Structure-based de novo drug design using 3D deep generative models.

Area of Science:

• Computational chemistry
• Artificial intelligence in drug discovery
• Molecular modeling

Background:

• Deep generative models offer data-driven approaches for de novo molecule design.
• Generating 3D molecular structures within target binding sites remains a significant challenge.
• Existing methods often require extensive expert knowledge.

Purpose of the Study:

• To introduce DeepLigBuilder, a novel deep learning method for 3D de novo drug design within protein binding sites.
• To develop a robust model capable of generating chemically valid and drug-like 3D molecules.
• To enable structure-based drug design and lead optimization.

Main Methods:

• Development of Ligand Neural Network (L-Net), a graph generative model for 3D molecule generation.
• Integration of L-Net with Monte Carlo tree search for structure-based design.
• Training L-Net on drug-like compounds from ChEMBL for chemical and conformational validity.

Main Results:

• DeepLigBuilder successfully generated novel, drug-like compounds with high predicted affinity for SARS-CoV-2 main protease.
• The generated molecules exhibited binding features similar to known inhibitors.
• Demonstrated the model's capability for structure-based de novo drug design and lead optimization.

Conclusions:

• DeepLigBuilder represents a state-of-the-art approach for structure-based de novo drug design.
• The model merges deep generative capabilities with atomic-level interaction evaluation.
• The L-Net framework is adaptable for generating functional molecules with desired properties.