Three descriptor model sets a high standard for the CSAR-NRC HiQ benchmark

Area of Science:

• Computational chemistry
• Cheminformatics
• Drug discovery

Background:

• Predicting ligand binding free energies is crucial for drug discovery.
• Existing methods require complex models and extensive data.
• The CSAR-NRC HiQ benchmark dataset provides a standardized testbed for evaluating prediction models.

Purpose of the Study:

• To develop a robust and interpretable model for predicting ligand binding free energies.
• To identify key molecular descriptors that govern binding affinity.
• To validate a proteochemometric approach using a stepwise multiple-linear regression (MLR) model.

Main Methods:

• Utilized a proteochemometric approach with distance-dependent atom-type pair descriptors.
• Employed a bagged stepwise multiple-linear regression (MLR) model with complexity reduction.
• Evaluated model performance on the CSAR-NRC HiQ benchmark dataset using cross-validation (R(2)(cv) = 0.55, MUE(cv) = 1.19, RMSE(cv) = 1.49).

Main Results:

• Identified three significant descriptors for predicting binding free energies.
• Descriptor 1: Count of protein atoms (4.5-6 Å shell around ligand heavy atoms, excluding O/P).
• Descriptor 2: Count of sulfur atoms near tryptophan.
• Descriptor 3: Count of aliphatic ligand hydroxy hydrogens.
• Descriptors 1 and 2 positively correlate with binding energy; descriptor 3 negatively correlates.

Conclusions:

• A simple, robust MLR model can effectively predict ligand binding free energies.
• The identified descriptors offer valuable insights into the physicochemical basis of ligand-protein interactions.
• The findings suggest that proteochemometrics can be a powerful tool for accelerating drug discovery efforts.