Large-Scale Prospective Virtual Screening Using a Co-Folding-Rescoring Workflow for NPM1-TYK2 Inhibitor Discovery

Conventional large-scale structure-based virtual screening typically relies on fixed-receptor docking, in which ligand poses are sampled against a predefined receptor conformation. Protein-ligand co-folding models can instead generate candidate complex conformations by jointly modeling ligand placement and local receptor-ligand structural adaptation. Although recent benchmark studies have demonstrated the promise of AlphaFold3-like approaches for retrospective virtual screening, their practical application to target-specific, large-scale prospective screening with experimental validation remains limited. Here, we developed and prospectively applied a co-folding-rescoring workflow for NPM1-TYK2 inhibitor discovery. Using an in-house set of 4608 compounds experimentally annotated in Ba/F3-NPM1-TYK2 cells, we evaluated 13 pipelines combining AlphaFold 3, Protenix, Glide, or AutoDock Vina pose generation with scoring functions or confidence metrics. AlphaFold 3 followed by AutoDock Vina rescoring achieved the numerically highest BEDROC value, closely comparable to Protenix + Vina, and was selected for screening a commercial library of more than 300,000 compounds. AF3-generated complex conformations were subsequently rescored with Vina and prioritized by physicochemical assessment, clustering, and binding-mode inspection. Of 30 compounds tested experimentally, FV13 was identified as a cell-active primary hit, with a TYK2 kinase IC50 of 21.70 μM and a Ba/F3-NPM1-TYK2 cellular IC50 of 17.07 μM. Subsequent analog exploration yielded TX14, exhibiting an approximately 9.4-fold improvement in biochemical TYK2 inhibitory potency (IC50 = 2.31 μM). Additional selectivity profiling and Western blot analyses further characterized these compounds and indicated attenuation of downstream STAT5 phosphorylation at higher concentrations. Together, these findings demonstrate the feasibility of integrating co-folding, empirical rescoring, and target-specific experimental data for large-scale prospective hit discovery.