Fungi are the hidden architects of our ecosystems, acting as everything from helpful partners for plants to aggressive decomposers that recycle dead wood. However, many fungi don't stick to just one job; they can switch lifestyles depending on their environment.
Understanding this flexibility is vital for predicting how forests and farms will react to climate change. Unfortunately, the information researchers need is buried in decades of scientific papers that would take too long to comb through manually.
A new study led by NAU doctoral student Beatrice M. Bock demonstrates how AI can solve this problem. By using a specialized language model called BioBERT , Bock developed an automated workflow that assesses scientific abstracts and accurately identifies whether a fungus has a single lifestyle or a dual, flexible one.
The study was recently published in Research Ideas and Outcomes.
A high-accuracy hack
Bock said that for years, mycologists have relied on manual databases to track what different fungi do in the environment. While these tools are essential, they are difficult to keep updated as new research is published every day.
"Manually identifying fungal versatility from the literature is time-consuming and difficult to scale," Bock said. "By using machine learning, we can now scan thousands of papers in just a few minutes to flag species that might be switching roles—such as a fungus that normally helps a plant grow but also turns into a decomposer when the plant dies."
The pilot study tested four different AI models to see which was best at understanding the nuances of biological language. The top-performing model, BioBERT, achieved nearly 90% accuracy in identifying fungal lifestyles.
What did BioBERT have that the other models didn't? For one, it had the power of capitalization. Bock found that "cased" models—those that recognize capital letters—performed significantly better than those that did not. That's likely because capital letters often signal species' scientific names, like Fusarium, which are crucial for AI to understand the context of the research.
The path ahead
Bock said that in a commitment to transparency, she has made all the code and data available for free online , allowing other scientists to build upon her work and track traits in other organisms, like insects or plants.
While Bock's study focused on a small group of papers as a proof-of-concept, it opens the door for much larger projects. Future versions of the tool could predict how a fungus's behavior might change under specific environmental conditions, such as drought or extreme heat.
"As fungal trait databases continue to grow in importance for biodiversity assessments, automated text mining offers a path toward more efficient, consistent and comprehensive trait annotation," Bock said.
