AI Designed 391 Plant Immune Systems. Only 71 Worked, and That Is Still a Breakthrough

Plant breeders usually search nature for genes that can recognise a disease. Researchers in China have now tested a faster idea: design a new immune sensor for the pathogen they want to stop, then improve it inside a living plant.

KEY NUMBER
71 of 391 initial designs worked as intended, an 18.2% success rate that shows both the promise and the current limits of AI-guided biology.

Why crop immunity needs a faster route

Plants do not have mobile immune cells like humans, but their cells carry receptors that detect molecules produced by attackers. Breeders can move useful resistance genes into crops, yet this process depends on finding the right gene in nature. Pathogens also evolve, sometimes overcoming protection that took years to develop.

The study, published in Science on July 23, 2026, explores a more programmable approach. Instead of waiting for a suitable receptor to be discovered, the team tried to build one around a chosen pathogen protein.

How the synthetic receptors were built

The researchers started with a class of plant immune sensors called nucleotide-binding leucine-rich repeat receptors, or NLRs. Some NLRs contain a small interchangeable region that acts like a bait. When a pathogen interacts with it, the receptor can trigger the plant's defence response.

Using AlphaFold 3 and BindCraft, the team designed new proteins intended to bind specific pathogen targets. Those binding modules were inserted into a rice immune receptor called Pikm-1. The resulting constructs were described as synthetic plant immune receptors, or SPIRs.

The targets came from a broad range of plant enemies, including viruses, bacteria, fungi and oomycetes. Of 391 designs, 71 recognised the intended target and activated immunity. That is not a high hit rate in everyday terms, but biological engineering rarely starts with certainty. A platform that produces dozens of functional candidates can still save substantial time.

AI was only the first filter

The result is also a useful correction to the idea that AI can simply print a working biological machine. Many designs were inactive. Others were autoactive, meaning they switched immunity on without the pathogen. Constant immune activation can stunt a plant and reduce yield.

To improve weak candidates, the researchers used directed evolution inside plant cells. Their method, called geminivirus replicon-assisted in planta directed evolution, generated variations and selected better performers. It strengthened immune activity while reducing unwanted self-activation.

An optimised receptor was then placed into Nicotiana benthamiana, a plant widely used in laboratory research. The modified plants resisted Tomato brown rugose fruit virus, an important pathogen of tomato and pepper crops.

What this could change

The researchers say new receptor candidates can be generated within weeks. If that pace survives crop development and regulation, breeders could respond more quickly to emerging disease strains instead of beginning a long search through natural diversity.

The approach could also broaden protection beyond one familiar virus. Because designers choose the protein target, the same workflow could in principle be adapted to many pathogens. It is best understood as a candidate generator, however, not an instant vaccine for plants.

What happens next

The most important tests will move the receptors into real crops and expose them to varied field conditions. Scientists will need to measure not only disease resistance, but also growth, yield and whether protection remains effective as the pathogen changes.

The 320 unsuccessful first designs are not a footnote. They reveal why human testing and biological evolution remain essential. The breakthrough is not that AI solved plant disease alone; it is that computation, synthetic biology and evolution can now work as a pipeline for building immunity that nature did not hand us ready-made.