CHAMPAIGN, Ill. — Scientists used machine-learning and super-resolution microscopy to overcome a challenge that has stymied research into the natural history of grasses for decades. Their method allowed them to detect subtle differences among grass pollen grains and trace changes in grass diversity and the proportions of the two main photosynthetic types of grasses at one site over a period of 25,000 years.
"Open grasslands are a relatively recent ecosystem in Earth history, with open-habitat grasses present in the Eocene, about 40 million years ago," the researchers wrote in a report in the Proceedings of the National Academy of Sciences. "Grasses were potentially the first plants domesticated about 12,000 years ago and today include several of the world's most important staple foods, such as wheat, rice, maize, barley, sorghum and millet."
But scientists face a massive challenge when trying to classify pollen fossils: grass pollens tend to all look alike, said University of Illinois Urbana-Champaign plant biology professor Surangi Punyasena , who led the new research with former Ph.D. student Marc-Élie Adaimé , now a postdoctoral researcher at the Smithsonian's Office of Digital and Innovation.
Unlike pollen from other flowering plants, which can be distinguished by their shapes, spikes, grooves or pore arrangements, pollen grains from different grass species look remarkably similar under a standard light microscope.
"As paleobotanists and paleontologists, we're restricted to working with the morphology of pollen grains, which are one of the main parts of the plant that can be fossilized," Punyasena said.
The field is not lacking data; the problem lies in finding objective measures to classify it.
"Within a small cubic centimeter of sediment, you could have thousands, potentially millions of pollen fossils," she said. "But the level at which we were able to analyze it before machine learning was limited by human ability."
Light microscopy could not distinguish characteristic features of a pollen grain's surface. Electron microscopy could detect more features, but in an expensive, labor-intensive manner.
This limited scientists' ability to explore and understand the evolution and distribution of grasses, Punyasena said.
In earlier studies, she and her colleagues made advances in using super-resolution microscopy to reveal some of the hidden features of grass pollens.
"Super-resolution microscopy works sort of like a confocal microscope, where you use a laser to illuminate one point at a time," Punyasena said. "But algorithmically, it's capturing all the scattered light and calculates it back to the point of origin. You get close to electron microscopy quality, but the process is much faster, much easier."
When analyzing the images, Adaimé "recognized that there were small differences in both the patterning and the complexity of the patterning on the surface of the pollen grains, and also the cell wall thickness," Punyasena said.
Using images from several identifiable grass species, Adaimé trained a machine-learning model to recognize these differences. He then developed a statistical method that uses the patterns the model learned to estimate species diversity in samples containing pollen from multiple species. When tested on samples with known species compositions, the method produced estimates that closely tracked their actual diversity.
"That's essentially the definition of machine learning: for a computer model to be able to learn patterns and apply that learning to new data without being given explicit rules," Adaimé said.
The model used "convolutional neural networks," which are loosely inspired by how networks of neurons in the human brain process visual information, Adaimé said. "The artificial neurons in these networks are obviously much simpler than actual nerve cells."
The model also distinguished between C3 and C4 grasses, which differ in how they concentrate carbon dioxide in their tissues to perform photosynthesis. (C4 grasses use CO2 more efficiently, enhancing their photosynthetic capacity and water-use efficiency.) Previous research found that C4 species allocate more biomass to roots and produce less dense leaves than C3 species do. Adaimé and Punyasena hypothesize that C4 grasses may also invest fewer resources in individual pollen grains than C3 grasses do, potentially contributing to thinner pollen walls and less complex surface patterns. Some evidence supports this idea, but more work must be done to confirm it, Adaimé said.
While the model cannot identify individual grass species in a sample, it can accurately estimate the diversity of species.
The researchers next used their model to analyze a core sample representing 25,000 years of pollen deposited in sediment from a lake bottom on Mt. Kenya.
"We saw how pollen diversity was substantially lower during the last ice age, especially between about 21,000 and 18,000 years ago, around the end of its coldest and harshest period, which experts call the 'Last Glacial Maximum,'" Adaimé said. "That period is particularly fascinating for biologists and Earth scientists because the CO2 levels in the atmosphere were extremely low."
Pollen diversity increased after that, coinciding with increasing CO2 levels and temperatures, he said.
The proportion of C3 to C4 grasses had no obvious association with atmospheric carbon dioxide or temperature over time, the researchers found. The proportion of C4 grasses was higher "during the final stretch of the ice age and then very gradually decreased," he said. "So, the fraction of C4 grasses decreased, and C3 grasses took over as the climate got warmer and CO2 got higher."
The study demonstrates that this approach can detect changes in species diversity and distinguish C3 and C4 grasses in assemblages of ancient fossil pollen, allowing scientists to convert these fossils into useful data more quickly and reliably, Punyasena said. Further studies aim to improve on the method and eventually expand this research to the analysis of pollen and spores from all land plants.
"It is satisfying to see that there is so much more information to be unlocked," Adaimé said. "Now we have a way to begin unraveling the history of grasslands, and perhaps even the deep evolutionary history of grasses, from clues hidden in subtle differences among their pollen grains."
Punyasena is an affiliate of the Carl R. Woese Institute for Genomic Biology at the U. of I.
