The interactive figures in this article are best viewed on a computer.
The surface of a pollen grain is covered with spines, ridges, networks and other structures, known as its ornamentation. Even when microscopes were still crude, pollen ornamentation fascinated botanists.
To this day, ornamentation is described mainly by qualitative categories such as echinate (spiny), lophate (ridged), psilate (smooth) and reticulate (net-like). Yet because ornamentation varies continuously, categories can only divide it into a few discrete bins: variation within each is lost, and which category a grain falls into often depends on the judgement of the observer.
Can ornamentation be quantified with as few descriptors as possible? This article describes our attempt, in which we use persistent homology from topology to quantify ornamentation as three interpretable scalars: intensity, density and polarity.
1. Limitations of spherical harmonic analysis
To describe shape quantitatively, one naturally thinks of Fourier analysis, whose counterpart for closed surfaces in three dimensions is spherical harmonic analysis. It writes the surface as a linear combination of orthogonal basis functions on the sphere, and the higher the degree, the finer the detail it can express.
As the degree rises, the reconstruction approaches the original surface. But because spherical harmonic analysis can only describe closed surfaces, the reconstruction is always closed, however high the degree. The original surface has openings, which the reconstruction always fills in with a continuous surface, and at higher degrees, protrusions that do not exist in the original surface also appear near the openings. These are oscillations that a truncated expansion produces where the surface changes abruptly, known as the Gibbs phenomenon in mathematics. Moreover, the information about ornamentation is spread over a large number of coefficients, and individual coefficients are hard to match to recognisable morphological features. Spherical harmonic analysis is therefore ill suited to building interpretable descriptors of ornamentation.
2. From watershed algorithm to persistent homology
Pollen is so diverse in form that any preselected set of basis functions would probably struggle to describe all grains equally well, so we need a method that makes no assumptions about shape and considers only the relief of the surface. The watershed algorithm from image processing, which is often used for image segmentation (for example, to separate touching cells in a micrograph), offers such an intuition. It treats a greyscale image as terrain, with bright areas as highlands and dark areas as valleys. Imagine flooding the surface from every local minimum, with the water rising at the same rate everywhere, so that each valley first fills to form a pool of its own. Where neighbouring pools meet lies a watershed, and the watersheds divide the image into catchment basins.
For a nearly spherical object such as pollen, we can imagine a ball of water expanding from the centre of the grain. Depressions in the surface are flooded first and form pools, while protrusions stand above the water as islands until the ball grows large enough to submerge them. In this way the complex three-dimensional surface becomes a function on the sphere, which greatly simplifies the problem.
Persistent homology, a tool from topology, formalises this intuition. Denote the pollen surface by ρ, the distance from the centre to the surface in each direction, and let a spherical water surface of radius r sweep through the grain. We call the part below the water, {ρ≤r}, the depression channel, whose connected components are the pools, and the part above it, {ρ≥r}, the protrusion channel, whose connected components are the islands. As the water rises, a new pool appears whenever the level reaches the bottom of a depression; when the level reaches the saddle between two pools, they join, and the younger pool merges into the older one and vanishes. As the water falls, the peaks of protrusions emerge as islands, and when two islands meet at a saddle, the younger one vanishes. The range of water levels over which a pool or island survives, from its appearance to its disappearance, is its persistence pi, and the channel it belongs to is denoted σi, with +1 for protrusions and −1 for depressions. Each pool and each island thus corresponds to a pair (pi,σi). The first component to appear in each channel never vanishes and is left out of the analysis.
These pairs can be drawn either as a persistence barcode, in which each pool or island is a horizontal bar, or as a persistence diagram, in which each is a point whose horizontal and vertical coordinates are the water levels at which it appears and disappears.
Unlike spherical harmonic analysis, persistent homology tolerates openings in the surface and is robust to small noise on it.
3. Three interpretable descriptors
The two channels together give N pairs (pi,σi). Letting πi=pi/∑jpj be the fraction of the total persistence contributed by each pool or island, we define three interpretable scalars, intensity, density and polarity, as follows.
Intensity I, the logarithm of the mean persistence, reflects the average amplitude of the ornamentation. Density D, the Shannon entropy of the fractions, reflects how finely the pattern is divided. Polarity P reflects the difference between the two channels and thus shows whether the ornamentation is dominated by protrusions or by depressions.
To get a feel for how these three scalars reflect ornamentation, we made the schematic pollen below. Its relief height, number of ornamental elements and fraction of protrusions can each be adjusted, and each corresponds to one of our scalars. As the relief height increases, intensity I increases and the point cloud in the persistence diagram moves away from the diagonal as a whole. As the number of elements increases, density D increases, more points appear in the persistence diagram, and the point cloud becomes more diffuse. As the fraction of protrusions increases, polarity P increases, and points move from the depression channel to the protrusion channel.
On real pollen, the three scalars likewise reflect ornamentation. The surface of the daisy (no. 1) is covered with tall, sharp spines, and the points of its protrusion channel lie far from the diagonal, so its intensity is high and its polarity positive. The surface of rough bluegrass (no. 11) is nearly smooth, and its persistence diagram has few points, all close to the diagonal, so both its intensity and its density are low and its polarity is near 0. The surface of cuckooflower (no. 16) is covered in a fine network, and its depression channel has many tightly clustered points, so its density is high and its polarity negative.
4. Pollen ornamentation morphospace
Taken as axes, intensity, density and polarity span a pollen ornamentation morphospace, in which each pollen grain is a point and grains with similar ornamentation lie close together. We placed 162 nearly spherical angiosperm pollen grains from the 3D Pollen Library in this space.
The three axes are almost uncorrelated, and each captures one aspect of ornamentation. Representative grains of the same type cluster together, and different types occupy different regions of the morphospace. Intensity separates strong relief from weak: echinate and lophate pollen score highest and psilate pollen lowest. Density separates coarse patterns from fine ones: reticulate pollen scores highest, echinate and lophate pollen fall in between, and psilate pollen scores lowest. Polarity separates protrusions from depressions, being positive for echinate and lophate pollen, negative for reticulate pollen and close to zero for psilate pollen.
Does this morphospace really capture ornamentation? We tested it on two sets of pollen whose ornamentation is known: theoretical pollen whose exine pattern is generated by a physical model, with its period set by the wavenumber ℓ0 (the number of ornamental elements around the grain), and 19 real echinate pollen grains of Asteraceae, whose spines can be counted one by one.
For theoretical pollen, D and P are related to the wavenumber ℓ0 by
where h(P) is the binary entropy of the persistence fractions of the two channels1.
For real echinate pollen, the number of spines is related to the wavenumber by n=ℓ02/π, which gives
When recovered from D and P, wavenumbers of theoretical pollen have a mean error of 6.1%, and spine numbers of real pollen 11.9%.
5. Pollination syndrome and evolution
A quantitative morphospace offers a way to test qualitative theories and hypotheses from traditional biology: once morphology becomes coordinates in a space, a theory’s predictions about morphology can be written as hypotheses about coordinates and tested statistically. Here we take the pollination syndrome as an example and examine it in the pollen ornamentation morphospace. The pollination syndrome holds that a given kind of pollinator shapes a suite of floral traits that go together; flowers of wind-pollinated plants, for example, tend to be small and inconspicuous and produce no nectar. For pollen, it predicts that wind-pollinated pollen tends to be smoother and more aerodynamic, so that it can travel far on the wind, whereas insect-pollinated pollen bears more conspicuous ornamentation that helps it cling to insects. Although the framework is simple and useful, it has always been contested, and a large-scale comparative study showed that the differences in ornamentation between the two kinds of pollen only partly match this prediction. With a quantitative morphospace, we go further and measure these differences along the three axes of intensity, density and polarity.
Of the 162 grains in the morphospace, we determined the pollination mode for 116 species from the Ecological Flora of the British Isles, 95 of them insect-pollinated and 21 wind-pollinated, and then compared the ornamentation of the two kinds of pollen on a dated phylogenetic tree.
Because wind-pollinated species are concentrated in a few branches of the tree, a direct comparison that ignores relatedness could mistake similarity inherited from shared ancestry for an effect of pollination mode. We therefore compared the two kinds of pollen with phylogenetic generalized least squares (PGLS), which discounts the similarity between closely related species according to the tree. Without correction, insect-pollinated pollen scores higher than wind-pollinated pollen in intensity, density and polarity, but after phylogenetic correction only the difference in intensity remains significant, indicating lower relief in wind-pollinated pollen.
The morphospace can also be used to trace the evolution of ornamentation. We estimated the ancestral states of 28 Asteraceae species under a Brownian motion model and placed the species together with their ancestors in the morphospace. Starting from a common ancestor, the three subfamilies diverge in different directions: Cichorioideae towards finer ornamentation dominated by depressions, Asteroideae towards coarser ornamentation dominated by spines, and Carduoideae towards lower relief. Of the three axes, only density carries a significant phylogenetic signal2.
6. Discussion
The morphospace is a central concept of theoretical morphology. There are broadly two ways to construct a morphospace: the theoretical way defines the axes in advance, while the empirical way obtains them by ordinating specimens, for example with principal component analysis. The classic example of a theoretical morphospace is Raup’s shell coiling morphospace, which founded theoretical morphology. Raup built this space to capture the “spectra of gradation” between qualitative categories such as turriculate and turbiniform in shells. Our pollen ornamentation morphospace pursues a similar goal. Its axes are likewise defined in advance and contain no free parameters, so no subjective choices are needed and the computation is fully automatic. Moreover, the axes can be read directly as the intensity, density and polarity of ornamentation, whereas the axes of an empirical morphospace have only statistical meaning, and their biological meaning must be interpreted afterwards. Finally, because the three axes capture distinct aspects of ornamentation, the differences along each axis can be analysed separately. By contrast, whether the axes of classic morphospaces are independent has long been debated; even in Raup’s space, the axes are mathematically correlated.
This method can also be seen as a generalisation of a classic technique in palynology, LO analysis. Before scanning electron microscopy came to palynology, LO analysis was the main means of identifying ornamentation under the light microscope. As the focal plane is gradually lowered through the grain, protrusions turn first bright and then dark (the LO pattern), while depressions turn first dark and then bright (the OL pattern), where L and O stand for the Latin words for light (lux) and darkness (obscuritas). For nearly spherical pollen, the spherical water surface locally approximates the focal plane, so islands correspond to the LO pattern and pools to the OL pattern. The difference is that persistent homology covers every water level and does not depend on the judgement of an observer.
One limitation of our approach lies in the size and taxonomic coverage of the sample. Most palynological records are two-dimensional images from light and scanning electron microscopy. The 3D Pollen Library we used is the largest public three-dimensional pollen dataset to date, but three-dimensional data remain limited in palynology as a whole. The method itself also has limitations. First, it assumes that pollen is nearly spherical, so for markedly non-spherical pollen, the overall shape may be mistaken for ornamentation; the two poles of an ellipsoidal grain, for example, would be treated as two long spines. Second, to keep the framework simple, we used only zero-dimensional persistent homology, which does not record the shape or arrangement of ornamental elements and therefore cannot distinguish ornamentation that differs only in these respects. Incorporating higher-dimensional homology could make up for this shortcoming.
We thank Anthony J. Hayes and the Bioimaging Research Hub at Cardiff University for generously sharing the 3D Pollen Library with the scientific community, without which this research would not have been possible. We also thank Asja Radja and colleagues for making their model code publicly available. The pollen models and the images generated from them are licensed under CC BY-NC and are not released under this article’s CC BY 4.0 licence.