General Binding Affinity Guidance for Diffusion Models in Structure-Based Drug Design.

Area of Science:

• Computational chemistry
• Drug discovery
• Artificial intelligence in medicine

Background:

• Structure-based drug design (SBDD) utilizes computational methods to create molecules targeting specific proteins.
• Diffusion models show promise in SBDD but often lack precise control over binding affinity.
• Existing models may not sufficiently prioritize binding affinity during ligand generation.

Purpose of the Study:

• To introduce BADGER, a novel binding-affinity guidance framework for diffusion models in SBDD.
• To enhance the control and effectiveness of diffusion models in generating high-affinity ligands.
• To enable the design of drug candidates with improved binding characteristics.

Main Methods:

• BADGER employs two strategies: classifier guidance using gradient-based affinity signals and classifier-free guidance integrating affinity conditioning into training.
• The framework is designed as a plug-and-play module for existing diffusion models.
• Extended to multiconstraint guidance optimizing for binding affinity, drug-likeness (QED), and synthetic accessibility (SA).

Main Results:

• BADGER demonstrated up to a 60% improvement in ligand-protein binding affinity compared to previous methods.
• Achieved controllable ligand generation guided by binding affinity.
• Successfully designed realistic and synthesizable drug candidates by optimizing multiple constraints.

Conclusions:

• BADGER significantly advances diffusion model capabilities in structure-based drug design.
• The framework offers a powerful tool for generating potent and specific drug candidates.
• Multiconstraint optimization enables the design of drug-like and synthesizable molecules with high binding affinity.