Deep learning and molecular dynamics reveal promising EZH2 inhibitors for epigenetic cancer targeting

Fine-tuned REINVENT generative model integrated with structure- and ligand-based computational approaches was applied to identify novel EZH2 inhibitors. The generated molecules closely reproduced the structural and physicochemical property distributions of ChEMBL compounds, confirming their relevance to drug-like chemical space. Classification performance was robust (ROC-AUC ≥ 0.87 across training, test, and validation sets), enabling the prioritization of 511 PAINS-filtered compounds for structure-based screening. Molecular docking identified four lead compounds (161, 225, 234, and 383) with binding affinities (-5.616 to -6.403 kcal/mol) stronger than the reference inhibitor Tazemetostat (-4.481 kcal/mol). Hydrogen bonding analyses and MM/GBSA free energy estimates further supported the stability of these complexes. QSAR modeling (R² = 0.7233, Q² = 0.6901) and DFT-derived descriptors provided mechanistic insights into electronic structures, donor-acceptor characteristics, and chemical reactivity. Pharmacokinetic predictions indicated that compounds 234 and 225 possessed superior oral absorption, permeability, and reduced cardiotoxic liability compared with Tazemetostat, while all hits satisfied Lipinski's rule of five. Molecular dynamics simulations confirmed that ligand binding stabilized the protein structure, with compound 383 achieving the lowest RMSD (3.58 Å) and RMSF (1.68 Å), outperforming the reference drug. Overall, compounds 234, 225, and 383 emerged as the most promising scaffolds: compound 234 for oral bioavailability, compound 225 for high absorption and safety, and compound 383 for conformational stability. These findings highlight the power of combining deep generative modeling with computational drug design to accelerate EZH2 inhibitor discovery.