Novel computed tomography-based tools reliably quantify plant reproductive investment

New computed tomography-based tools allow plant reproductive investment to be quantified, and their use is demonstrated by studying the differences in pollen/ovule ratios across inflorescences of deceptive versus rewarding orchid species.


Introduction
It should be evident to most human beings that the mere existence of genders is a harbinger of conflicts for resources. Gender conflicts are nowhere more acute than in hermaphroditic organisms where both genders have to draw from the resource pool of the same organism to maximize fitness (Charnov, 1979;Lloyd, 1979). In the overwhelmingly hermaphroditic flowering plants, counting the pollen grains and ovules of flowers allows us to understand how much a plant invests in the male vs. female part of its fitness.
Pollen counting methods fall into three groups (Costa and Yang, 2009): counting with the naked eye, particle counters, and image processing algorithms. Counting visually usually involves spreading samples on specialized slides with a grid, and counting a sub-sample (e.g. (Jorgensen, 1967) (Kearns and Inouye, 1993)), which is then extrapolated (Kannely, 2005). Pollen grains tend to settle disproportionately on grids, especially if the grains are large or still aggregated, which may then produce incorrect estimates (Kannely, 2005).
Electronic or laser-based counters physically detect pollen grains in order to count them.
A particle counter counts every particle; unfortunately, this may include debris and aggregated pollen (Kearns and Inouye, 1993). Image processing automates pollen counting from pollen grain images. This requires software to scan images for objects and then count each object as a unit (e.g. (Bechar et al., 1997), (Aronne et al., 2001)). All three approaches require sample destruction, including proper de-aggregation of the grains, which can be hard to achieve.
Ovules are counted either before fertilization (as ovules) or after fertilization and maturation (as seeds). At the ovule stage, counting is manual after dissection, either directly or on photographs (Nazarov, 1989), either on the whole ovary or an a stretch of the latter (ovules per mm are then extrapolated to the length of the whole gynoecium). At the seed stage, methods fall into two groups: (1) extrapolations based on counting an aliquot of known weight (Salisbury, 1942), or on a portion of a line of dry seeds (Darwin, 1862), or on the surface of a liquid (Burgeff, 1936), or within a suspension of seeds (Proctor and Harder, 1994;Sonkoly et al., 2016). (2) use of particle counting devices (DuBois, 2000). All these approaches require sample destruction and rely on the assumption that all ovules have been fertilized, which is hard to test.
In summary, these traditional methods for counting pollen and ovules show limitations with very large numbers of pollen grains (that are tightly aggregated) and ovules (especially when fertilization rates are unknown). Given these limitations, novel and more reliable methods to count pollen and ovules are needed. Contrasting agents for X-Ray Computed Tomography (CT), especially phosphotungstic acid (PTA), semiselectively accumulate in both pollen and ovules, possibly due to the higher protein content of their cells/tissues relative to their surroundings (Bellaire et al., 2012;Hayat, 2000;Staedler et al., 2013). By selecting only the brightest voxels (3D pixels) of the 3D model, i.e., the areas that absorb the more X-Rays (X-Ray data is traditionally displayed in negative), it is possible to segregate pollen and ovules from their surroundings tissues.
This process is called greyscale thresholding. Obtaining the volume of such a selection by counting voxels is straightforward. Provided that the average volume of a grain or ovule is available, CT can thus be used to count ovules and pollen in case where classical methods are at their limits.
These limits are nowhere more evident than in the study of orchids, which both possess enormous numbers of ovules and -tightly aggregated-pollen grains (Darwin, 1862).
Orchids are remarkable plants in many ways: not only are they the largest family of flowering plants (with up to 30 000 species), but also almost a third of them offer no reward to their pollinators (Ackerman, 1986;Porsch, 1909;Van der Pijl and Dodson, 1966). The presence of a reward, or lack thereof, influences dramatically pollinator behaviour: in rewarding taxa, the pollinators tend to visit several or all open flowers of an inflorescence during a visit, and visit the same inflorescence repeatedly in order to harvest its rewards ((Van der Cingel, 1995); Fig. 1A, B). In taxa with deceptive flowers, however, pollinators tend to learn to avoid deception and to visit only the first open flowers they encounter while they are still naive ((Jersáková and Kindlmann, 1998;Van der Cingel, 1995); Fig. 1D). Consequently, in rewarding plants, the fruit set is usually higher and fruits are spread across the inflorescence ((Neiland and Wilcock, 1995); Fig.   1C), whereas in deceptive plants, only the first flowers to open tend to bear fruits ((Jersáková and Kindlmann, 1998;Nilsson, 1980;Vogel, 1993); Fig. 1E). In orchids in general, the ratio of pollen to ovules (P:O) increases from the bottom to the top of inflorescences (Kopylov-Gus' kov Yu et al., 2006;Nazarov and Gerlach, 1997;Salisbury, 1942). Due to decreased pollinator visits to top flowers, we hypothesise that the decrease in ovule number (increase in P:O) should be stronger across inflorescences of deceptive flowers than across inflorescences of rewarding flowers ( Fig. 1F-I). The Orchidinae, to which most European orchids belong, are a good system to test this hypothesis because their pollination biology and phylogenetic relationships are very well understood (Claessens and Kleynen, 2011;Inda et al., 2012;Van der Cingel, 1995). Known phylogenetic relationship between studied species is needed in order to control for potential phylogenetic constraints.
This study aims at (1) establishing new methods for pollen and ovule counting that can be used even for flowers with many, densely aggregated pollen grains and ovules, and (2) demonstrating the potential of these methods by focussing on species of European orchids to uncover if the differences of pollinator behaviour in rewarding versus deceptive plants lead to different patterns of reproductive investment at the level of the inflorescence.

Material
We sampled three rewarding and five deceptive species of the subtribe Orchidinae ((Inda et al., 2012); see table S1). We collected three flowers for two to four inflorescences per species, for a total of 76 flowers (table S2).

Collection method
Open flowers (including pollinia, i.e. the pollen aggregates of orchids) or buds close to anthesis were collected from the bottom, middle and top sections of inflorescences and were immediately fixed in 1% phosphotungstic acid in formalin-acetic acid-alcohol (1% PTA / FAA; Fig. 2A). The flowers were removed from the plants with razor blades and tweezers.

Sample preparation
The sampled flowers and buds were de-aerated for 20-30 minutes with a water-jet vacuum pump. Total infiltration time in 1% PTA / FAA was four to six weeks during which the solution was changed twice. The long infiltration time allowed saturation of the sample with contrasting agent. In order to further optimize for space constraints during mounting for CT-scanning and to prevent the formation of air bubbles, the petals and sepals were removed with tweezers and dissecting scissors (Fig. 2B). The mounting was then performed in a 200µl pipette tip (Standard UNIVERSAL (Art. No.: B002.1), Carl Roth GmbH+Co KG) as described in (Staedler et al., 2013) with the following differences: (1) before mounting, the bottom end of the pipette tip was cut off. This allowed us to enter the tip with a preparation needle and to drag down the samples to optimize the used space without displacing or breaking pollinia. (2) the samples were not washed with 70% EtOH before scanning, but immediately mounted in PTA / FAA ( Although the tension produced by the shape of the tubes stabilized the samples, they were additionally stabilized by gluing them to the straw with epoxy glue (UHU Plus, UHU GmbH & Co. KG). The batch was then placed in an in-house batch holder and fixed with epoxy glue, itself fixed to the sample table (see Fig. 2E). Epoxy glue was then added between all the abovementioned parts to stabilize them. In this way, we were able to programme and scan batches of up to five flowers sequentially.

Scanning
Scanning was performed with a MicroXCT-200 system (Zeiss Microscopy). Scanning conditions are summarised in table S2. 3D reconstructions were performed via the software XMReconstructor 8.1.6599 (Zeiss Microscopy). In order to have repeatable greyscale values from one scan to another, byte scaling and CT scaling were used. Byte scaling is a procedure in which minimum and maximum greyscale values are set for the whole reconstructed scan volume (Xradia, 2010). Byte scaling was used for the reconstruction of scans of gynoecia. CT scaling is a procedure by which greyscale values are scaled to the values of two reference materials (Candell, 2009); in our studies, air and a solution of 1% PTA in FAA were used. CT scaling requires scanning of a dummy (or phantom) filled with the reference material and the presence of air on both sides of the sample during the whole scan (Fig. 2F). For this reason a pipette tip was filled with FAA+PTA, sealed on the bottom end with paraffin wax, shortened at the top and sealed with parafilm (Fig. 2F). For one dummy scan to be used on sample scans, the following parameters have to remain constant: number of projection images, voltage, objective type, beam hardening coefficient, and source filter ((Candell, 2009) ; Fig. 2G) . In practice, for all the scans in which we used CT scaling, projection images, voltage, objective type, beam hardening coefficient, and source filter were constant, whereas only voltage varied (Table S2). We thus carried out a dummy scan for each voltage value.
Calibration was performed as described in the manual (Candell, 2009). The density of FAA+PTA was measured seven times by using 4x a 100ml and 3x a 50ml measuring cylinder and an arithmetic mean of 93.7g/100ml was determined and used for calibration.
Pollinia were reconstructed with CT scaling.

Data processing
Two different procedures were developed to count pollen (summarised in Fig. 3A-D) and ovules (summarised in Fig. 3E-I). The pollen count procedure involves a high and a low resolution scan. In the low resolution scan, the whole pollinium is visible so that whole pollen volume is accessible. However in the low resolution scan, individual pollen grains are not resolved due to the strong aggregation of the grains. A high resolution scan ( Object Counter" (Bolte and Cordelieres, 2006). Subsequently, the pollen volume of the subset of the high resolution scan was measured in the low resolution scan (via the cropping function; Fig. 3C). Finally, we calculated the number of pollen grains per pollen volume for the low resolution scan, and extrapolated that value to the whole flower ( Fig.   3D), in order to obtain the total pollen grain number.
The ovule count procedure only involved one scan because the ovules can be distinguished on scans of the whole gynoecium. The ovules were segmented away from the rest of the ovary (Fig. 3E-G). On a subset of the ovules (Fig. 3H), the ovules were counted manually by using the landmark function (Fig. 3I). Average ovule volume was thereby calculated. The total ovule volume was then divided by the average ovule volume in order to obtain total ovule number per flower.

Statistical tests
We investigated differences in the number of ovules, the number of pollen grains, and the value of the pollen to ovule ratio (P:O), between rewarding and deceptive species, among species, and within inflorescences. We therefore tested the effects of the factors "interaction" (deceptive and rewarding), "species" and "flower position" (bottom, middle, and top) on the number of ovules, on the number of pollen grains, and on the P:O per flower, respectively. Analyses were performed with the software R (R_Core_Team, 2014). Because our data is nonparametric (see Supplementary data S1), we used a nonparametric analysis of variance (npANOVA), with the function Adonis (vegan) (Oksanen et al., 2013). We first generated a distance matrix with the function dist(stats) using the Euclidean distance index, and then performed the npANOVA using 9999 permutations (Anderson, 2001). Post hoc tests were performed with the same function, with a Bonferroni correction for multiple comparisons. Finally, we used the same test to look for an effect of the interaction between the factors "species" and "flower position". If the effect between the factor "species" and "flower position" is significant, it means that the number of pollen or ovules does not vary in the same direction within the inflorescence for each species. If this effect is not significant, it means that the number of ovules or pollen grains varies in the same direction within the inflorescences for all species. The variance of the P:O per species was compared between rewarding and deceptive species with a non-parametric Mann-Whitney-Wilcoxon (MWW) test using the function wilcox.test(stats).

Phylogenetic analyses
We used a pruned version of the time-calibrated phylogeny of Inda et al. (Inda et al., 2012) for phylogenetic comparative analyses (Fig. 5E). We did not include Dactylorhiza majalis in our phylogeny, because of its allotetraploid origin (Hedren et al., 2001

Development of a new method
We present a set of two approaches to count numerous high contrast objects in plant tissues via X-Ray CT. One method is presented for cases where it is possible to resolve all the objects to be counted on a scan of the whole tissue. The other method is applicable when this is not the case. Ovules can be resolved on scans of the whole ovary; therefore, an estimation of the average volume of an ovule and an ovule count can be carried out on the same data ( Fig. 3E-I). At least in the special of case of Orchidaceae, where pollen grains are relatively small and aggregated in compact pollinia, individual pollen grains cannot be resolved on the same scan data as overview scans of the entire pollinium. Two scans are therefore necessary (see Fig. 3A-D). A high resolution scan of a subset of the tissue has to be performed in order to estimate the number of pollen grains inside this subset (Fig. 3B, C); this number will then be used to estimate total object number on the overview scan (Fig. 3D). On the high resolution scan, image processing and automatic counting methods are presented that allow the straightforward processing of scan data (Fig. 4).

Phylogenetic analyses
Individually, the traits used in the analyses exhibit a λ nearly always zero, indicating no phylogenetic signal except for the "mean pollen number" (λ=1; see Table S4). In contrast to the phylogenetic signal estimates for the individual variables, the estimated maximum likelihood values of λ for two of the regression models ("mean ovule number" ~ "pollination strategy" (deceptive or rewarding) and "P:O variance" ~ "pollination strategy") were zero, indicating no phylogenetic signal in the residual errors of the models, and hence results that are equivalent to conventional ordinary least squares analyses. However, two other regression models ("mean pollen grain number" ~ "pollination strategy" and "P:O" ~ "pollination strategy") were under a Brownian motion model of evolution with an estimated maximum likelihood value of λ = 1. The "pollination strategy" was only strongly associated with the "P:O variance" (table S5).

Development of a new method
Traditionally, pollen and ovule counts have relied on destructive sampling and sub- (2) extrapolating values from massulae assumes that, within the same species, massulae numbers and pollen grain numbers are always tightly correlated, which we show not to be correct (see Fig. 5B). Counting methods relying on the whole volume of pollen are therefore more precise than methods relying on counts of massulae. Our methods also do not require de-aggregation of the pollen grains. Pollen aggregation evolved at least 39 times in angiosperms, including in some of their most species-rich lineages (e.g. the legumes and the orchids), probably because it promotes male fitness in response to infrequent pollinator visits (Harder and Johnson, 2008). The counting method we present allows to bypass the requirement of de-aggregation of the pollen grains, and can provide accurate counts for any type of pollen.
For the species we studied, the seed counts from the literature were mostly obtained from flowers that were naturally pollinated (see table 1 and citations therein). We therefore assume that, in species with deceptive flowers, only 10% of all fruits came from flowers from the top third of the inflorescence. In species with rewarding flowers, we assume that the seeds were counted from all portions of the inflorescences equally. With these assumption our counts are on average 4,5% higher than the values from literature (see table 1 and citations therein), although variation is very large and makes comparisons difficult. Our higher counts are probably due to our method counting all ovules, not only those that were fertilized. We could not test the latter possibly due to the specificities of orchid seed development (see Supplementary data S2). In the gynoecia of flowers at the top of inflorescences, extensive gaps in the ovule distribution were often noticed ( Fig. S1 A-E). These gaps could affect the chemotactic signalling from ovules to guide pollen tubes (Okuda et al., 2009;Takeuchi and Higashiyama, 2016), which could lead to some ovules not being fertilized, even if the pollen loads would be sufficient to do so.
And finally, unlike all the previously existing methods that require the destruction of the sample, our processing methods allow to measure both pollen and ovule numbers from the same flower with minimal destruction (perianth removal). The column and ovary are left intact after counting, and could be used for other analyses: shape analysis, histology, etc. The methods we present are flexible and could be used to count any high contrast objects inside plant tissues, such as crystals (druses or raphides) or stone cells (sclereids).

Reproductive investment and presence/absence of floral reward
There is no significant difference in pollen grain number per flower between rewarding and deceptive species, probably because the two rewarding species D. viridis and D.
fuchsii produced significantly less pollen than all the other species, whereas the third rewarding species, P. bifolia, was one of the species producing the most pollen grains per flower (Fig. 5A). There are also no significant trends of increase or decrease across inflorescences. Moreover, pollen grain number appears to follow a Brownian Motion model of evolution. Taken together these data suggest that pollen grain number is not under strong selection.
Ovule numbers strongly differ between species with rewarding and deceptive flowers.
Deceptive flowers contain more ovules than rewarding flowers ((Sonkoly et al., 2016); this study), which is possibly an adaptation that enables deceptive inflorescences to have the same seed set as rewarding inflorescences, despite lower fertilization rates (Sonkoly et al., 2016). Ovule numbers do not show any phylogenetic signal, possibly due to high homoplasy in this character state, i.e. rapid evolution in adaptation to different reproductive ecologies. In both rewarding and deceptive species, ovule numbers decreased from bottom to the top of the inflorescences.
Due to lack of difference in pollen grain numbers and strong difference in ovule numbers across pollination strategy and positions in inflorescence, differences in P:Os are driven by ovule numbers. Increase in P:Os from bottom to top of inflorescences have been observed in orchids (Kopylov-Gus' kov Yu et al., 2006;Nazarov and Gerlach, 1997;Salisbury, 1942), and other taxa, see e.g. (Thomson, 1989). Significant differences in variance of P:O between deceptive and rewarding flower and the well-supported strong phylogenetic correlation of these two characters (P:O variance and pollination strategy) highlight the differences in constraints acting on the reproductive investment in deceptive versus rewarding species. Given that deceptive flowers contain significantly more ovules than rewarding flowers ((Sonkoly et al., 2016); this study), and given that usually only the lower flowers on the inflorescences of deceptive flowers are fertilized (Jersáková and Kindlmann, 1998;Nilsson, 1980;Vogel, 1993), it seemed likely that deceptive inflorescences are under strong constraint to stringently decrease ovule numbers from bottom to top of the inflorescence in order to efficiently allocate resources. In rewarding flowers there are much fewer ovules, and fertilization occurs across the whole inflorescence. It seems thus likely that constraints acting on reproductive allocation in rewarding inflorescences is much weaker, which could explain the much larger variance in P:Os found in rewarding inflorescences.
Ultimately, the new tools we present will allow much broader comparative studies of the diverse groups of angiosperms in which pollination by deceit takes place, and will allow us to better quantify plant reproductive investment.

Supplementary data
Supplementary data S1 Choice of ANOVA test Supplementary data S2 Loss of contrast of developing seeds in selected European orchid species and comparison with Arabidopsis thaliana Table S1 Collected species, locality, collection date, breeding strategy, pollinators, and permits.    Movie S1 3D models of pollinium subset high resolution scan data after each steps of image processing, from raw reconstructed data to machine countable objects. Tables   Table 1 Mean pollen and ovule             completely missing, and one is present only for less than 25% of the length of the ovary