PARP1 inhibitors discovery: innovative screening strategies incorporating machine learning and fragment replacement techniques.

Abstract

PARP1, the most prominent member of the PARP family, mediates DNA repair and cellular stress responses. PARP inhibitors (PARPi) show clinical promise in treating BRCA1/2-mutated or homologous recombination-deficient tumors, particularly in breast and ovarian cancers. However, acquired resistance remains a significant therapeutic challenge. This study developed a PARP1 inhibitor discovery pipeline integrating machine learning with conventional virtual screening methods. We introduced a novel strategy called fragment replacement to generate new compounds with optimized properties. Using the Maybridge compound library, we developed machine learning models to predict inhibitor activity. The Random Forest classifier demonstrated superior performance (AUC = 0.971, accuracy = 0.915) in tenfold cross-validation. This machine learning-driven approach outperformed conventional virtual screening in terms of efficiency. Subsequently, we conducted virtual screening using 2D fingerprints, shapes, and docking to retain the top-ranked ligands based on a standardized score (Z2-score). XP docking and ADMET prediction were used to select two molecules with strong drug-like properties. Fragment replacement was employed to reconstruct 101 new compounds with improved drug-like characteristics and increased activity. After validation, we identified three hits with docking scores between - 11.802kcal/mol and - 10.808kcal/mol, which were superior to the positive control Talazoparib (docking score: - 9.103kcal/mol). MD simulations assessed the binding stability of the compounds to proteins, with all three selected compounds exhibiting good binding stability, thus identifying them as potential candidates for development as PARP1 inhibitors.