Weeds are learning to survive the chemicals farmers rely on, and those chemicals raise their own health worries. An MIT-led team has been awarded up to $53 million by ARPA-H, the US health research agency, to use generative AI to design replacement weed killers with human health in mind. It is a good example of AI-driven molecule design moving from medicine into agriculture, with a clear test of whether the predictions hold up in a greenhouse.
The stakes are large. More than 1 billion pounds of herbicide are used in the United States each year. Weeds are developing resistance to the most common products, including glyphosate, which works by blocking a plant enzyme called EPSP synthase, and atrazine. Exposure to herbicides may contribute to higher risks of cancer and metabolic and endocrine disorders, and farm workers are especially exposed. MIT says new alternatives could also reduce US dependence on imported phosphorus, a key ingredient in glyphosate. Christopher Voigt, head of MIT's Department of Biological Engineering and the project's principal investigator, warns that resistance plus supply-chain risk could leave the country without essential farm chemicals.
The project, called PLOW, under ARPA-H's STREAM programme, is split into three teams. The first, led by MIT's Connor Coley and Regina Barzilay with Alexander Tropsha of the University of North Carolina, will build generative models to propose molecules, either chemical compounds or small interfering RNA, that disrupt the same plant enzymes current herbicides target while carrying low risk of resistance or harm to people and ecosystems. The models are expected to generate billions of candidates. A screening team at MIT and the Broad Institute will test them at high throughput and feed results back to improve the models. A third team, including the University of Tennessee, the University of Illinois and the start-up Robigo, will run greenhouse trials, with the first tests of RNA-based candidates due by the end of the programme's first year. The team will also engineer microbes to break down agrochemicals before they reach food, soil or water.
The design loop, generate, screen and feed back, is the same one behind AI drug discovery, and Barzilay's group has used generative models for hard-to-treat diseases. What changes here is the target: plants rather than people.
The caveats are significant. This is a funding announcement, not a result. Up to $53 million is a ceiling, not a guarantee, and no molecules have yet been reported. Generating billions of candidates is easy; finding ones that are effective, cheap to make, safe and slow to provoke resistance in real fields is the hard part, and regulatory approval of any new herbicide takes years.