Carpio_2025_Int.J.Mol.Sci_26_5579

Type 2 diabetes mellitus (T2DM) is a complex and prevalent metabolic disorder, and dipeptidyl peptidase 4 (DPP4) inhibitors have proven effective, yet the identification of novel inhibitors remains challenging due to the vastness of chemical space. In this study, we developed DPPPRED-IV, a web-based ensembled system integrating both binary classification and continuous regression Quantitative Structure Activity Relationships (QSAR) models to predict human DPP4 inhibitory activity. A curated dataset of 4 676 ChEMBL compounds was subjected to genetic algorithm descriptor selection and multiple machine learning algorithms; classification models were combined via a soft voting ensemble, while regression models estimated IC(50) values. All models underwent external 10-fold cross-validation and applicability domain analysis. The final models were integrated into a user-friendly web server, allowing predictions from SMILES inputs. Experimental testing of 29 MolPort compounds at 1.5 microM confirmed that 14 predicted actives exhibited significant inhibition, supporting the tool's performance in early-stage screening. DPPPRED IV is freely available within the ChemoPredictionSuite and offers a resource to accelerate decision making, reduce costs and minimize animal use in T2DM drug discovery.