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

Area of Science:

• Medicinal Chemistry
• Computational Chemistry
• Artificial Intelligence in Drug Discovery

Background:

• Designing central nervous system (CNS) drugs is challenging due to the blood-brain barrier.
• Deep learning, especially Transformer models, shows promise for de novo molecular design.

Purpose of the Study:

• To present CNSGT, a novel generative framework for CNS drug design.
• To overcome limitations of traditional molecular representations like SMILES.
• To generate molecules targeting specific CNS targets, exemplified by dopamine transporter (DAT) inhibitors.

Main Methods:

• Developed CNSGT, integrating variational autoencoders (VAE) with self-attention mechanisms.
• Pretrained the model on large molecular datasets and fine-tuned using transfer learning.
• Evaluated generated molecules using CNS drug-likeness (MPO score), synthetic accessibility (SAScore), molecular docking, and dynamic simulations.

Main Results:

• CNSGT generated chemically valid molecules with high CNS drug-likeness (CNS MPO score >4).
• Generated molecules demonstrated improved synthetic accessibility (SAScore <3).
• Promising binding affinities (Glide docking score < -8 kcal/mol) and stable binding conformations were observed.

Conclusions:

• CNSGT effectively captures molecular structure and semantic relationships, overcoming SMILES limitations.
• The framework shows potential for expanding chemical space and accelerating CNS drug discovery.
• Generated molecules exhibit favorable properties for drug development, validated by theoretical synthetic route analysis.