Deep generative models for 3D molecular structure

Area of Science:

• Computational chemistry
• Drug discovery
• Artificial intelligence in chemistry

Background:

• Deep generative models are increasingly popular for chemical design.
• Historically, these models operated in 2D, but explicit 3D molecular generation is a recent focus.
• Dozens of 3D generative models have been developed, outputting atom types and coordinates.

Purpose of the Study:

• To review and discuss explicit 3D molecular generative models for chemical design.
• To highlight the capabilities and limitations of current 3D generative models.
• To propose a unified benchmark for evaluating generation and suggest future research directions.

Main Methods:

• Review of published explicit 3D molecular generative models.
• Analysis of model outputs, including atom types and coordinates.
• Evaluation of model performance in terms of docking scores, efficiency, and conformational realism.

Main Results:

• 3D generative models can generate molecules with docking scores comparable to known actives when guided by structural information.
• These models exhibit lower computational efficiency and generation throughput compared to 1D/2D models.
• Current 3D models may produce unrealistic molecular conformations.

Conclusions:

• Explicit 3D generative models show significant potential for structure-based chemical design.
• Further research is needed to improve computational efficiency and conformational accuracy.
• A unified benchmark is crucial for standardized evaluation and advancement of these models.