- Published on
Bioorganic Chemistry – ML Design of Crop Safe Herbicides.
- Author
-
-
- Name
- Posts
- Posts
-
We've published work in Bioorganic Chemistry applying our AI drug-design pipeline to a very different target, herbicide discovery. The molecule here is 4-hydroxyphenylpyruvate dioxygenase (HPPD), one of the few commercially successful modern herbicide targets, and the goal was to generate genuinely new inhibitor scaffolds rather than variations on the existing triketone chemistry. Data is what makes this hard. Generative AI models are usually trained on large pharmaceutical datasets, but specialized enzyme targets like HPPD have far fewer known active compounds, and agricultural molecules have to satisfy physicochemical requirements like solubility, foliar uptake, and systemic movement in plants that pull the useful chemical space away from where medicinal-chemistry-trained models are comfortable.
Our approach pairs a DrugEx-RNN generative model with reinforcement learning, using docking-based binding scores as the reward signal. Computer-aided drug design (CADD) generates new data for any molecule the model proposes, but a full docking calculation runs far too slowly to evaluate every candidate across a reinforcement learning loop that samples thousands of structures per round, so we approximate the docking reward with a Bayesian ridge regressor over molecular fingerprints that is fast enough to guide broad exploration of chemical space in real time. We then periodically run actual CADD calculations on promising candidates and fold those results back into the regressor, so the reward model sharpens as generation proceeds rather than drifting away from ground truth. The generated compounds scored 10 to 20% better than known HPPD inhibitors on docking and occupied regions of chemical space that neither the commercial herbicides nor the ChEMBL reference set had explored.
To show that these results were biologically meaningful, we synthesized three of the designed compounds and assayed them. The best, TP-054, inhibited HPPD with a Ki of 44 nM, about 3.5 times more potent than the commercial herbicide topramezone, and it showed systemic herbicidal activity against grass and broadleaf weeds while leaving maize unharmed at high application rates. That combination of potent weed control and crop safety is exactly the balance selective herbicides need, and it confirms that the pipeline's in silico gains translated all the way through to bioavailability and target engagement in whole organisms. The method isn't specific to HPPD, and should extend to any enzyme target with enough structural data for docking.