Spatio-temporal learning from molecular dynamics simulations for protein-ligand binding affinity prediction.

Area of Science:

• Computational chemistry
• Structural biology
• Machine learning

Background:

• Protein-ligand binding affinity prediction is challenging due to limited, biased training data for deep learning (DL) models.
• Existing DL models struggle with generalization, often showing poor performance on unbiased datasets.
• Protein-ligand interactions are time-dependent, making molecular dynamics (MD) simulations a promising approach for richer data.

Purpose of the Study:

• To develop a novel dataset and machine learning models for accurate protein-ligand binding affinity prediction.
• To address the limitations of current DL models in generalization and bias.
• To integrate molecular dynamics (MD) simulations into the prediction pipeline.

Main Methods:

• Created MDbind, a dataset of 63,000 protein-ligand interaction simulations.
• Developed novel neural networks trained on MD simulation data.
• Utilized MD simulations as a data augmentation technique for model training.

Main Results:

• Achieved state-of-the-art performance on benchmark datasets (PDBbind v.2016, FEP dataset).
• Demonstrated reduced prediction bias when models were trained on full MD simulations.
• Showcased the effectiveness of MD simulations in improving binding affinity prediction accuracy.

Conclusions:

• MDbind dataset and associated neural networks significantly advance protein-ligand binding affinity prediction.
• Integrating MD simulations enhances model generalization and reduces bias.
• This approach offers a more robust method for understanding and predicting molecular interactions.