Protein-ligand binding affinity prediction exploiting sequence constituent homology.

Area of Science:

• Computational chemistry and cheminformatics
• Machine learning applications in drug discovery

Background:

• Molecular docking is crucial for estimating binding affinities, with machine learning enhancing accuracy.
• Existing methods often rely on complex spatial and categorical information from protein-ligand complexes.
• Evaluation typically uses PDBbind datasets, notably the Comparative Assessment of Scoring Functions (CASF) benchmarks.

Purpose of the Study:

• To develop a simplified yet effective approach for binding affinity estimation.
• To demonstrate that minimal descriptors can accurately predict binding affinity.
• To validate the method on established benchmarks and datasets lacking spatial information.

Main Methods:

• Utilizing a small set of ligand and protein descriptors.
• Employing gradient boosting trees for affinity prediction.
• Testing on CASF 2007, 2013, and 2016 datasets, and a large ChEMBL-derived dataset.

Main Results:

• The developed method shows high performance on CASF datasets, excelling on CASF2016.
• Achieved superior performance compared to existing approaches on benchmark datasets.
• Demonstrated effectiveness even when the precise binding conformation is unknown.

Conclusions:

• A parsimonious set of descriptors combined with gradient boosting trees offers a powerful tool for binding affinity prediction.
• This approach simplifies affinity estimation and expands applicability to diverse datasets.
• The methodology provides a robust and efficient alternative for computational drug discovery.