Machine-Learning- and Knowledge-Based Scoring Functions Incorporating Ligand and Protein Fingerprints

Area of Science:

• Computational chemistry
• Drug discovery
• Machine learning

Background:

• Molecular docking is crucial for structure-based drug design (SBDD), aiding rational drug discovery.
• Current docking methods accurately predict ligand binding structures but struggle with precise binding affinity prediction.
• Accurate binding affinity prediction is essential for reducing drug discovery costs.

Purpose of the Study:

• To develop a novel machine learning-based scoring function for improved drug discovery.
• To integrate ligand and protein structural information into a knowledge-based potential of mean force (PMF) score.
• To enhance the accuracy of protein-ligand binding affinity prediction.

Main Methods:

• Developed a new scoring function incorporating ligand and protein structural fingerprints as descriptors within a PMF score.
• Employed machine learning techniques, specifically least absolute shrinkage and selection operator (LASSO) and light gradient boosting machine (LightGBM), for regression analysis.
• Validated the scoring function using a test dataset to assess prediction accuracy against experimental values.

Main Results:

• The novel scoring function achieved a Pearson correlation coefficient of 0.79 for binding affinity prediction against experimental values.
• This performance surpasses that of conventional scoring functions used in molecular docking.
• Identified key descriptors contributing to the improved prediction accuracy, offering chemical insights.

Conclusions:

• The developed machine learning-based scoring function significantly enhances the prediction of protein-ligand binding affinity.
• This approach offers a more accurate and efficient tool for rational drug discovery.
• The findings provide valuable insights for future development of computational drug design strategies.