A geometry-aware generative framework integrating GPS-VAE and Transformer-SELFIES for structure-based de novo drug...

Area of Science:

• Computational chemistry
• Drug discovery
• Bioinformatics

Background:

• Structure-based de novo drug design (SBDD) aims to create novel ligands for specific protein targets.
• Current deep generative models face challenges in capturing protein topology, ensuring chemical validity, and optimizing multiple objectives.
• Janus Kinase 2 (JAK2) and Dopamine D2 Receptor (DRD2) are key targets in drug development.

Purpose of the Study:

• To develop an advanced generative framework for de novo drug design.
• To overcome limitations of existing deep generative models in SBDD.
• To generate chemically valid and high-affinity ligands for JAK2 and DRD2.

Main Methods:

• Developed a graph interaction transformer variational autoencoder (GPS-VAE) for feature extraction.
• Utilized Transformer-SELFIES to ensure 100% chemical validity of generated molecules.
• Integrated a variational projection network and STONED evolutionary algorithm for refinement.

Main Results:

• The GPS-VAE framework accurately captured geometric dependencies and topological features of protein binding sites.
• Generated fragment-like and lead-like scaffolds with high predicted ligand efficiency (>0.5).
• Discovered novel macrocyclic structures targeting JAK2, demonstrating the framework's efficacy.

Conclusions:

• High-quality data representation and evolutionary search significantly improve de novo drug design efficiency.
• The proposed framework offers a robust solution for generating valid and effective drug candidates.
• This approach advances the field of structure-based drug design by addressing key computational challenges.