Machine Learning Methods as a Cost-Effective Alternative to Physics-Based Binding Free Energy Calculations.

Area of Science:

• Computational chemistry
• Drug discovery
• Machine learning in pharmacology

Background:

• Accurate prediction of protein-ligand binding affinity is crucial for drug discovery.
• Physics-based methods face challenges with forcefield accuracy and conformational sampling.
• Machine learning (ML) methods are emerging as powerful tools for binding affinity prediction.

Purpose of the Study:

• To retrospectively evaluate and compare the performance of state-of-the-art physics-based and machine learning methods for binding free energy calculations.
• To assess the strengths and limitations of different computational approaches across diverse protein targets and ligand series.

Main Methods:

• Retrospective binding free energy evaluations were performed on 172 compounds across four protein targets and five congeneric ligand series.
• Comparison included various physics-based methods with differing complexity and sampling, alongside available state-of-the-art machine learning models.
• Performance was assessed based on the accuracy of rank ordering ligands and predicting binding affinities.

Main Results:

• Physics-based methods demonstrated strong performance when ligand perturbations involved the solvation region.
• Physics-based methods showed limitations in accurately capturing large conformational changes within protein active sites.
• Machine learning-based methods provided a cost-effective alternative, but their predictive accuracy was contingent on the quality and quantity of available training data.

Conclusions:

• Both physics-based and machine learning methods have distinct advantages and limitations in predicting protein-ligand binding free energies.
• The choice of method should consider the specific characteristics of the system, such as the nature of ligand perturbations and conformational flexibility.
• Machine learning models show promise as efficient tools, emphasizing the need for robust experimental datasets for reliable model training and application in drug discovery.