Structure-Based Drug Design with a Deep Hierarchical Generative Model.

Area of Science:

• Computational Chemistry
• Drug Discovery
• Machine Learning

Background:

• Expanding chemical libraries and improved virtual screening methods impact early drug design.
• Scalability limitations persist in screening-based methods due to computational constraints and vast chemical space.
• Machine learning models learn drug-target relationships from data to overcome these limitations.

Purpose of the Study:

• Introduce DrugHIVE, a deep hierarchical variational autoencoder for enhanced molecular generation.
• Demonstrate DrugHIVE's superior speed and performance compared to state-of-the-art generative models.
• Highlight DrugHIVE's applicability to a broad range of drug design tasks and targets.

Main Methods:

• Developed DrugHIVE, a deep hierarchical variational autoencoder architecture.
• Evaluated DrugHIVE on standard generative benchmarks against autoregressive and diffusion models.
• Assessed DrugHIVE's performance in virtual screening efficiency and various drug design tasks.

Main Results:

• DrugHIVE outperformed state-of-the-art methods in speed and performance on generative benchmarks.
• The hierarchical design provided improved control over molecular generation.
• The method demonstrated scalability and applicability to AlphaFold-predicted structures.

Conclusions:

• DrugHIVE significantly accelerates drug design by enabling efficient, controlled molecular generation.
• The approach enhances virtual screening and supports diverse tasks like de novo design and scaffold hopping.
• DrugHIVE extends high-quality drug-like molecule generation to a majority of the human proteome, including previously intractable targets.