Effects of seed morphology and elaiosome chemical composition on attractiveness of five Trillium species to seed‐dispersing ants

Abstract Morphological and chemical attributes of diaspores in myrmecochorous plants have been shown to affect seed dispersal by ants, but the relative importance of these attributes in determining seed attractiveness and dispersal success is poorly understood. We explored whether differences in diaspore morphology, elaiosome fatty acids, or elaiosome phytochemical profiles explain the differential attractiveness of five species in the genus Trillium to eastern North American forest ants. Species were ranked from least to most attractive based on empirically‐derived seed dispersal probabilities in our study system, and we compared diaspore traits to test our hypotheses that more attractive species will have larger diaspores, greater concentrations of elaiosome fatty acids, and distinct elaiosome phytochemistry compared to the less attractive species. Diaspore length, width, mass, and elaiosome length were significantly greater in the more attractive species. Using gas chromatography–mass spectrometry, we found significantly higher concentrations of oleic, linoleic, hexadecenoic, stearic, palmitoleic, and total fatty acids in elaiosomes of the more attractive species. Multivariate assessments revealed that elaiosome phytochemical profiles, identified through liquid chromatography–mass spectrometry, were more homogeneous for the more attractive species. Random forest classification models (RFCM) identified several elaiosome phytochemicals that differed significantly among species. Random forest regression models revealed that some of the compounds identified by RFCM, including methylhistidine (α‐amino acid) and d‐glucarate (carbohydrate), were positively related to seed dispersal probabilities, while others, including salicylate (salicylic acid) and citrulline (L‐α‐amino acid), were negatively related. These results supported our hypotheses that the more attractive species of Trillium—which are geographically widespread compared to their less attractive, endemic congeners—are characterized by larger diaspores, greater concentrations of fatty acids, and distinct elaiosome phytochemistry. Further advances in our understanding of seed dispersal effectiveness in myrmecochorous systems will benefit from a portrayal of dispersal unit chemical and physical traits, and their combined responses to selection pressures.


K E Y W O R D S
ants, chemical ecology, insect-plant interaction, myrmecochory, oleic acid Kruger, & Linsenmair, 2004;Heil, Fiala, Kaiser, & Linsenmair, 1998;Paiva, Buono, & Lombardi, 2009;Rickson, 1976;Shenoy, Radhika, Satish, & Borges, 2012;Webber, Abaloz, & Woodrow, 2007), and these food resources are typically not as attractive to generalist ants as they are to the plants' specific ant mutualists Davidson, Foster, Snelling, & Lozada, 1991;Gonzalez-Teuber & Heil, 2009;Heil et al., 2004Heil et al., , 1998. Furthermore, ant preferences for specific chemical compounds in EFN are highly species-specific, suggesting mutualists exert strong stabilizing selection on EFN chemotypes. Although the myrmecochory is a facultative mutualism rather than an obligate mutualism as in tropical myrmecophytic systems, it is possible that seed-dispersing ants nevertheless exercise a similar stabilizing selection on elaiosome phytochemical profiles such that a certain narrow chemotype is defined. This hypothesis has not been tested in myrmecochorous systems; in particular, studies are lacking that characterize the metabolomic profiles of elaiosome phytochemistry and assess whether single compounds or entire phytochemical profiles affect diaspore retrieval and dispersal to the nest by ants.
Despite the fact that three main aspects of diaspores (morphological, nutritional, chemical) have been linked to seed dispersal by ants, rarely have multiple diaspore traits been investigated in the same study (but see Leal et al., 2014), and the most important mechanisms governing ant preference remain unclear. We explore whether differences in diaspore morphology, elaiosome fatty acids, or elaiosome phytochemical profiles explain the differential attractiveness of diaspores of plants in the genus Trillium (Order Liliales, Family Melanthiaceae) to ants, primarily of the keystone seed-dispersing genus Aphaenogaster. Although a number of ant species disperse seeds of myrmecochores in eastern North America (Gaddy, 1986;Zelikova, Sanders, & Dunn, 2011), Aphaenogaster ants are the primary seed dispersal vector for most temperate ant-dispersed flora in this region and are responsible for approximately 74% of myrmecochore seed collection in forests where Aphaenogaster have been reported (Ness et al., 2009). Our study system is comprised of five eastern North American Trillium species, two of which are geographically widespread and three of which are narrowly endemic; herein, pairs of widespread and endemic species co-occur in multiple sites across the southern Appalachian region of the United States. In this system, Aphaenogaster ants display a preference for dispersing the diaspores of the geographically widespread trilliums compared to their co-occurring, endemic congeners (Miller & Kwit, 2018). Using field-based observations of seed dispersal probabilities from Miller and Kwit (2018), we ranked each of the five study species from least attractive to most attractive (see Materials and Methods below) and use this scale of attractiveness as a comparative framework for exploring diaspore morphology and elaiosome chemistry throughout the present study.
Our first objective is to quantify interspecific differences in diaspore morphology among our study species. We predict that the more attractive species of Trillium will have larger diaspores overall, and higher elaiosome mass and elaiosome-seed mass ratios than less attractive species. Our second objective is to assess interspecific differences in concentrations of elaiosome fatty acids. We predict that concentrations of key fatty acids, including oleic acid, will be higher in the elaiosomes of the more attractive species. Our third objective is to assess interspecific differences in elaiosome phytochemical profiles. We predict that elaiosome phytochemical profiles of the more attractive species will be more tightly clustered in multivariate space than those of less attractive species, reflecting greater stabilizing selection by ant dispersers.

| Study species and sites
Eastern North America is a biodiversity hotspot for the plants in the genus Trillium-a group of myrmecochorous (ant seed dispersed) perennial understory forest herbs-with at least 29 species occurring in this region (Freeman, 1975;Ohara, 1989). In the southern Appalachian region, many species of Trillium are sympatric. Although co-occurring myrmecochore species often temporally stagger fruiting (Gordon, Meadley-Dunphy, Prior, & Frederickson, 2019;Warren, Giladi, & Bradford, 2014), the sympatric species studied here have overlapping flowering and fruiting phenology (Miller & Kwit, 2018).
Due to the spatial proximity of plants at our study sites, foraging ants occasionally come across mature diaspores of co-occurring congeners at the same time (C. N. M., personal observation), resulting in potential interspecific competition for dispersal services. We investigated the importance of morphological and chemical diaspore attributes in determining seed attractiveness to foraging ants using the five species of Trillium studied in Miller and Kwit (2018). These included the geographically widespread species Trillium catesbaei Elliott and T. cuneatum Raf., as well as the range-restricted, endemic species T. lancifolium Raf., T. discolor Wray ex Hook, and T. decumbens Harbison. Pairs of these species co-occur in multiple locations throughout the southern Appalachians (see Miller & Kwit, 2018).
We located six study sites in spatially distinct forest stands (average size of 130 ha; stands separated from one another by at least 3 km) containing sympatric populations of species pairs, and one additional site containing only T. lancifolium to supplement a lack of available mature fruits for this species at the other sites (n = 7 sites; Table 1). Five sites were located in northwest Georgia, and two sites were located near the borders of Georgia, North Carolina, and South Carolina. All sites were low-lying, mesic, deciduous forests with moderate to thick canopy cover. Sites in northwest Georgia were located in the Limestone Valley soil prov- We collected mature fruits just prior to natural dehiscence from the study species at their respective sites during summer 2018 for use in morphological and chemical analyses. Upon return to the laboratory, we stored fruits at −20°C for 1 month prior to diaspore morphological analyses and 4 months prior to chemical analyses. To organize the species by their overall attractiveness to ants, we used seed dispersal probabilities calculated for each species based on in situ observations conducted during a previous study in the same study system (Miller & Kwit, 2018). In that study, the proportion of seeds removed by ants (primarily of the genus Aphaenogaster) in 1 hr from natural seed depots was averaged for 15 individuals of each species at three study sites; the proportion across sites for each species was then averaged (n = 45 hr of observation/species). We interpret these averaged proportions as the probability of seed dispersal, a measure of attractiveness for each species of Trillium and a comparative framework for the present study.

| Diaspore morphology
To address our first study objective, quantifying interspecific differences in diaspore morphology, we measured diaspore and elaiosome length, width, and mass for 26 diaspores of each species, representing multiple individuals from each study site (n = 130). We took the fresh mass of entire diaspores (g) and then measured the length and width of diaspores (mm) using digital calipers (Mitutoyo Digimatic Caliper, 0.01 mm resolution). We then removed the elaiosome from each seed using a straight razor and repeated the above measurements for the elaiosome. Elaiosome-seed mass ratios were calculated by dividing the mass of the elaiosome by the mass of the entire diaspore.
After evaluating each response variable (diaspore length, diaspore width, diaspore mass, elaiosome length, elaiosome width, elaiosome mass, and elaiosome-seed mass ratio) for normality by generating normal probability (Q-Q) plots and histograms to visualize model residuals, we determined that the residuals of all response variables were normally distributed. Each response variable was then compared for the five study species using a linear mixed-effects model in the package lmerTest (Kuznetsova, Brockhoff, & Christensen, 2017) in R (R Core Team, 2017). This test uses the Satterthwaite's method for approximating degrees of freedom for t and F tests, which is more conservative than residual degrees of freedom approximations; therefore, the assumption of homogeneity of variances across samples can be relaxed (Keselman, Algina, Kowalchuk, & Wolfinger, 1999). Each model included one morphological trait as the response variable, species of Trillium as the fixed effect, and study site as a random effect to account for the potential effect of environmental heterogeneity TA B L E 1 Study species collected at field sites in June 2018, and their relative status as attractive or less attractive to ant dispersers F I G U R E 1 Empirically-derived averaged probability of seed dispersal by ants, primarily of the genus Aphaenogaster, for five species of Trillium at seven sites in the southeastern U.S. Bars represent standard error. Blue indicates lower seed dispersal probabilities (i.e., "least attractive" species) while pink indicates highest seed dispersal probabilities (i.e., "most attractive" species) or intraspecific population. Tukey post hoc tests were performed to identify the pairwise direction of effects using the package multcomp (Hothorn, Bretz, & Westfall, 2008). To visualize trends, which were impacted by the random effect of study site, we generated model residuals from linear models with the effect of site partialled out. Boxplots display these model residuals, which depict the isolated effect of species on diaspore morphology traits.

| Elaiosome fatty acid profiles
To address our second study objective, quantifying interspecific differences in concentrations of elaiosome fatty acids among species of Trillium, we assessed triglyceride, diglyceride, and free fatty acid forms using gas chromatography-mass spectrometry (GC-MS). Detailed methodology relating to the GC-MS analysis can be found in Data S1.
We produced standard curves by running five concentrations (10, 1, Total fatty acid concentrations and concentrations of fatty acids in free, di-, and triglyceride forms were compared for the five study species using linear mixed-effects models in the package lmerTest in R, as above. Concentrations (% fresh weight) of five individual fatty acids and of total fatty acids were evaluated for normality prior to running linear mixed-effects models; in each case, the assumption of normality was violated. Therefore, we applied data transformations to raw concentrations to approximate normality (logarithmic or square-root; see Keene, 1995;Osborne, 2002). As in the Diaspore Morphology methods above, our use of the Satterthwaite's method for approximating degrees of freedom allowed us to relax the assumption of homogeneity of variances. Study site was included as a random effect in each model to account for the effects of environmental heterogeneity or interspecific population. Species and fatty acid form (free, di-, or triglyceride) were included as fixed effects.
Interactions between species and form were tested for each model; in no case was the interaction term significant, so interactions were dropped from all models. Tukey post hoc tests were performed to identify the pairwise direction of effects between species and between fatty acid forms. To visualize trends, we generated boxplots using model residuals as in the Diaspore Morphology methods above.

| Elaiosome phytochemical profiles
To address our third study objective, assessing interspecific differences in elaiosome phytochemistry, we characterized and compared profiles of both known and unknown elaiosome metabolites using liquid chromatography-mass spectrometry (LC-MS). Each of the five study species was represented by six sample replicates (N = 30).
Samples were taken from 2 to 6 individuals from the representative study sites (Table 1). Detailed methodology relating to the LC-MS analysis can be found in Data S1. LC-MS produced relative concentrations of phytochemical compounds across samples, which were compared using z-transformed peak areas from extracted ion chromatograms.
To visualize clustering of elaiosome phytochemical profiles across species, we performed a constrained redundancy analysis (RDA) using the vegan package (Oksanen et al., 2017) in R with species of Trillium as the explanatory variable and peak areas of the partial phytochemical data (known compounds; n = 122) as the response.
RDA, a method to summarize the variation in a set of response variables that can be explained by a set of explanatory variables, is a constrained version of principal components analysis wherein canonical axes (i.e., linear combinations of response variables) must also be linear combinations of the explanatory variables (Legendre & Legendre, 1998). We did not explicitly account for variation in the phytochemistry data that may have been explained by study site because RDA partitions the total variance of the data into constrained variances (i.e., variation in the response matrix that is redundant with the variation in the explanatory matrix) and unconstrained variances (i.e., variation in the response matrix that is not redundant with the variation in the explanatory matrix), which can be compared to determine how much of the variation in the response is accounted for by the explanatory variable (Legendre & Legendre, 1998). We used all six sample replicates per species, for a total of 30 samples. The model ran for 1,000 permutations and the significance of the explanatory variables and the first two RDA axes were assessed using ANOVA. We repeated this procedure for the full phytochemical data (known compounds + unknown features; n = 7,552) detected by LC-MS.
To better understand the major chemical drivers of the multivariate spread of the partial (i.e., known compounds) and full (i.e., known + unknown compounds) phytochemical data, we constructed four random forest models using the package randomForest (Liaw & Wiener, 2002) in R. Random forest is a supervised machine learning technique that builds a classification tree by repeatedly splitting the data based on whether or not they fall above or below a threshold value of each explanatory variable in the model (Biau, 2012;Bielby, Cardillo, Cooper, & Purvis, 2010). Random forest ranks the relative importance of different predictors in distinguishing among the levels and provides a measure of prediction accuracy, cross-validation, for correctly classifying unknown samples into groups according to the predictor variables. By training the model using a subset of samples, random forest is able to classify the remaining samples with higher accuracy than would be achieved by always guessing the most common category.
In our analysis, random forest classification models (RFCMs) were implemented to identify compounds with the highest relative importance in distinguishing among the five species of Trillium, a categorical level. Two RFCMs were used evaluate the partial and full phytochemical data sets, respectively. Peak areas of the known compounds were used as the set of possible predictors for the partial RFCM (n = 122), whereas peak areas of all compounds (known + unknown) were used as the set of possible predictors for the full RFCM

| Diaspore morphology
Diaspore length, width, and mass differed significantly among the five Trillium species (F 4,15 = 4.98, p = .009; F 4,13 = 7.79, p = .002; F 4,15 = 9.03, p < .001, respectively). Elaiosome length also differed significantly among species (F 4,15 = 5.59, p = .006). Post hoc comparisons revealed that the second most attractive species, T. cuneatum, had significantly greater diaspore length than the least attractive species, T. decumbens; that the two most attractive species, T. catesbaei and T. cuneatum, had significantly greater diaspore width than the least attractive species; and that the two most attractive species had significantly greater diaspore mass than the least attractive species (Table S1; Figure 2). Post hoc comparisons revealed that the most attractive species, T. catesbaei, had significantly greater elaiosome length than all of the other species except for T. lancifolium.

| D ISCUSS I ON
In this study, we investigated the differences in diaspore morphology, elaiosome fatty acids, and elaiosome phytochemisty among five species of Trillium with different levels of attractiveness to seed-dispersing ants in the southern Appalachian region of North America.
Of the morphology metrics considered in our study, diaspore length, diaspore width, diaspore mass, and elaiosome length were significantly different among species, and post hoc tests revealed that values of these traits tended to increase with seed attractiveness.
These findings provide support for the established hypothesis that seed size is a key trait determining the probability of seed dispersal , with larger diaspores overall being preferred by seed-dispersing ants (Hughes & Westoby, 1992;Takahashi & Itino, 2015).
Contrary to our prediction, elaiosome mass, elaiosome width, and elaiosome-seed mass ratio were not significantly different among species. These results are not consistent with the results of several studies showing that ants prefer larger elaiosome biomass and/or elaiosome-seed mass ratios (Leal et al., 2014;Levine et al., 2019;Mark & Olesen, 1995), although they do support the finding that ants do not exert significant selection on elaiosome size in Helleborous foetidus . Considering that the most attractive species in our study, T. catesbaei, had smaller average elaiosome mass and smaller elaiosome-seed mass ratios than all of its congeners, these aspects of elaiosome morphology do not appear to be the most important factors contributing to the attractiveness of Trillium seeds to ants. Larger diaspores likely enhance attractiveness of seeds to ants, but ants may not always prefer seed species with larger elaiosome mass or elaiosome-seed mass ratios.
The most attractive species also had significantly higher concentrations of oleic acid than the less attractive species, providing support for the hypothesis that this compound acts as a behavior-releasing signal that stimulates ants to pick up and carry items to or from the nest (Brew et al., 1989;Marshall, Beattie, & Bollenbacher, 1979;Qiu et al., 2015;Skidmore & Heithaus, 1988).
Oleic acid is the most abundant fatty acid in plant and animal tissue and the biosynthetic precursor of linoleic and linolenic acids (Christie, 2005), essential nutrients that are not synthesized by hymenopterans (Barbehenn, Reese, & Hagens, 1999;Canavoso, Jouni, Karnas, Pennington, & Wells, 2001;Dadd, 1973;Hagen, Dadd, & Reese, 1984). As the main constituent in insect hemolymph, oleic acid in the form of diolein is of particular nutritional importance for ant larvae (Fischer et al., 2008;Municio, Odriozola, & Pérez-Albarsanz, 1975;Thompson, 1973 (1980) found that elaiosomes of Datura discolor (Solanaceae) were conspicuously absent of oleic acid, whereas palmitic acid, stearic acid, linoleic acid, and linolenic acid were all present. We found that linoleic, hexadecanoic, stearic, and palmitoleic acids were also present in higher concentrations in the more attractive species of Trillium, so these fatty acids likely contribute to the overall attractiveness of seeds to ants.
Our prediction that elaiosome phytochemical profiles of the more attractive species of Trillium would be distinct from their less attractive congeners was supported. Although the partial RDA shows that all study species were independently clustered in multi- by these endemic species (Miller & Kwit, 2018)  in seeds that had a higher probability of being dispersed by ants.
The potential roles played by the remaining compounds selected by random forest models is unknown, but could be the focus of further investigations.
The defense trade-off hypothesis posited by Cipollini and Levey (1997), and further explored by Schaefer, Schmidt, and Winkler (2003), may explain the presence of phytochemicals that appear to reduce the attractiveness of trillium seeds. Plants must balance attracting dispersers and repelling the mortality agents that exposed fruits/seeds come into contact with prior to and following dispersal. Whereas some compounds present in elaiosomes might deter ant dispersers, this cost may be off-set by better defenses against granivores or microbial pathogens. In myrmecochores, this could be of particular importance, given the likelihood that rodents will prey on seeds that are not removed by ants within a few hours of dehiscence (Heithaus, 1981). The importance of microbial pathogens has not been investigated thoroughly in myrmecochore systems, so deterrent compounds in elaiosomes might also play a role in defending the seed against fungal infections. The defense trade-off hypothesis would imply that mortality selection agents (i.e., granivores, F I G U R E 4 Constrained RDA evaluating general elaiosome phytochemical profiles (n = 6 samples per species; n = 30), with relative abundance of 122 known phytochemical compounds as the response and species as the fixed effect. Ellipses represent the standard deviation of the points around the centroid for each species. Blue = least attractive species, pink = most attractive species. The overall model is significant (p < .001), meaning species is a significant predictor of phytochemical composition of elaiosomes and several other morphologically similar species including A. picea (Umphrey, 1996), are difficult to reliably distinguish from one another and may be characterized by polyphyly (DeMarco & Cognato, 2016). Our study system encompasses the ecotone between A.
picea and A. rudis (Warren et al., 2016). Although both species are effective seed dispersers, they may be marked with differences in seed dispersal effectiveness that are not accounted for in this paper (see Warren, Bahn, & Bradford, 2011); as such, we acknowledge that some of the variability in seed dispersal probabilities across species of Trillium may be due to differences in the seed disperser assemblages at each study site. However, that pairs of the study species co-occur within multiple sites and thus were exposed to identical dispersal assemblages during the in situ observations of seed dispersal likely minimizes these differences.
Based on our results, the trillium seeds that were most attractive to ants are characterized by larger diaspores overall (but not necessarily by larger elaiosome biomass or elaiosome-seed mass ratios), and by elaiosomes with high concentrations of the α-amino acid methylhistidine and the carbohydrate d-glucarate, low concentrations of salicylate, xanthine, histidine, citrulline, and pantothenate, and high concentrations of free oleic, linoleic, hexadecenoic, stearic, palmitoleic, and total fatty acids. Many of these traits are likely correlated, and thus there is some redundancy in our multi-dimensional description of an attractive myrmecochore seed. Our results regarding diaspore and elaiosome morphology and elaiosome fatty acids compliment the work of many previous studies, but we provide novel insights into the potential roles played by previously-unknown components of the broader elaiosome phytochemical profile. Previous work in other Trillium species corroborates our conclusion that seed-dispersing ants respond to a complex of characters, including morphological and chemical seed attributes (Gunther & Lanza, 1989; Lanza et al., Least attractive M ost attractive 1992). Further advances in our understanding of seed dispersal effectiveness in myrmecochorous as well as other animal-mediated seed dispersal systems will require a portrayal of dispersal unit chemical and physical traits, and their combined responses to selection pressures.

ACK N OWLED G M ENTS
We thank E. Schilling, J. Clark, and J. Fordyce for their input and suggestions throughout the project. We thank J. Fordyce for help with statistical analyses. Thanks to the late T. Patrick, and the Georgia DNR for assistance in locating study sites. We thank K. Lawhorn for field assistance. This project was

DATA AVA I L A B I L I T Y S TAT E M E N T
Data are available from the Dryad Digital Repository: https ://doi. org/10.5061/dryad.hhmgq nkcz