Repurposing nucleos(t)ide analogs as purine nucleoside phosphorylase inhibitors through Gaussian accelerated...

Purine nucleoside phosphorylase (PNP) is a key enzyme in the purine salvage pathway and a promising therapeutic target for T-cell-mediated disorders such as leukemia, lymphoma, psoriasis, and rheumatoid arthritis. We used an integrated computational drug-repurposing approach to identify nucleos(t)ide analogs with inhibitory potential against human PNP. A library of approximately 800 nucleos(t)ide analogs underwent machine learning-assisted virtual screening, and the top 20 candidates were further evaluated by molecular docking. The four highest-ranked compounds, based on docking scores (-7.8 to -10.0 kcal/mol), were selected for 500 ns long Gaussian accelerated molecular dynamics (GaMD) simulations. We estimated binding free energies using trajectory-based MM-PBSA calculations and supplemented these with AI-assisted affinity predictions from SG-ML-PLAP and K-DEEP. Among the compounds, 6-hydroxy-flavin adenine dinucleotide (C2) showed the most favorable binding profile, although flavin adenine dinucleotide (C1) had the best docking score (-10.0 kcal/mol). C2 achieved a trajectory-based PBSA binding free energy of -17.23 ± 3.45 kcal/mol, significantly better than the reference inhibitor Immucillin-H (-8.11 ± 2.56 kcal/mol). The other lead compounds, C1, C3, and C4, also demonstrated favorable PBSA binding free energies of -12.90 ± 5.05 kcal/mol, -11.71 ± 3.64 kcal/mol, and -13.19 ± 3.56 kcal/mol, respectively. These results indicate that combining machine learning-assisted screening, molecular docking, enhanced molecular dynamics simulations, MM-PBSA calculations, along with AI-based affinity predictions is an effective strategy for identifying promising nucleos(t)ide analogs for future experimental validation as human PNP inhibitors.