Breeding the Next Wheat

Norman Borlaug spent roughly twenty years breeding the semi-dwarf wheat varieties that triggered the Green Revolution. Working in Mexico from 1944, crossing thousands of plants manually, evaluating field performance across seasons, selecting and backcrossing promising lines—it was painstaking empirical science that produced, by the early 1960s, varieties that quadrupled yields per acre when combined with nitrogen fertilizer and irrigation. The first deliveries of Mexican wheat to Pakistan in 1965 helped avert what many demographers had projected would be a catastrophic famine.

The timelines haven’t changed much since then. Developing a new crop variety from initial cross to commercial release still typically takes 10-15 years for major cereals. The breeding cycle has a fixed biological component: you have to grow the plants, wait for them to flower, cross them, harvest seed, grow the next generation. Each year of field testing produces one cycle of phenotypic data.

AI has not eliminated the biology. It has dramatically compressed the decision-making—which plants to cross, which lines to discard, which genomic markers predict the phenotypes you want—and in doing so, is beginning to meaningfully accelerate the timeline.

The foundation of modern AI-assisted breeding is genomic prediction. The idea is straightforward: rather than waiting to grow a plant to maturity to measure its yield, drought tolerance, disease resistance, or protein content, you read its genome and use a statistical model to predict those traits from the DNA sequence. This allows breeders to screen thousands or millions of candidate lines at the genomic level and advance only the most promising to field testing—reducing the number of field trials required and focusing expensive field resources on candidates already predicted to perform well.

Genomic prediction models have been in use in animal breeding since the 1990s, where they transformed dairy cattle selection (the genetic improvement rate in dairy cattle approximately doubled after genomic selection was introduced). The application to crops is more recent and more complex, because crop breeding involves more distinct environments, more traits of interest, and more complex gene-by-environment interactions.

The current generation of deep learning approaches to genomic prediction—treating the genome as a sequence and using transformer-based architectures to learn long-range interactions between genetic variants—is substantially more accurate than earlier GBLUP (genomic best linear unbiased prediction) methods for complex polygenic traits. A 2024 paper from INRAE (France’s national agricultural research institute) demonstrated that transformer-based genomic prediction for wheat grain protein content improved prediction accuracy by 15-20% compared to GBLUP, with larger gains in populations with limited historical phenotyping data.

For a breeder deciding which of 50,000 candidate lines to advance to field trials, a 15-20% improvement in prediction accuracy translates directly to fewer expensive field cycles wasted on lines that would have failed.

The other major AI application in plant breeding is image-based phenotyping: using computer vision on field or greenhouse images to measure traits that previously required manual measurement. Leaf area, canopy architecture, flag leaf angle, grain-filling rate, root architecture (from transparent gel media or X-ray CT imaging)—all of these phenotypes are slow and expensive to measure manually at scale and can now be captured automatically.

A field phenotyping platform—cameras on a ground robot or drone measuring a breeding nursery weekly—can generate more phenotypic data in a season than a traditional breeding program generates in a decade. The data quality is different (automated measurements have different error profiles than manual ones), but the volume enables statistical approaches that weren’t possible with thin datasets.

LemnaTec’s Scanalyzer systems, Field Pheno platforms from companies like Phenospex, and increasingly custom systems built by large seed companies like Bayer Crop Science and Corteva are making high-throughput phenotyping standard in advanced breeding programs. The challenge is analysis: generating terabytes of image data per season and extracting meaningful phenotypic estimates from it is a machine learning problem at the frontier of what current models can do reliably.

The gene editing layer adds another dimension. CRISPR-Cas9 and related tools allow targeted modification of specific genomic locations—turning a gene off, introducing a specific variant from a wild relative, or modifying regulatory regions to change expression levels. AI contributes to this pipeline through guide RNA design (predicting which guide RNAs will edit the target site efficiently and avoid off-target edits), prediction of the functional effect of specific edits on trait phenotypes, and identification of which genomic targets are most promising candidates for improvement.

Pairwise, a startup spun out of the Broad Institute, used AI-guided CRISPR to develop a cherry tomato with lower levels of a bitter compound (chlorogenic acid) and a strawberry with longer shelf life. These are relatively simple single-gene modifications. The more ambitious targets—improving nitrogen use efficiency, encoding drought tolerance, enhancing photosynthetic efficiency—involve complex polygenic traits and are harder to achieve through gene editing alone.

The productivity ceiling that plant breeders are trying to push through is significant. Global crop yield growth has been slowing since the 1990s. Wheat yield growth was approximately 3% per year during the Green Revolution; it has been around 1% per year since 2000. Maize has held up somewhat better but faces similar trends. Meanwhile, climate change is projected to reduce average crop yields in major production regions by 2-6% per decade under moderate warming scenarios, according to the most recent IPCC assessment.

The arithmetic is uncomfortable: yield growth is decelerating while climate stress is accelerating. The trajectory implies that without significant improvement in genetic material, food production per capita will stagnate or decline in stressed regions by mid-century.

Whether AI-accelerated breeding can maintain the yield growth trajectory needed to stay ahead of this combination depends partly on the biological constraints and partly on the institutional ones. The biological constraints are real: photosynthetic efficiency, theoretical yield potential, and water use efficiency all have physical limits. The crops being grown today are not that far from theoretical optimum for several key metrics.

The institutional constraints are equally significant. Plant variety development and registration—the regulatory process that a new variety must pass before commercial deployment—typically takes 2-5 years in most jurisdictions, after the breeding process has produced the candidate. This regulatory pipeline was designed for conventionally-bred varieties and is poorly adapted to the faster generation of candidates that AI-assisted breeding enables. You can breed faster than you can register.

Additionally, the seed industry is highly concentrated. Four companies—Bayer (Monsanto), Corteva (DuPont Pioneer and Dow AgroSciences), Syngenta (ChemChina), and BASF—control roughly 60% of global commercial seed sales. Their breeding programs are the best resourced and will adopt AI tools fastest. But the crops and traits they prioritize are determined by commercial return, not by the food security needs of subsistence farmers in sub-Saharan Africa growing sorghum or cassava.

Borlaug’s wheat reached Pakistan in 1965 through a combination of scientific breakthrough and vigorous international political commitment—the Rockefeller Foundation, the Mexican government, CIMMYT, and USAID all participated in getting the technology to the farmers who needed it. The institutional infrastructure for that kind of diffusion is weaker now, not stronger.

CGIAR, the international agricultural research consortium that carries on Borlaug’s mission, has an annual budget of roughly $1 billion—less than Bayer’s annual R&D expenditure on just one crop. The AI-powered breeding tools that commercial seed companies are using are largely proprietary. The germplasm collections that form the genetic raw material for breeding improvements are maintained in public genebanks, but access to the computational tools to exploit them is not equally distributed.

The next wheat—the climate-adapted, drought-tolerant, high-protein variety that will feed people in regions where current varieties fail as temperatures rise—is being bred right now, somewhere in a server farm as much as a field trial. Whether it reaches the farmers who most need it depends on choices being made in company boardrooms and agricultural development budgets that have nothing to do with the algorithm’s elegance.

Borlaug won the Nobel Peace Prize in 1970. The recognition was that food security is peace. The same logic applies to the tools that maintain food security as the climate changes. They have to work everywhere, not just in the markets that can pay for them.