Pocket-Conditioned Generation and Down-Selection of Non-Oxime Reactivators

Acetylcholinesterase reactivation is a structure-constrained design problem in which chemical plausibility, access to central nervous system, pocket compatibility, and reactivation geometry must be considered together. We report a computational workflow for generating and down-selecting candidate non-oxime reactivators in the paraoxon-inhibited human acetylcholinesterase pocket. For generating the candidates, the first-round utilitezed pocket-conditioned DrugFlow, but due to solubility issues, the second-round candidates were output form FlexFlow, DrugFlow-OOD, DrugFlow preference alignment, and pocket-specified BoltzMol-1. A property-guided DrugDiff output provided a protein-unconditioned baseline. Although 93.9% of the DrugDiff subset passed Lipinski criteria for druglikeness, its median synthetic accessibility (SA) score signalled that it was difficult to synthesize compared against other generative approaches. A strict predicted blood-brain barrier and solubility gate retained 31.0% of Base DrugFlow, 61.0% of FlexFlow, 27.5% of DrugFlow-OOD, 3.8% of DrugFlow-PA, and 35.1% of BoltzMol-1. The second-round top-ten lists had lower SA scores and higher quantitative estimate of drug-likeness (QED) values than the DrugDiff shortlist, while DrugDiff had fewer aggregate heuristic violations. This work compares the best performing generative drug discovery models with suggested down sleection approaches for human acetylcholinesterase reactivation.