Machine Learning‐Guided Discovery of Natural MDM2 Inhibitors: A Multistage In Silico Pipeline from Screening to ADMET Profiling

Abstract

ABSTRACT Tumor protein p53 (TP53) and mouse double minute two homolog (MDM2) regulate each other via an autoregulatory feedback loop that is frequently disrupted by MDM2 overexpression or mutation, a hallmark in sarcomas, glioblastomas, and breast carcinomas. In the absence of FDA‐approved MDM2 inhibitors, a multi‐stage in silico strategy is applied to identify novel candidates from COCONUT, a comprehensive natural product library. Using experimentally validated ChEMBL data, 40 machine‐learning models are trained and evaluated; the best RandomForestClassifier selects 116 compounds from approximately 700,000 after sequential PAINS, Brenk, and Lipinski filtering. Docking‐based screening prioritizes two leads with binding energies of (CNP0492204.2) and (CNP0385629.2), both engaging in key interactions and featuring a novel quinazolindione‐based peptidomimetic scaffold. Molecular dynamics simulations confirm stable binding to MDM2: CNP0492204.2 induces local N‐terminal loop flexibility, whereas CNP0385629.2 favors deep burial and ‐stacking with Tyr56. Despite distinct modes, MM/GBSA calculations indicate comparable binding free energies (: and ), consistent with mechanistically distinct yet similarly potent inhibition. Density functional theory characterizes electronic features and reactivity, and ADMET profiling indicates favorable drug‐like properties with low predicted toxicity. Overall, CNP0492204.2 and CNP0385629.2 emerge as potential MDM2 inhibitors that merit in vitro and in vivo validation and early preclinical development studies.