May 25, 2024 - Science

AI pierces the secret life of plants

Illustrated collage of an aerial image of a crop in a field juxtaposed against lines of code, with pieces of wheat emerging from the code surrounded by abstract squares.

Illustration: Aïda Amer/Axios

The latest AI tools and computing advances are providing a more detailed view of plants and their interactions with the world that could help breeders develop more resilient crops and farmers plan for a far different future.

Why it matters: A growing global population needs to be fed using less land — all under the pressures of degrading soil, pests, disease and climate change.

  • "These are huge challenges, and we're trying to address them on many different scales — from the epidemiological to the molecular," says Jake Harris, a professor of plant biology at the University of Cambridge.
  • "AI is going to help across the board," he says.

State of play: AI algorithms aren't new to plant science. Robots roam fields taking photos of plants and, using AI deep learning methods, detect disease and analyze different attributes of plants in an effort to bring precise, consistent data collection to agriculture.

  • Machine learning algorithms analyze plant traits and predict which genetic combinations will produce a plant with the desired traits — often aiming for a resilience to drought or disease that doesn't slash a crop's yield.

But new AI-based tools are allowing researchers to unravel the inner workings of plant biology that were previously hidden in a complex web of molecular interactions.

  • "It wasn't economically feasible to do plant structural biology at scale," says Google DeepMind researcher John Jumper, who leads the team behind AI-powered protein structure predictor AlphaFold.
  • One paper found the structures of less than 2% of the proteins in plant biology's model species, Arabidopsis thaliana, were known, compared with about 10 times as many for human proteins. AlphaFold bumped that coverage of the plant's proteins to more than 60%, though with varying degrees of quality.
  • Harris is using AlphaFold to try to understand the chemical modifications that are made to plant DNA when they are exposed to pathogens, drought and other stresses. Those modifications store information for the plant to respond the next time it is stressed — but how the cell does that was "previously invisible" information, Harris says.

Zoom out: Other recent AI advances are allowing plant scientists to look beyond a plant's genes and proteins and consider other key factors — such as the soil, climate, and farm management practices — that are involved in the production of a plant.

  • In a recent preprint paper, researchers at the University of Kentucky reported using machine learning algorithms to predict the genetic makeup of a grapevine from the microbiome of the soil. They could, for instance, tell whether a vine was shiraz or cabernet sauvignon.
  • That indicates a plant's genetics have an effect on the microbiome in the soil and that crops could be bred to be better hosts for beneficial microbes that could reduce the amount of water or chemicals used to grow a crop, says study co-author Carlos Rodriguez Lopez, a professor of horticulture at the University of Kentucky.

Generative AI methods and advances in the specialized processors AI needs are driving an effort at Michigan State University, led by agricultural engineering professor Daniel Uyeh, to build a digital twin of an apple tree.

  • The team is planning to use LiDAR-equipped iPhones, cameras and other sensors to build a collection of images of trees growing under a slew of different conditions, including temperature and relative humidity, to "mimic everything about a tree," Uyeh says.
  • Ultimately the idea is that a farmer can capture data about their own orchard and, using the digitally reconstructed tree, construct future climate and field conditions for their crops that they can plan for. Data about those future conditions would be produced using generative AI tools.
  • Generations of farming knowledge are being challenged by rapidly evolving climate change, Uyeh says.
  • The researchers' big goal is to model an entire orchard, while developing tools along the way to help farmers use the wealth of information about plants and crops that's already available.

Yes, but: AI's integration into plant biology faces hurdles, particularly around the data that is available and its quality.

  • Unlike the massive corpus of digitized text used to train chatbots, or even the rafts of biological data generated from genome sequencing, there aren't digitized images of different physical aspects of apple trees.
  • It's also difficult to find scientists who understand both biology and computer science — and to attract them to plant science, which often doesn't offer salaries and funding at the level of medical research, Rodriguez Lopez says.

What to watch: Large language models (LLMs) that power chatbots and other AI tools that people increasingly interact with are being developed to decipher the language of DNA and proteins.

  • That could allow scientists to identify how different regions of a genome interact with one another and control different traits of a plant — the way different sentences in a chapter string together a story, says Mohsen Yoosefzadeh Najafabadi, a research associate at the University of Guelph.
  • It would "save so much time and resources because there is no need to grow plants in the field and select them. We can take DNA from a seed and see if a sentence is available and select the best seed."
Go deeper