Interaction-constrained 3D molecular generation using a diffusion model enables structure-based pharmacophore...

Area of Science:

• Computational chemistry
• Drug discovery
• Artificial intelligence in medicine

Background:

• Structure-based drug design faces challenges in generating molecules that maintain critical protein-ligand interactions.
• Existing methods often struggle to balance molecular generation with specific interaction constraints.

Purpose of the Study:

• To introduce DiffPharma, a conditional diffusion model framework for structure-based pharmacophore modeling.
• To generate novel 3D molecules that satisfy predefined protein-ligand interaction requirements.

Main Methods:

• Utilized a conditional diffusion model integrated with a semantic fusion architecture.
• Incorporated multiple interaction-specific neural networks to capture diverse molecular interactions (e.g., hydrogen bonds, hydrophobic interactions).
• Validated through ligand generation for AKT kinase 1, serine β-lactamase, and SARS-CoV-2 main protease.

Main Results:

• Achieved up to 0.9 residue-level interaction similarity, outperforming baseline models.
• Generated ligands for target proteins successfully preserved key interaction features.
• Molecular dynamics simulations and MM/GBSA calculations indicated structural stability and favorable binding tendencies for generated molecules compared to a reference ligand.

Conclusions:

• DiffPharma effectively generates 3D molecules that preserve essential protein-ligand interactions for drug design.
• The framework demonstrates generalizability across different protein targets.
• The study provides a practical, AI-driven solution for structure-based drug discovery with open-source implementation.