Physics-Guided Deep Generative Model for New Ligand Discovery.

Area of Science:

• Computational chemistry
• Molecular modeling
• Drug discovery

Background:

• Structure-based drug discovery seeks molecules targeting specific proteins.
• Deep learning models generate drug-like molecules but often neglect physical binding principles.
• Incorporating underlying physics is crucial for realistic molecular formation and binding.

Purpose of the Study:

• To develop a physics-guided deep generative model for novel ligand discovery.
• To condition molecule generation on binding site and physics-based binding mechanism features.
• To improve the accuracy and efficiency of drug discovery processes.

Main Methods:

• Developed a hybrid deep generative model integrating physics-based features.
• Conditioned the model on receptor binding sites and ligand-receptor interaction physics.
• Evaluated the model on large protein-ligand complexes and small host-guest systems.

Main Results:

• The physics-guided model generated stronger binders, with over 75% exceeding original ligand affinity.
• Achieved an average binding affinity (ΔG) improvement of 1.88 kcal/mol over state-of-the-art methods.
• Generated ligands exhibited more feasible conformations and orientations compared to traditional deep learning models.

Conclusions:

• Physics-guided deep learning offers a promising approach for enhanced ligand discovery.
• The model demonstrates superior performance in predicting high-affinity binders.
• Future work includes expanding datasets and incorporating more biophysical insights.