GB-Score

GB-Score predicts ligand–protein binding affinity using Gradient Boosting Trees applied to distance-weighted interatomic contact features for structure-based scoring in drug discovery.

Key Features:

• Algorithmic Foundation: GB-Score employs the Gradient Boosting Trees algorithm for regression-based prediction of binding affinities.
• Featurization Methodology: It uses distance-weighted interatomic contact featurization representing distances between ligand and protein atom types.
• Training and Validation: The model is trained on PDBbind 2019v general and refined sets excluding the core set, with the PDBbind core set serving as an independent test set.
• Performance Metrics: On the CASF-2016 benchmark test GB-Score attains a Pearson correlation coefficient of 0.862 and an RMSE of 1.190 for scoring power.

Scientific Applications:

• Structure-based drug design: Provides quantitative binding affinity predictions to support identification and prioritization of potential drug candidates.
• Docking scoring and ranking: Can be applied to score and rank docked ligand poses based on predicted binding affinity.
• Analysis of crystal complexes: Enables assessment of binding affinities for experimentally determined protein–ligand crystal structures.

Methodology:

Featurization via distance-weighted interatomic contacts and model training using Gradient Boosting Trees on PDBbind 2019v (general and refined sets excluding the core set); evaluation on the PDBbind core set and the CASF-2016 benchmark.