In-Pocket 3D Graphs Enhance Ligand-Target Compatibility in Generative Small-Molecule Creation: A Dopamine D2 Receptor...

Area of Science:

• Computational chemistry and cheminformatics
• Artificial intelligence in drug discovery
• Structural biology and molecular modeling

Background:

• Structure-based drug discovery relies on protein-ligand complexes, but 3D information is often missing in generative models.
• Developing generative models that incorporate explicit 3D structural context is crucial for improving drug design.
• The dopamine D2 receptor (DD2R) serves as a relevant model system for evaluating protein-ligand interactions.

Purpose of the Study:

• To present a novel graph-based generative modeling technology encoding explicit 3D protein-ligand contacts.
• To evaluate the efficacy of this 3D generative approach using the dopamine D2 receptor (DD2R) as a model.
• To demonstrate how incorporating structural context enhances small molecule generation within a realistic binding environment.

Main Methods:

• Developed a graph-based generative model using a conditional variational autoencoder for activity-specific molecule generation.
• Integrated putative contact generation to predict molecular interactions within the target-binding pocket.
• Evaluated generated molecules based on docking scores, stereochemistry, and recoverability in chemical databases.

Main Results:

• Molecules generated using the 3D procedure showed higher compatibility with the DD2R binding pocket compared to a 2D method.
• The 3D method yielded better docking scores, expected stereochemistry, and higher recoverability in commercial databases.
• Predicted protein-ligand contacts were frequently among the highest-ranked docking poses with a high recovery rate.

Conclusions:

• The proposed 3D graph-based generative model effectively encodes protein-ligand contacts for enhanced drug discovery.
• Incorporating the structural context of a protein target significantly improves the generation of small molecules for realistic binding environments.
• This approach offers a promising direction for advancing structure-based drug design using deep learning.