MIT Lands Up to $52.9M From ARPA-H to Design Safer Herbicides With Generative AI

MIT will lead a project backed by up to $52.9 million from ARPA-H, the U.S. Advanced Research Projects Agency for Health, to design safer herbicides with generative AI and to build biological systems that break down agricultural chemicals. The institute announced the award on November 10, 2026. ARPA-H lists September 18, 2026 as the award start date.

Christopher Voigt, head of MIT's Department of Biological Engineering, serves as principal investigator, and MIT holds the prime award. The project, called Precision Learning for On-target Weed control (PLOW), sits inside ARPA-H's Systems for Tracking and Resilience in Efficient Agricultural-input Management program, known as STREAM. The U.S. Department of Health and Human Services launched STREAM on August 17, 2026, framing it as an effort to protect farm workers, consumers and the environment while maintaining crop yields.

The scale of the commitment places PLOW among the larger federal bets on generative AI applied to chemistry rather than text or code. Its test is unusually demanding: a candidate library counts for nothing unless laboratory screening and greenhouse studies confirm that molecules work under real agricultural conditions while meeting cost, resistance and safety constraints.

Why does U.S. agriculture need new herbicide chemistry?

PLOW targets a structural tension in American farming. More than 1 billion pounds of herbicide are applied each year to prevent significant crop losses and to enable regenerative practices that protect soil. Repeated use, however, has allowed some weeds to acquire resistance mutations that blunt the chemicals' effect.

Glyphosate illustrates the problem. It kills weeds by interfering with EPSP synthase, an enzyme plants need to produce several critical amino acids. Resistance mutations make that mechanism less effective, eroding the tool on which much of modern row-crop agriculture depends.

Health concerns add a second design constraint. Herbicide exposure may contribute to increased risks of cancer, metabolic and endocrine disorders and other adverse outcomes, MIT's announcement states. Farm workers face exposure through frequent herbicide use, while many U.S. residents encounter small amounts through food and drinking water. PLOW's researchers therefore include human and ecological harm among their candidate-selection criteria alongside weed-killing performance.

Supply security forms the third objective. The intended alternatives would reduce U.S. dependence on imported phosphorus, a critical ingredient in glyphosate. Voigt said the combination of resistance and insecure precursor supply chains could leave the country without agrochemicals that are "absolutely essential for getting the high crop yields and economical prices that the United States now sees."

How does the generative AI pipeline work?

PLOW's first research group will build models that propose either chemical compounds or small interfering RNA (siRNA) molecules designed to disrupt the plant enzymes targeted by glyphosate and other herbicides. Connor Coley and Regina Barzilay of MIT and Alexander Tropsha of the University of North Carolina lead that work.

The models are intended to optimize candidates across several dimensions simultaneously:

• effectiveness against target enzymes;
• low synthesis cost;
• low risk of inducing weed resistance;
• low risk of harming humans;
• low risk of ecological damage.

The team expects the models to generate billions of possible candidates. That scale makes the next stage decisive. Angela Koehler of MIT and Jaime Cheah of the Broad Institute will lead high-throughput screening to determine which molecules actually hit their targets. The screening data will then flow back to the computational group to improve its models, creating an iterative design-and-test loop rather than treating AI output as a finished herbicide.

The structure matters for how the field reads the results. Generative models have produced large candidate libraries before; the open question has been whether experimental validation can keep pace. PLOW wires the two together by contract.

What happens after the models propose candidates?

Once the pools are narrowed, a third group will evaluate each molecule under different agricultural conditions in greenhouse studies. That group is led by:

• Neal Stewart and Scott Lenaghan at the University of Tennessee at Knoxville;
• Patrick Tranel at the University of Illinois at Urbana-Champaign;
• Tim Willson at UNC;
• Robigo founder and CEO Andee Wallace.

The first greenhouse tests of initial candidates based on RNA interference (RNAi) are expected to be complete by the end of the program's first year. That timeline gives the project an early, concrete checkpoint: RNAi-based candidates reach agricultural evaluation before the full chemical pipeline does.

Can engineered biology clean up what is already in the field?

PLOW also addresses chemicals already in use, not only the ones it hopes to replace. The researchers plan living microbial systems in which engineered bacteria and fungi convert agrochemicals and formulation ingredients into biomass before those substances reach food, soil or waterways. The approach is intended to provide a scalable way to reduce environmental and human exposure.

The degradation work gives the project a second deliverable that does not depend on new herbicides reaching the market. Even if candidate molecules fail the greenhouse stage, the microbial systems would still target exposure from products farmers apply today.

Where does PLOW fit in the STREAM program?

Taken together, PLOW's discovery and degradation work covers two of the four technical areas in STREAM's program plan: AI-assisted discovery of next-generation herbicides, and the removal or degradation of agricultural chemicals. The other two areas cover different ground:

• precision and nonchemical weed control, including laser ablation and miniature autonomous weeders;
• improved monitoring of herbicides and associated chemicals.

ARPA-H, created to pursue high-risk health breakthroughs outside conventional funding channels, is extending that mandate to agricultural inputs on the argument that herbicide exposure is a public-health problem — for farm workers first, and for the general population through residues in food and drinking water.

What is at stake for generative AI in the sciences?

The award ties generative AI's scientific value to a demanding experimental test. Billions of AI-proposed candidates are useful only if screening identifies molecules that hit their targets, and screened hits are useful only if greenhouse studies show they perform under agricultural conditions while meeting cost, resistance and safety constraints. Each stage can reject the output of the previous one.

That framing separates PLOW from demonstrations that stop at computational benchmarks. Barzilay, who co-leads the modeling group, has built her career on connecting machine learning to molecular discovery; here the connection runs all the way to greenhouse soil.

The program's first hard date is close. With RNAi greenhouse tests due by the end of year one, the project will show early whether its design-and-test loop produces candidates that survive contact with agriculture — and whether generative AI can deliver molecules judged not by novelty alone, but by yield, cost and human safety at the same time.