Directional ΔG Neural Network (DrΔG-Net): A Modular Neural Network Approach to Binding Free Energy Prediction.

Area of Science:

• Computational chemistry
• Structural biology
• Drug discovery

Background:

• Protein-ligand binding free energy is crucial for drug discovery.
• Alchemical methods are accurate but computationally expensive.
• Data-driven methods are fast but lack transferability.

Purpose of the Study:

• Introduce DrΔG-Net (Dragnet), an equivariant graph neural network.
• Blend ligand-based and protein-ligand data-driven approaches.
• Develop a global scoring function for binding affinity prediction.

Main Methods:

• Utilize a 3D fingerprint representation of ligands and protein complexes.
• Employ an equivariant graph neural network architecture.
• Apply transfer learning for system-specific optimization.

Main Results:

• Dragnet predicts binding affinity for arbitrary protein-ligand complexes.
• Achieves performance comparable to 2D ligand-based methods via transfer learning.
• Demonstrates that a few experimental affinities can optimize Dragnet for specific scaffolds.

Conclusions:

• Dragnet offers a hybrid approach balancing speed and accuracy in binding affinity prediction.
• Transfer learning with minimal data or physics-based predictions can tune Dragnet effectively.
• Highlights current limitations in data-driven binding free energy modeling.