Integrated multi-stage screening assisted discovery and optimization of spirooxindole MDM2 inhibitors.

The rational design of novel MDM2 inhibitors with superior biochemical properties represents the most consequential outcome of contemporary computer-aided drug discovery. Here, we report the discovery of spirooxindole-based p53-MDM2 interaction antagonists through a systematic, integrated multi-stage screening workflow that couples deep-learning-driven molecular remodeling with hierarchical physics-based assessment. Starting from 53 reported spirooxindoles, we first constructed a ligand-based pharmacophore model (ADHHR_1) and a validated PLS-5 3D-QSAR model (Q2 = 0.719) to quantify the microscopic contributions of hydrogen-bond donors, hydrophobic groups, and electron-withdrawing moieties to binding affinity. Leveraging the Modof graph-remodeling network trained on 1970 azaindoles, we generated 22 lead-like spirooxindole analogues (N01-N22). Hierarchical virtual screening via molecular docking identified 12 analogues surpassing the clinical reference Nutlin 3a in predicted affinity, the top candidate N14 achieved a binding energy of -9.2 kcal/mol versus -8.4 kcal/mol for Nutlin 3a. Sequential 500 ns molecular dynamics simulations, coupled with MM-PBSA free-energy decomposition, revealed that the N14@MDM2 complex possesses a markedly more favorable binding free energy (-41.5 kcal/mol) than Nutlin 3a (-35.42 kcal/mol), concomitant with enhanced conformational stability evidenced by lower backbone RMSD fluctuations and a tightly maintained radius of gyration. Per-residue interaction profiling further distinguished N14 by its unique contacts with Lys51 and His96, supplementing the shared hydrophobic anchor residues (Leu54, Ile61, Met62). Collectively, these biochemically supported findings nominate N14 as a priority lead with predicted advantages over the reference standard, establishing a generalizable AI-driven paradigm for protein-protein interaction inhibitor discovery that balances computational efficiency with predictive rigor.