Robust scoring functions for protein-ligand interactions with quantum chemical charge models

Area of Science:

• Computational Chemistry
• Structural Biology
• Drug Discovery

Background:

• Ordinary least-squares (OLS) regression is commonly used for protein-ligand scoring functions but is sensitive to outliers.
• Accurate atomic charge determination is crucial for modeling electrostatic interactions in biomolecular association.
• The AutoDock4 scoring function previously used OLS and a basic Gasteiger charge method.

Purpose of the Study:

• To investigate if more rigorous quantum chemical charge models can enhance the statistical performance of AutoDock4 scoring functions.
• To compare the impact of different atomic charge models on protein-ligand free energy regression.
• To develop improved scoring functions for protein-ligand interactions.

Main Methods:

• Employed quantum chemical methods: restrained electrostatic potential (RESP) and Austin-model 1-bond charge correction (AM1-BCC) for atomic partial charges.
• Utilized robust regression analysis and outlier exclusion techniques.
• Developed a new protein-ligand free energy regression model using AM1-BCC charges for ligands and Amber99SB for proteins.

Main Results:

• Achieved a root-mean-squared error of 1.637 kcal/mol on the training set and 2.176 kcal/mol on an external test set.
• Demonstrated an 87% success rate in binding pose prediction for external decoy sets (RMSD < 2 Å).
• The new scoring functions showed weak class-dependency (hydrophobic, hydrophilic, mixed).

Conclusions:

• More rigorous charge models, specifically AM1-BCC, significantly improve AutoDock4 scoring function performance.
• Robust regression and outlier exclusion are vital for developing accurate and reliable scoring functions.
• The developed scoring functions offer high accuracy in both binding free energy prediction and pose prediction.