De novo design with deep generative models based on 3D similarity scoring

Area of Science:

• Computational Chemistry
• Drug Discovery
• cheminformatics

Background:

• Ligand-based drug design often relies on 2D similarity metrics.
• Exploring novel chemical space for drug targets like Dopamine receptor type 2 (DRD2) remains a challenge.

Purpose of the Study:

• To evaluate a deep generative model incorporating 3D shape and pharmacophore similarity for molecular design.
• To assess the model's capability in discovering new lead compounds and enabling scaffold hopping.

Main Methods:

• A deep generative model was trained using reinforcement learning with 3D shape and pharmacophore similarity scoring.
• The model was applied in a retrospective study using Dopamine receptor type 2 (DRD2) and haloperidol as a starting point.
• Generated outputs were compared against 2D Quantitative Structure-Activity Relationship (QSAR) based models.

Main Results:

• The 3D similarity-enabled generative model successfully discovered novel leads without additional information.
• The model demonstrated efficiency in scaffold hopping and generating diverse novel series.
• Orthogonality and greater diversity were observed in outputs from 3D similarity models compared to 2D QSAR models.

Conclusions:

• 3D shape and pharmacophore similarity are valuable components for deep generative models in molecular design.
• Combining 3D and 2D scoring components leads to more efficient exploration of desirable chemical space.
• This approach offers a powerful tool for discovering novel drug candidates and exploring new chemical entities.