From mining germplasm and predicting crosses to finding genes for editing, AI is moving deeper into the breeding process. Researchers say the biggest gains may depend as much on better data as better models.
AI platforms around the world are already changing how crop breeders analyze data, evaluate germplasm and make breeding decisions.
Decades ago, modern breeding relied heavily on large plots and wide-scale phenotype selection. Over the past 20 years, tools such as genome mapping, marker-assisted selection and gene editing have expanded what breeders can identify and accomplish. AI is adding another layer, allowing breeders to analyze much larger and more complex datasets and, increasingly, predict outcomes before making a cross or planting a field trial.
AI is already being used to better understand trait physiology. In 2019, a University of California-Davis team published results from using AI to predict metabolic pathways in tomatoes. But understanding trait biochemistry is only one area where AI is affecting breeding.
AI platforms can now speed genetic marker scoring, help select genes for editing, model breeding decisions and predict cross outcomes without field trials. U.S. firms Heritable Agriculture and Avalo, for example, both offer AI modeling. These and other leading commercial platforms will be examined in Part 4 of this series.
Researchers around the world are also beginning to describe AI as a new stage in crop breeding. Geneticists at the Chinese Academy of Sciences recently outlined what they call Breeding 5.0, a stage in which AI can digitally mine and evaluate germplasm and transform “static gene banks into dynamic, deeply-analyzable intelligent data sources.”
They describe AI as coming to “deeply ‘understand germplasm,’ not merely by identifying genetic markers but also by decoding its architecture, plasticity, regulatory logic and environmental interactions.” This “germplasm intelligence,” they argue, could enable predictive trait modeling, optimized parental design and more targeted selection.
AI may also help breeding programs make better use of information they already possess.
“AI can help us leverage existing data, clean it up and extract useful information,” explains Heritable Agriculture CEO Brad Zamft. “It’s safe to say most breeding companies have decades and decades of siloed data, from R&D notebooks to breeding data and agronomy notes all the way to sales data. Yes, you need a special team that understands how to navigate data in these complicated environments, to access and organize and convert the data into a form that AI can use, but there would be a lot of value in having all a breeding company’s data analyzed by AI for a variety of purposes.”
AI could also have a large impact on automating the processing of breeding data, says Jacob Washburn, supervisory research plant geneticist for the Midwest Area at USDA’s Agricultural Research Service.
“There are still a number of human bottlenecks that will likely be increasingly automated, some of that likely using AI,” he says. “Even some apparently simple steps like accurately identifying plots in a field and tracking them over successive dates still require significant human intervention, at least in the public sector.”
Where Will AI Make the Biggest Difference?
The larger question is what those capabilities will ultimately mean for crop improvement. Will AI primarily help breeders make incremental gains in areas such as photosynthetic efficiency, or could it contribute to larger advances, including perennial versions of annual crops or corn and wheat capable of fixing nitrogen?
Mariano Alvarez, co-founder and chief science officer at Avalo, sees potential for both.
“AI’s ability to map the broader context of gene interactions across individuals and populations is where the real magic happens,” he says.
On the subject of “huge revolutionary achievements,” Alvarez points out that many of these traits already exist in the wild.
“They are not that ‘revolutionary’ for nature,” he explains, “but we have not yet been able to harness that power for agriculture. Essentially, nature can do almost anything, and with AI platforms…we can start to harness the full potential of natural innovation.”
Washburn also sees potential in both areas, although he leans toward incremental gains.
“There may be some surprises with large breakthroughs that have evaded breeders for a long time,” he says, “but based on the information we currently have, it’s reasonable to anticipate that the greatest gains in breeding will continue to be through small improvements that together add up to significant impacts.”
In Zamft’s view, achieving larger advances such as nitrogen fixation in staple crops will require models that work at different scales. Some need to “understand” the genomic and environmental context of field trials around the globe, while others allow researchers to target specific genes or base pairs in a given variety.
“This allows you to optimize very complex traits like yield that are controlled by thousands of genes as well as identify causative genes for ‘simple’ traits,” he says. “These are traits controlled by a few or a few dozen genes and if you try to take the whole genome approach, you’ll get a lot of unintended consequences that take decades to fix. Corn ear number, for example, can be optimized by focusing on just a few genes.”
Better AI Still Depends on Better Data
The usefulness of AI in breeding depends heavily on the quantity, quality and diversity of the data behind it.
Alvarez and others expect future progress to rely on combining drone-derived phenotype data with genomic, transcriptomic and metabolomic data collected across multiple environments and seasons.
“Creating large, complex data ‘flywheels’ is the future of breeding,” says Alverez, “but they have to be scalable, and the data streams are only part of the picture. They have to be cheap and interpretable data streams in order to be truly impactful.”
Washburn agrees that data availability will help determine how useful AI tools become and which breeding programs can adopt them.
“Large companies, breeding networks and other groups that are able to aggregate large amounts of data will be able to leverage the newest AI methods the fastest and to the greatest benefit,” he notes. “At the same time, innovations that allow AI models to perform well with less data will be critical to enabling the use of AI in breeding, especially for public breeding programs, and programs focused on crops with less use/economic impact.”
That creates a particular challenge for public breeders and programs working in smaller crops. The organizations with the most data may benefit first, while advances that allow models to work with smaller datasets could determine how broadly AI spreads through the breeding community.
Zamft says public research infrastructure remains essential to agricultural innovation. Seed companies have built substantial research capabilities, but academic and public-sector research still feeds discoveries into commercial breeding programs. As AI increases the value of large datasets, he expects pressure to grow for new approaches to sharing data.
Why AI May Push Breeders to Share More Data
For Zamft, collecting data at scale is a prerequisite for the kind of modeling AI enables.
“Heritable Agriculture has developed our model through using public data, running our own trials and working with partners,” he says, “But I’ve long been an advocate of rip-the-bandage-off all-out data collection efforts, which would probably best come in the form of a public-private partnership.”
Heritable Agriculture is currently in a partnership with The Gates Foundation funding. Its scientists are collaborating with researchers around the world to generate multi-omic datasets from drought and well-watered field trials of tropical maize at sites in Sub-Saharan Africa.
Heritable Agriculture then uses its AI to mine the dataset for genes associated with drought resistance. The company also feeds those results into its models to create digital twins of crops and run digital trials. At the same time, the raw data and selected genes will be globally accessible to breeders.
The effort illustrates a larger tension surrounding AI in crop breeding. Better models can expand what breeders are able to predict, select and test, but the technology still depends on the information available to it. As AI moves further into breeding programs, the ability to collect, organize and share useful data may become as important as the algorithms themselves.
