Accelerated development of 4-HPPD inhibitors using a hybrid deep learning approach with Bayesian-guided reinforcement learning.

Generative AI methods for enzyme inhibitor design are constrained by imbalanced training data and the difficulty of simultaneously optimizing binding affinity and drug-like physicochemical properties. We describe a hybrid approach combining a DrugEx-RNN generative model with Bayesian-guided reinforcement learning, using predicted docking scores as the reward signal, to generate novel inhibitors of 4-hydroxyphenylpyruvate dioxygenase (HPPD). The Bayesian reward approximation allows effective optimization despite limited active training examples. Generated compounds achieved docking scores 10-20% beyond known HPPD inhibitors and occupied novel regions of chemical space relative to both commercial inhibitors and ChEMBL reference sets. Three compounds were synthesized and assayed; the most active, TP-054, inhibited HPPD with a Ki of 44 nM, a roughly 3.5-fold improvement over the commercial inhibitor topramezone (Ki = 157 nM). TP-054 also showed systemic herbicidal activity against Echinochloa crusgalli and Portulaca oleracea with no measurable phytotoxicity to Zea mays at 300 g ai/ha, confirming that the computationally optimized binding translated to bioavailability and target engagement in whole organisms. The approach is general to any enzyme target with sufficient structural data for docking.