Learning protein-ligand binding affinity with atomic environment vectors.

Area of Science:

• Computational chemistry
• Drug discovery
• Machine learning in cheminformatics

Background:

• Protein-ligand binding affinity prediction is crucial for drug discovery.
• Machine learning and deep learning methods are advancing scoring function performance.
• Classical scoring functions have limitations in accuracy and scope.

Purpose of the Study:

• To develop and evaluate an Atomic Environment Vector (AEV)-based scoring function (AEScore) for protein-ligand binding affinity prediction.
• To assess AEScore's performance on benchmark datasets for binding affinity, docking, and virtual screening.
• To explore hybrid approaches combining AEScore with classical scoring functions.

Main Methods:

• Utilized Atomic Environment Vectors (AEVs) and feed-forward neural networks to create AEScore.
• Evaluated AEScore on the CASF-2016 benchmark dataset for binding affinity prediction.
• Investigated combining AEScore with AutoDock Vina using a transfer learning approach.

Main Results:

• AEScore achieved an RMSE of 1.22 pK units and a Pearson's correlation coefficient of 0.83 on the CASF-2016 benchmark, comparable to state-of-the-art methods.
• AEScore showed limitations in docking and virtual screening tasks without explicit training.
• The hybrid [Formula: see text]-AEScore model demonstrated an RMSE of 1.32 pK units and a Pearson's correlation coefficient of 0.80, retaining docking and screening capabilities.

Conclusions:

• AEV-based neural networks offer a competitive approach for predicting protein-ligand binding affinity.
• Hybrid models integrating novel machine learning scoring functions with classical methods can enhance overall performance.
• Further development is needed to optimize AEScore for docking and virtual screening applications.