De Novo Structure-Based Drug Design Using Deep Learning.

Area of Science:

• Computational Chemistry
• Drug Discovery
• Artificial Intelligence

Background:

• Ligand-based deep learning methods for de novo drug design require initial target-specific ligand datasets, which are often unavailable for novel proteins.
• Designing molecules against novel targets is hindered by the challenge of obtaining sufficient ligand data.

Purpose of the Study:

• To develop a deep learning-based method for de novo drug design that utilizes only the target protein's active site structure.
• To generate novel drug molecules by learning active site features and employing conditional generation and reinforcement learning.

Main Methods:

• A graph attention model was employed to learn amino acid features within protein active sites of known protein-ligand complexes.
• Learned active site features were integrated with a pretrained generative model for conditional molecule generation.
• A bioactivity prediction model, optimized via reinforcement learning, refined the conditional generative model.

Main Results:

• The method successfully generated molecules similar to known inhibitors for Janus kinase 2 (JAK2) and dopamine receptor D2 (DRD2).
• The graph attention model identified key active site residues influencing molecule generation.
• Generated molecules exhibited pharmacophoric features comparable to known inhibitors.

Conclusions:

• The proposed deep learning method effectively designs novel drug molecules using only target protein active site information.
• This approach addresses the data scarcity challenge in drug design against novel protein targets.
• The integration of graph attention, conditional generation, and reinforcement learning offers a promising strategy for targeted drug discovery.