MIT Wins Up to $53M for AI That Could Replace Glyphosate

When people talk about generative AI, they usually think of chatbots and coding assistants. But one of the best-funded stories right now is unfolding in chemistry: models are learning not to write text but to invent molecules. MIT has just received up to $53 million from ARPA-H for exactly this kind of task, and the target is not drugs but herbicides.

What Problem Is at Stake

In the US, more than a billion pounds of herbicides are used every year. They protect crops from weeds and make soil-friendly farming possible. But they come with two problems. First, weeds acquire resistance mutations. Glyphosate, for example, inhibits the enzyme EPSP synthase, without which a plant cannot produce several important amino acids. Mutations let weeds bypass this block, and the product stops working. Second, there are health risks: herbicides are linked to an increased risk of cancer and of endocrine and metabolic disorders. Farmers receive the largest dose, but traces of the chemicals end up in the food and water of everyone else.

Add geopolitics. Phosphorus, a key component of glyphosate, is largely imported by the US. Christopher Voigt, head of MIT's Department of Biological Engineering and leader of the PLOW project, says bluntly that weed resistance combined with a vulnerable supply chain could leave the country without basic agrochemistry.

How It Works: The "Model – Screening – Greenhouse"

Loop

The project is split into three teams, and the main thing about it is not a single model but a closed loop.

1. Generative models. The group of Connor Coley, Regina Barzilay and Alexander Tropsha will predict molecules or small interfering RNAs (siRNAs) that switch off the same key plant enzymes as current herbicides. The models must account for several goals at once: cheap synthesis, low risk of resistance emerging, and minimal harm to humans and the environment.
2. High-throughput screening. The models are expected to produce billions of candidates. The group of Angela Koehler and Jamie Chia at the Broad Institute will test the libraries for target hits. Screening results will return to the computational team to improve the models. In essence, this is active learning on real laboratory data.
3. Greenhouses. Once the candidate pool narrows, a third team will test how the molecules perform under different agronomic conditions. It includes researchers from the Universities of Tennessee, Illinois and North Carolina, and the startup Robigo. The first tests of RNAi candidates are due to be completed by the end of the program's first year.

A separate track is devoted to living microbial systems. These are engineered bacteria and fungi that break down agrochemicals before they reach food, soil and water, converting them and the components of formulations into biomass.

What Really Matters Here

First, the limitations. This is a funding announcement, not a result. The material contains no model architectures, no metrics, and no evidence that the generative approach will find better molecules than classical screening. The road from a greenhouse candidate to a registered herbicide is long, and regulation lies ahead, which AI will not remove. So don't expect a product on the shelf in the next few years.

But for those building AI products, there are three useful signals here.

• The value is in the feedback loop. A model on its own is worth nothing until it is fed fresh experimental data. The winner is whoever has built a lab or a real-world environment into the training loop.
• Multi-objective optimization is becoming the norm. What's needed are not just potent molecules, but ones that are cheap to make, hard for mutations to "hack," and safe to use. If you are designing a generative system, build constraints into the objective function from the start.
• Agrochemistry is opening up as a market for AI. Government money and a partnership of universities, the Broad Institute and the startup Robigo show that the niche for founders at the intersection of biology, chemistry and ML is real and funded.

If you work with AI in science, watch for the first data on RNAi candidates: they will show whether this loop works in practice.