Improving the binding affinity estimations of protein-ligand complexes using machine-learning facilitated force field...

Area of Science:

• Computational chemistry
• Structural biology
• Drug discovery

Background:

• Scoring functions are crucial for predicting protein-ligand (PL) complex formation in structure-based drug design.
• Existing scoring functions face challenges in accurately assessing binding affinities for both metallo and non-metallo PL complexes.

Purpose of the Study:

• To introduce Bappl+, a novel scoring function designed to predict binding affinities for both metallo and non-metallo protein-ligand complexes.
• To evaluate the performance of Bappl+ against state-of-the-art scoring functions.

Main Methods:

• Development of the Bappl+ scoring function incorporating a machine-learning model.
• Training Bappl+ on an enlarged dataset of protein-ligand interactions.
• Validation of Bappl+ performance using Pearson correlation coefficient and evaluation on target-specific proteins.

Main Results:

• Bappl+ achieved a high Pearson correlation coefficient of approximately 0.76, outperforming existing scoring functions.
• The use of a machine-learning model and an expanded training dataset significantly contributed to Bappl+'s improved accuracy.
• Performance evaluation on target-specific proteins identified limitations and areas for future refinement.

Conclusions:

• Bappl+ demonstrates superior performance in predicting binding affinities for metallo and non-metallo protein-ligand complexes.
• The Bappl+ methodology offers a valuable tool for ranking candidate molecules in the drug discovery process.
• Further improvements can be achieved by addressing identified limitations and expanding the training data.