In their current study, the research team led by Susanne Dunker developed a novel automation method for pollen analysis. To do this, they combined the high-throughput of image-based flow cytometry, a method of particle analysis, with a form of artificial intelligence (AI) known as deep learning - and thus designed a highly efficient analysis tool that, in addition to precise species identification, also makes it possible to quantify the pollen grains contained in a sample. Image-based flow cytometry is a method mainly used in medicine for the analysis of blood cells, which is now also used for pollen analysis. "A pollen sample to be examined is first placed in a carrier fluid, which then flows through a narrowing channel," Susanne Dunker explains this method. "Due to the constriction, the pollen grains are separated and lined up like on a string of pearls. In this way, each pollen grain travels individually through the built-in microscope element - that can be up to 2,000 pollen per second." Two normal microscopic images are supplemented with ten fluorescence microscopic images per pollen. In this process, after being excited by lasers with light of specific wavelengths, the pollen itself emits light. "In which range of the colour spectrum and where exactly the pollen fluoresces is sometimes very specific. This gives us further characteristics that can help identify the plant species in question," says Susanne Dunker. In deep learning, an algorithm abstracts the original pixels of an image more and more in successive steps to ultimately extract the features specific to a species. "The combined use of microscopic images, fluorescence features and high-throughput has never been seen before in pollen analysis - this is actually an absolute novelty." For example, the analysis of a low-complexity sample takes four hours on the microscope; with the new method, it takes only 20 minutes. The novel high-throughput analysis method is therefore registered as a UFZ patent, and Susanne Dunker received the UFZ Technology Transfer Award for it in 2019. The pollen samples examined in the New Phytologist study came from 35 meadow plant species, including, for example, yarrow, sage, thyme and various clover species such as white, mountain and meadow clover. The researchers took a total of around 430,000 pictures, which formed the basis for a database. In cooperation with the Technical University of Ilmenau, this was converted into a highly efficient tool for pollen identification using deep learning. In subsequent investigations, the researchers tested the accuracy of their new method: they compared unknown pollen samples from the 35 plant species with the database. "The result was more than satisfactory, the accuracy was 96 percent," says Susanne Dunker. Even species that are difficult to distinguish even for experts using a microscope could be reliably identified. The new method is therefore not only extremely fast, but also highly precise. In the future, the new method of automated pollen analysis will play a central role in important research questions concerning plant-pollinator interactions. How important are certain pollinators such as bees, flies or bumblebees for a plant species? What would be the consequences of losing a pollinating insect species or a plant? "We are now able to qualitatively and simultaneously quantitatively evaluate pollen samples on a large scale. We are constantly expanding our pollen database of insect-pollinated plants for this purpose," says Susanne Dunker. She wants to expand the database to at least those 500 plant species whose pollen is relevant for the honey bee as food.
