NovoMol: Recurrent Neural Network for Orally Bioavailable Drug Design and Validation on PDGFRα Receptor

Longer timelines and lower success rates of drug candidates limit the productivity of clinical trials in the pharmaceutical industry. Promising de novo drug design techniques help solve this by exploring a broader chemical space, efficiently generating new molecules, and providing improved therapies. However, optimizing for molecular characteristics found in approved oral drugs remains a challenge, limiting de novo usage. In this work, we propose NovoMol, a novel de novo method using recurrent neural networks to mass-generate drug molecules with high oral bioavailability, increasing clinical trial time efficiency. Molecules were optimized for desirable traits and ranked using the quantitative estimate of drug-likeness (QED). Generated molecules meeting QED's oral bioavailability threshold were used to retrain the neural network, and, after five training cycles, 76% of generated molecules passed this strict threshold and 96% passed the traditionally used Lipinski's Rule of Five. The trained model was then used to generate specific drug candidates for the cancer-related PDGFR{\alpha} receptor and 44% of generated candidates had better binding affinity than the current state-of-the-art drug, Imatinib (with a receptor binding affinity of -9.4 kcal/mol), and the best-generated candidate at -12.9 kcal/mol. NovoMol provides a time/cost-efficient AI-based de novo method offering promising drug candidates for clinical trials.