dating
Molecular dating analyses relied on BEAST v. 2  and uncorrelated lognormal relaxed clock models unless otherwise stated. We used the GTR + G substitution model with four rate categories and a Yule tree prior. For both our plant and ant trees, MCMCs were run for 20 million (Lecanopteris) or 40 million (Philidris, Hoya-Dischidia, Hydnophytinae) generations, with parameters and trees sampled every 10 000 generations. We used Tracer v. 1.6  to check that the effective sample size (ESS) of all parameters was more than 200, indicating that runs had converged. After discarding 20% as burn-in, trees were summarized in TreeAnnotator v. 1.8 (part of the BEAST package) using the option ‘maximum clade credibility tree’, which is the tree with the highest product of the posterior probability of all its nodes. The final tree was visualized in FigTree v. 1.4 . Calibrations details are provided in the electronic supplementary material, Materials and Methods.
(b) Ancestral area reconstruction
Based on the geological history of Southeast Asia [38,39], we defined nine geographical units relevant to the focal clades: (i) Fiji and Vanuatu (since there was only a single species from Vanuatu); (ii) the Solomon Islands; (iii) the Bismarck archipelago, including New Ireland, New Britain, Normanby Island and D'Entrecasteaux Islands; (iv) Northern Papua New Guinea inclusive of and limited by the Maoke range in Indonesian Papua and the Bismarck range in Papua New Guinea; (v) South Papua New Guinea (south of the Maoke range in Indonesian Papua and the Bismarck range in Papua New Guinea) and Australia; (vi) Sulawesi, Moluccas and the lesser Indonesian Islands (limited by Java in the west); (vii) the Philippines; (viii) Sundaland (including Indochina, Sundaland, Borneo, Java, Sumatra that were connected until the Quaternary [38,39]; we also include continental Asia as far west as India since our focal clades are younger than 20 Ma, a time during which these area were constantly connected; and (ix) Pacific islands (Hawaii, Wallis and Futuna, French Polynesia), in which we also included New Caledonia because only two outgroups are native from this island.
To infer the ancestral areas of Philidris-inhabited AG plants and to probe whether ancestral areas of Philidris match those of their plant hosts, we used ancestral range reconstruction as implemented in the R package BioGeoBEARS [40,41] on the BEAST chronograms. BioGeoBEARS infers ancestral geographical ranges and permits comparison of three biogeographic models, namely dispersal–extinction–cladogenesis (DEC), dispersal–vicariance (DIVALIKE) and BAYAREA (BAYAREALIKE). Founder-event speciation is modelled via a speciation parameter j that can be added to each of the models. We selected the best model based on log likelihood values as well as the Akaike information criterion (ΔAICc). BioGeoBEARS statistics are shown in the electronic supplementary material table S5.
Data on interactions between dolichoderine ants and plant species in our three focal clades came from the field observations of Eva Kaufman  and the two first authors. A few additional interaction links came from field observations by Matthew Jebb in Papua New Guinea (including Indonesian Papua; , table 10.2) or from matching geography and morphological traits; the latter shown as dashed lines in figure 2. When the ant phylogeny includes multiple accessions per species, we only added putative links to a single specimen.
(d) Correlated evolution of dispersal by ants with other mutualistic traits and ancestral state reconstructions of mutualistic traits
To test whether the evolution of dispersal by ants correlates with other mutualistic traits, we used BayesTraits v. 2 , which allows detecting correlated evolution between pairs of discrete binary traits. We studied the following traits, all treated as binary: ant inhabitants (score ‘0’ when no associations with ants are formed or only facultative associations with generalist ants, ‘1’ for consistent association with one or few specialized ant partners), domatium growth (scored ‘0’ for diffuse growth, and ‘1’ for apical growth), entrance holes (scored ‘0’ when absent (no tuber) or smaller than 1 cm in diameter (when tuber present), and ‘1’ when larger than 1 cm in diameter), warts, absorptive structures inside Hydnophytinae domatia , (scored ‘0’ when present, and ‘1’ when absent), first domatium cavity (scored ‘0’ when it enlarges throughout development, ‘1’ when its development is determinate), post-anthetic sugar rewards (scored ‘0’ when absent, and ‘1’ when present), seed dispersal (scored ‘0’ for birds, other animals or gravity alone and ‘1’ for dispersal by ants). Trait states were coded based on Huxley , Jebb [42,45], Davidson & Epstein , Huxley & Jebb [46–49], Maeyama & Matstumoto , Chomicki & Renner  and Chomicki et al. , an unpublished revision of Hydnophytum from M. Jebb and C. R. Huxley, personal observations by G.C. 2014, 2015, personal communications to G.C. from M.H.P. Jebb (March 2015) and C. Huxley-Lambrick (October 2015). We tested the correlation of every pair of traits. We used the maximum clade credibility (MCC) tree from BEAST but pruned the outgroups and first ran a model of independent trait evolution and estimated the four-transition rate parameters α1, α2, β1, β2, wherein double transitions from state 0,0 to 1,1 or from 0,1 to 1,0 are set to zero. We then ran a model of dependent trait evolution with eight parameters (q12, q13, q21, q24, q31, q34, q42, q43). To compare these non-nested models, we calculated the Bayes factor score.
We used stochastic character mapping to infer possible histories of mutualistic traits not only at nodes but also along branches in the phylogenies. We relied on the function ‘make.simmap’ in the phytools package (v. 04–60) , which implements the stochastic character mapping approach developed by Bollback . We estimated ancestral states using: (i) an equal rate (ER) model and (ii) an all rate different (ARD) model, and then simulated 1000 character histories on the MCC trees from BEAST. We summarized the 1000 simulated character histories using the function densityMap (also in phytools).
(a) Phylogenetic relationships in Australasian and Southeast Asian dolichoderines and their host plants
The 5-gene phylogeny of Dolichoderinae ants recovered the genus Philidris as monophyletic with maximal statistical support while the internal topology had little statistical support (electronic supplementary material, figure S1), with the exception of a few nodes, importantly the Philidris nagasau clade from Taveuni Island (Fiji), which was strongly supported (maximum-likelihood bootstrap support (ML BS) = 81%, Bayesian posterior probability (pp) = 1). Nevertheless, we found strong biogeographic signal. Phylogenetic relationships in Lecanopteris and Hoya-Dischidia are discussed in previous studies [55,56]. Our 6-gene Hydnophytinae phylogeny yielded a strongly supported tree (electronic supplementary material, figure S2) and implies a single origin of the characteristic modified hypocotyl tuber domatium (figure 1a,d,f). The ant-dispersed lineages are recovered with high support: Myrmecodia (ML BS = 89%, pp = 0.99), H. formicarum (71/0.99), H. moseleyanum (100/1), Myrmephytum arfakianum (99/1), Philidris nagasau-inhabited Squamellaria clade (85/1), except for the Anthorrhiza chrysacantha/A. caerulea clade (60/0.55), probably because of missing data for A. chrysacantha.
(b) Times of origin of ant-dispersed epiphytes and their dispersers
Our molecular-clock dating estimates are congruent with previous estimates for the Apocynaceae , Rubiaceae  and leptosporangiate ferns . For the ants, no earlier molecular-clock dating is available; our secondary calibrations came from a 10-gene Dolichoderinae phylogeny calibrated with six fossils . The most recent common ancestor of Philidris originated at 13.2 ± 7 Ma, older than the age of the oldest ant-dispersed Hydnophytinae lineage (Myrmecodia) (6.3 ± 3 Ma), but Philidris clades of younger ages also inhabit Myrmecodia (figure 2). The ant-dispersed Anthorrhiza clade dates to 5.8 ± 2 Ma. The Fijian P. nagasau clade dates from 3.5 ± 2 Ma, matching the age of its obligate plant host species in the genus Squamellaria dating to 2.5 ± 1.5 Ma (figures 2 and 3). The Hydnophytum moseleyanum group originated some 4.9 ± 2 Ma. Epiphyte seed dispersal independently evolved in Anonychomyrma more recently, some 2.6 ± 2 Ma (figures 2 and 3). The crown age of Lecanopteris is 4.8 ± 2 Ma, and that of ant-dispersed Dischidia 4.9 ± 3 Ma. A minimum of four AG Hoya lineages originated during the last 2 Myr (figures 2 and 3).
(c) Biogeographic analysis
The biogeographic model comparison yielded the DEC + J model as best fitting for Philidris (−36.47 versus −44.17 for DEC), the Hoya-Dischida data (−79.44 versus −91.46 for DEC), the Hydnophytinae (LnL = −130.53 versus −170.05 for DEC) and Lecanopteris (−25.53 versus −27.07) (electronic supplementary material table S5). Philidris likely originated in Australia because during the Mid-Miocene, Southern New Guinea was still submerged [38,39]. The decrease in sea level in the Late Miocene resulted in the connection of Australia to the emerging Southern New Guinea [38,39], enabling Philidris to expand its range by ca 10 Ma and providing a stepping-stone for the colonization of Northern New Guinea between 10 and 5 Ma (figure 2), which at the time formed several disconnected landmasses [38,39]. From Australia-Southern New Guinea, Philidris colonized Sundaland around 10 Ma (figure 2), including today's Borneo, peninsular Malaysia, Sumatra, Java and continental Asia [38,39]. Fiji was apparently colonized by long-distance dispersal also from Australia-Southern New Guinea (figure 2), and since P. nagasau is endemic from Vanua Levu and Taveuni, the maximal age of the colonization is 4 Ma; Vanua Levu and Taveuni only emerged 0.8 Ma [61,62]. Alternatively, the colonization of Fiji could have happened as early as 6.5 ± 3 Ma (P. nagasau stem age, figure 2) on Viti Levu (the oldest, and largest Fijian island), and then Philidris would have gone extinct on this island.
The Hoya–Dischidia clade apparently originated in continental Asia, and underwent at least two long-distance dispersal events to Southern New Guinea/Australia in the Pliocene and Quaternary (figure 2). The Philippines were recurrently colonized from Sundaland during the same period (figure 2). The Hydnophytinae clade apparently originated in northern New Guinea some 15.5 Ma, at a time when it was separated from Australia . An early dispersal event occurred from New Guinea to the South Pacific archipelago of Fiji and Vanuatu (figure 2, ). The Solomon Islands were colonized by dispersal, not vicariance, since the Solomon arc had drifted from the Fiji–Vanuatu arc some 12 Ma [61,62]. While most of the Hydnophytinae radiation (ca 75%) is endemic from New Guinea, a number of dispersal events to islands off Moluccas, Philippines, Sulawesi and Sundaland occurred during the last 5 Myr. For the fern Lecanopteris, we were unable to infer its detailed biogeographic history owing to a few widespread species (figure 2) for which phylogeographic work would be required.
(e) Correlated evolution of mutualistic traits
Of the 21 pairwise correlations between each of the seven binary traits (ant inhabitants, domatium growth type, entrance hole size, wart presence or absence, first cavity enlargement, post-anthetic sugar rewards presence or absence, dispersal type), 11 were very strongly correlated (Bayes factor > 10), one was strongly correlated (5 < Bayes factor < 10), three were positively correlated (Bayes factor > 2) and six were not correlated (figure 4). Dispersal type (presence or absence of dispersal by ants) was correlated with all mutualistic traits except the size of entrance hole. The plots summarizing the stochastic mapping of ancestral states are shown in figure 4.
(a) The spatio-temporal assembly of Australasian ant-gardens
The assembly of the 13 AG plant lineages studied here occurred independently in three areas, two of which already harboured Philidris (New Guinea/Australia and Sundaland) and one where the inferred time of colonization by Philidris coincides with the inferred evolution of AG plant lineages (Fiji) (figures 2 and 3). Surprisingly, the Philidris crown age predates that of the oldest AG lineages by 6 Myr. Perhaps Philidris plant farming only evolved in the very Late Miocene, about 6.3 Ma, when the oldest domatium-bearing lineage, Myrmecodia, first diversified on New Guinea. This would imply that Philidris lineages initially did not nest in plant domatia. Indeed, years of field observations across Papua New Guinea (2004–2014) by M.J. have revealed much variation in these ants in terms of nest formation and acceptable nest spaces, mostly with low plant host specificity. Alternatively, previous AG communities dissolved and their plant lineages went extinct or are not sampled here. The second-oldest domatium-bearing lineage in our sample, Dischidia, is from Sundaland and dates to 3–5 Ma, followed by the 2–5 Ma old Squamellaria clade from Fiji, which today is exclusively and obligatorily inhabited by P. nagasau  and whose crown age matches that of Fijian Philidris (figures 2 and 3).
New AG lineages were repeatedly recruited following broadening of the host range. This appears to have occurred in New Guinean Hydnophytinae (H. moseleyanum species group) and also among distantly related groups, such as Hydnophytinae and Lecanopteris ferns, although it remains unclear whether ants disperse the spores of this fern . The non-domatium-bearing orchid Dendrobium insigne has also been recruited into AGs . In Sundaland, host broadening also led to multispecies gardens with non-domatium-bearing species, including orchids, Araceae, Apocynaceae, Gesneriaceae, Melastomataceae, Moraceae, Piperaceae, Polypodiaceae, Urticaceae and Zingiberaceae , growing together with domatium-bearing species from the three clades studied here (figures 2 and 3). Dispersal events must be responsible for the AG lineages now found in Sundaland, New Guinea and Australia (figures 2 and 3b) and led to the origin of an additional epiphyte-farming lineage within the Dolichoderinae genus Anonychomyrma (figures 2 and 3). Anonychomyrma and Philidris have different niches (dry vegetation versus rainforest, lowlands versus high altitude areas [42,44], M. Janda 2014, personal observation), which probably helped the expansion of AGs in New Guinea.
(b) Mutualism specialization following host broadening onto pre-adapted plant lineages
The correlation between the evolution of ant domatia and seed dispersal by ants (figures 2 and 4), together with the inference that the oldest Australasian ant-dispersed epiphyte lineages were domatium bearing, suggests that these AGs evolved from domatium-based symbioses, not food-only-based mutualisms, answering our second research question. Nevertheless, the older age of the Philidris lineage than any of the plant lineages implies the earlier existence of AGs without domatium-bearing plants or at least without the domatium-bearing lineages studied here. That dispersal by ants and rewards for ants (in the form of nectar) are correlated in the oldest domatium-bearing lineage here studied, the Rubiaceae clade Hydnophytinae (figure 4), suggests that food rewards may be a pre-adaptation for AG lineages, as they are more generally for farming mutualisms . Other pre-adaptations for the convergent evolution of AG plants and ants may include chemical compounds in the seeds [25–27], making them attractive to ants, and polydomy , an ant species' ability to form several interlinked nests, which allows negotiating the lag-time between seed-planting and rewards (domatia or food), essential in any farming mutualism. The recurrent recruitment of new epiphyte lineages into the AGs probably promoted the evolution of domatia, which enhance uptake of ant-derived nutrients [44,64,65]. This leads to the, perhaps counterintuitive, conclusion that the broadening of host ranges by the ants led to the specialization of new plant hosts.
Our study reveals the assembly of tropical Asian AGs through time and space. We inferred a minimum of 13 independent origins of AG plants during the last 6–7 Myr. Our dating and biogeographic analyses revealed that AGs evolved independently in Australasia and Southeast Asia where the often domatium-nesting Philidris ants were already present, with subsequent dispersal between these two areas and a uniquely specialized AG system evolving in Fiji. The host broadening by Philidris ants resulted in the recurrent entry of diverse (probably pre-adapted) plant lineages into the AG ‘adaptive zone’ (sensu Simpson ), resulting in the specialization of additional hosts. That the ants may be the key driver both in AGs (this study) and in terrestrial ant/plant symbioses  matches the overall asymmetry of ant/plant symbioses worldwide. Such symbioses involve 113 species of ants and 684 species of vascular plants , highlighting that generalists can drive mutualism specialization, consistent with the asymmetric specialization found in mutualistic networks where a subset of specialists interacts with generalists [68,69].
Data are available in TreeBase accessions http://purl.org/phylo/treebase/phylows/study/TB2:S1476 (Dischidia and Hoya) and Dryad accession http://dx.doi.org/10.5061/dryad.4tv0s  (Philidris, Hydnophytinae and Lecanopteris).
G.C. designed study and analysed the data; G.C. and M.J. generated the data; G.C. and S.S.R. wrote the manuscript with edits from M.J.; all authors provided reagents.
We declare we have no competing interests.
This work was supported by a grant from the German Research Foundation (DFG), RE 603/20, and grants from the Society of Systematic Biologists and the American Association of Plant Taxonomy to G.C. M.J. was supported by a grant from Czech Science Foundation: P505/12/2467 and the Marie Curie IOF PIOFGA2009-25448.
We thank Eva Kaufmann for Philidris samples, Matthew Jebb and Camilla Huxley-Lambrick for discussion, Andreas Wistuba for samples of cultivated plants, Jeremy Aroles for proofreading the manuscript, and William Baker for identifying the palm in figure 1c. M.J. extends his gratitude to the staff of the New Guinea Binatang Research Center for field assistance, to V. Novotny and N. Pierce for assistance with the Papua New Guinea project and to M. Borovanska and P. Matos-Maravi for assistance with molecular data.
One contribution to a special feature ‘Ant interactions with their biotic environments’.
Electronic supplementary material is available online at https://dx.doi.org/10.6084/m9.figshare.c.3670204.
- Received August 8, 2016.
- Accepted October 19, 2016.
- © 2017 The Author(s)
Published by the Royal Society. All rights reserved.