Binding Affinity Prediction by Pairwise Function Based on Neural Network.

Area of Science:

• Computational chemistry
• Structural biology
• Drug discovery

Background:

• Estimating protein-ligand binding affinity is crucial for drug discovery.
• Accurate prediction of binding affinity guides the design of novel therapeutics.
• Existing scoring functions often struggle with precise affinity predictions.

Purpose of the Study:

• To introduce a new computational method for estimating binding affinity.
• To leverage neural networks for predicting binding free energy from protein-ligand complex structures.
• To develop a more accurate and efficient scoring function for molecular interactions.

Main Methods:

• Developed a pairwise additive free-energy model for protein-ligand complexes.
• Utilized a neural network to calculate atom-pair contributions based on atomic properties and distance.
• Trained the neural network model on a subset of the PDBbind 2018 dataset.
• Validated the model using PDBbind 2018 and CASF-2016 benchmark datasets.

Main Results:

• The proposed model accurately predicts binding affinity.
• Achieved competitive performance compared to established scoring functions like AutoDock Vina.
• Demonstrated the effectiveness of the neural network-based pairwise contribution approach.
• The model shows good generalization capabilities on benchmark datasets.

Conclusions:

• The novel neural network approach provides a powerful tool for binding affinity estimation.
• This method offers improved accuracy and potential for broader application in drug design.
• The framework is extensible, allowing for future incorporation of additional factors to enhance predictive power.