Exploring protein-ligand binding affinity prediction with electron density-based geometric deep learning.

Area of Science:

• Computational chemistry
• Structural biology
• Drug discovery

Background:

• Structure-based drug design requires accurate prediction of protein-ligand binding affinity.
• Current deep learning methods may not fully capture physical interactions or can be biased.
• Electron density offers a fundamental physical representation of molecular interactions.

Purpose of the Study:

• To investigate the utility of bond-critical points from electron density for predicting binding affinity.
• To benchmark a geometric deep learning model using these points against existing methods.
• To critically analyze the role of electron density in deep learning for drug design.

Main Methods:

• Utilized a geometric deep learning model.
• Incorporated bond-critical points derived from electron density of protein-ligand complexes.
• Evaluated model performance on PDBbind and PDE10A datasets.
• Analyzed the correlation between electron density and binding affinity.

Main Results:

• Models achieved root-mean-squared errors of 1.4-1.8 log units (PDBbind) and 1.0-1.7 log units (PDE10A).
• Performance was comparable to benchmark methods, not showing significant advantages.
• Pearson correlation coefficients (r > 0.7) were observed between electron density and binding affinity for some targets.
• The utility of electron density for deep learning models was found to be context-dependent.

Conclusions:

• Electron density bond-critical points offer a physically grounded approach for protein-ligand interaction analysis.
• The direct application in deep learning for binding affinity prediction showed context-dependent utility.
• Further research is needed to optimize the integration of electron density features into computational drug design models.