Generating 3D molecules conditional on receptor binding sites with deep generative models.

Area of Science:

• Computational chemistry
• Structural biology
• Artificial intelligence in drug discovery

Background:

• Structure-based drug discovery aims to identify small molecules that bind to target proteins.
• Deep learning has generated molecules with desired properties but not yet 3D structures conditioned on protein binding interactions.

Purpose of the Study:

• To develop and present the first deep learning system for generating 3D molecular structures conditioned on a receptor binding site.
• To enable the sampling of conditional distributions of protein-ligand binding interactions for molecule generation.

Main Methods:

• Utilized a conditional variational autoencoder trained on atomic density grids of protein-ligand complexes.
• Employed atom fitting and bond inference to create valid 3D molecular conformations from generated densities.

Main Results:

• Generated 3D molecules whose properties significantly change when conditioned on mutated receptors.
• Demonstrated the ability to explore the generative model's latent space through sampling and interpolation.

Conclusions:

• This work presents a groundbreaking deep learning system for de novo 3D molecule generation based on protein targets.
• It paves the way for end-to-end prediction of stable, bioactive molecules using deep learning and protein structural information.