RELATION: A Deep Generative Model for Structure-Based De Novo Drug Design.

Area of Science:

• Computational chemistry
• Drug discovery
• Artificial intelligence in medicine

Background:

• Deep learning (DL) models are increasingly used for de novo molecular design.
• Existing DL models primarily focus on ligands, underutilizing target binding pocket 3D geometry.
• Integrating 3D structural information is crucial for designing effective and targeted molecules.

Purpose of the Study:

• To develop a novel 3D-based generative model for de novo molecular design.
• To leverage protein-ligand complex geometry in the molecular generation process.
• To design novel molecules with specific geometric properties and pharmacophore features.

Main Methods:

• Proposed the RELATION (REpresentation LEarning and ANalysis) model for 3D-based molecular generation.
• Developed the BiTL (Binding-aware Transfer Learning) algorithm to encode geometric features into a latent space.
• Employed pharmacophore conditioning and docking-based Bayesian sampling for efficient chemical space navigation.

Main Results:

• The RELATION model successfully generated novel molecular inhibitors for AKT1 and CDK2 targets.
• Generated molecules exhibited favorable predicted binding affinity and essential pharmacophore features.
• The model demonstrated efficient navigation of chemical space guided by 3D structural information.

Conclusions:

• The RELATION model represents a significant advancement in 3D-based de novo molecular design.
• Incorporating binding pocket geometry enhances the generation of targeted and effective drug candidates.
• This approach holds promise for accelerating the discovery of novel therapeutics.