Deep generative models for 3D structure-based drug design and molecular optimisation: a comprehensive survey

Area of Science:

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

Background:

• Structure-based drug design (SBDD) is crucial but challenged by diverse generative models.
• Lack of unified taxonomy and fragmented evaluation hinder progress in 3D deep generative models for drug design.
• Existing methods transition from diffusion models to flow matching and synthesizability-aware generation.

Purpose of the Study:

• To provide a comprehensive review and unified taxonomy of over 100 3D SBDD methods.
• To address evaluation inconsistencies and propose standardized geometric validity auditing.
• To identify key trends and open challenges in generative molecular design.

Main Methods:

• Proposed a unified taxonomy of four generative paradigms: autoregressive, diffusion, flow matching, and Bayesian Flow Networks.
• Systematically analyzed evaluation inconsistencies and introduced standardized geometric validity auditing.
• Identified four evolutionary trends: SE(3)-equivariant modeling, parallel generation, explicit conditioning, and synthesizability integration.

Main Results:

• A comprehensive review of >100 methods in 3D SBDD and molecular optimization.
• Standardized geometric validity auditing improves metric comparability.
• SE(3)-equivariant flow matching identified as a promising direction for foundation models.

Conclusions:

• A unified taxonomy and standardized evaluation framework are proposed for 3D SBDD.
• Key trends indicate a shift towards equivariant modeling, parallel generation, explicit conditioning, and integrated synthesizability.
• SE(3)-equivariant flow matching offers a promising balance of efficiency, fidelity, and multi-objective optimization for future drug design models.