MIT PLOW: generative AI designs new herbicides to replace glyphosate — $53M from ARPA-H

How PLOW designs a weed killer, step by step#

PLOW runs like a loop with four stations, and the AI sits at the front of it. The generative models do the inventing; the lab and the greenhouse do the judging; the results feed back into the models. Here is the pipeline, as MIT described it on November 10.

Step 1: propose. Generative AI models predict molecules that could kill weeds — either chemical compounds or small interfering RNA (siRNA) — that disrupt the same plant enzymes glyphosate targets. Glyphosate works by interfering with an enzyme called EPSP synthase, which plants need to produce several critical amino acids. PLOW aims at that same choke point, but with new molecular shapes. The team is led by Connor Coley, Regina Barzilay and Alexander Tropsha, and the models are also told to optimize for three things at once: low cost of synthesis, low risk of inducing resistance, and low risk of harming humans or causing ecological damage.

Step 2: screen. The models are expected to generate billions of possible candidates. A second team, led by Angela Koehler at MIT and Jaime Cheah at the Broad Institute, runs high-throughput screening on those libraries — checking which molecules actually hit their target, then sending all of that screening data back to the computational team so the models improve. It is the same generate-and-test loop that AI-designed enzymes went through in Manchester this October, where only 4 of 67 candidates worked — except here the candidate pool is measured in billions, not dozens.

Step 3: grow. Once the candidate pools are narrowed down, a third team starts greenhouse studies to see how each molecule performs under different agricultural conditions. That team is led by Neal Stewart and Scott Lenaghan at the University of Tennessee, Knoxville, Patrick Tranel at the University of Illinois, Tim Willson at UNC and Andee Wallace, CEO and founder of Robigo. The first greenhouse tests of initial RNAi candidates are expected to be complete by the end of the program’s first year.

Step 4: clean up. The project has a second product line entirely: living microbial systems that degrade agrochemicals before they reach food, soil or waterways. Engineered bacteria, fungi and srRNA will convert chemicals and formulation ingredients into biomass — a way to cut exposure even while herbicides are still in use.

The announcement#

MIT announced the project on November 10, 2026. The funding — up to $53 million — comes from ARPA-H’s STREAM program (Systems for Tracking and Resilience in Efficient Agricultural-input Management), which the agency launched on August 17, 2026 to develop safer, more effective weed-management technology. The award itself started September 18, 2026, according to ARPA-H. Christopher Voigt, head of MIT’s Department of Biological Engineering, is the principal investigator, and the team spans the Broad Institute of MIT and Harvard, the University of North Carolina at Chapel Hill, the University of Tennessee at Knoxville, the University of Illinois at Urbana-Champaign, and the startup Robigo.

The framing in MIT’s announcement is blunt. “With herbicides like glyphosate and atrazine, there’s a real crisis because of resistance mutations that are coming,” Voigt said. He tied the problem to supply chains as well as biology: combine the resistance mutations with “issues around national security of the supply chain of the precursors to herbicides,” and the U.S. could find itself short of chemicals it considers essential for high crop yields and economical food prices. One of the project’s stated aims is to reduce American dependence on imported phosphorus, a critical ingredient in glyphosate.

Why this lands now#

The numbers explain the urgency. More than 1 billion pounds of herbicide are used in the United States every year. Herbicides prevent significant crop loss and enable regenerative farming practices that protect the soil — but after many years of use, some weeds have acquired genetic mutations that let them resist the most common herbicides, which makes the chemicals less effective and threatens crop yields. Exposure is a second worry: MIT notes it may contribute to increased risk of cancer, metabolic and endocrine disorders, and that farm workers face the most exposure while many U.S. residents encounter small amounts through food and drinking water.

PLOW is also a sign of where AI-for-science money is going. The same month, DeepMind’s VP said AI science is hitting a “phase transition”, with new releases coming in weeks, and OpenAI bought Orchestra Bio to put AI agents in charge of drug-development work. Designing molecules was the first proven win for AI in biology — antibiotics, then enzymes, now herbicides. Agriculture is simply the next field where the generate-and-test loop has matured enough to bet $53 million on. And unlike a drug, a herbicide does not need a decade of clinical trials to reach the field, which is why the greenhouse clock — tests done by the end of year one — is the number to watch.

What to watch#

The first real checkpoint is the greenhouse. If the initial RNAi candidates clear their first-year tests, PLOW will have shown that AI-designed molecules survive contact with real plants — the step where most computational chemistry fails. If they do not, the feedback loop still gets its first full cycle of screening data, which is the point of building the loop in the first place.

Two caveats travel with the announcement. The funding is “up to” $53 million, the standard ARPA-H formulation, and the project is a research program, not a product: no farmer will spray a PLOW molecule next season. And the safety case is a design goal, not a result — the models optimize for low harm to humans and ecosystems, but that optimization has to survive the same screening and greenhouse gauntlet as everything else. The promise is a weed killer designed, from the first molecule, to be safer and harder to resist. The proof is a year away, in a greenhouse in Tennessee.