CNSGT: Generative Transformer for De Novo Drug Design Targeting the Central Nervous System.

Abstract

The design of novel central nervous system (CNS) drugs presents formidable challenges due to the restrictive nature of the blood-brain barrier, which imposes stringent physicochemical requirements. Recent advances in deep learning, particularly Transformer-based architectures, have shown great potential for de novo molecular design. In this study, we present CNSGT, a novel generative framework that integrates variational autoencoders (VAE) with self-attention mechanisms to address the complexity of CNS drug design. By overcoming the limitations of traditional SMILES-based representations, CNSGT effectively captures molecular structure and semantic relationships. The model is pretrained on large-scale molecular data sets and fine-tuned via transfer learning for target-specific generation, demonstrated on dopamine transporter (DAT) inhibitors. The results show that CNSGT generates chemically valid molecules with high CNS drug-likeness (CNS MPO score >4) and improved synthetic accessibility (SAScore <3). The generated molecules also exhibit promising binding affinities in molecular docking (Glide docking score < -8 kcal/mol) and dynamic simulation studies with stable binding conformations. And theoretically prove their good synthetic accessibility through synthetic route analysis by medical chemists, suggesting the model's potential for expanding the useful chemical space and accelerating CNS drug discovery.