Machine Learning-Driven De Novo Design of DDP-4Targeted Drug Molecules Using RNN-Based GenerativeModels and Reinforcement Learning

We present a fully in silico pipeline for the de novo design of dipeptidyl peptidase-4 (DPP- 4) inhibitors that integrates data-driven curation, transfer learning, and reinforcement learning (RL) within the REINVENT architecture. Activity records for human DPP-4 (CHEMBL284) were programmatically retrieved from ChEMBL, normalized to nanomolar units, filtered at IC50100 nM, standardized to canonical SMILES, and consolidated into a high-quality training table pIC50 values were computed and a top-100 reference set was exported for down- stream novelty control. A REINVENT prior was adapted to the DPP-4 chemical space via maximum-likelihood fine-tuning on 173 nonredundant, high-activity SMILES. The adapted generator was then optimized with an RL objective that combined predicted potency (pIC50), drug-likeness (QED), synthetic accessibility (SA), and novelty penalties relative to the top-100 reference inhibitors. Relative to the transfer-learned baseline, RL increased mean QED by ~ 10%, improved normalized synthetic accessibility (1 - SA)/10 by ~ 15%, and maintained diversity with ~60% novelty, while the composite reward showed a clear upward shift. Structure-based evaluation further corroborated these gains: 100 RL-generated molecules achieved a mean docking score of -9.8 kcal/mol, surpassing both pre-RL de novo samples (-7.7 kcal/mol) and the top 100 reference actives (-8.5 kcal/mol). These results demonstrate that RL fine-tuning can steer a pretrained generator toward DPP-4relevant regions of chemical space with improved developability surrogates and predicted binding. Future work will integrate ADMET predictors into the reward and prioritize wet-lab validation to confirm biochemical potency and advance selected designs toward lead optimization.