PIGNet: a physics-informed deep learning model toward generalized drug-target interaction predictions.

Area of Science:

• Computational chemistry
• cheminformatics
• Artificial intelligence in drug discovery

Background:

• Deep neural network (DNN) models show promise for drug-target interaction (DTI) prediction due to high accuracy and low computational cost.
• A significant challenge for current DNN-based DTI models is their limited generalization capability in practical drug discovery applications.

Purpose of the Study:

• To enhance the generalization of deep neural network models for drug-target interaction prediction.
• To develop a physics-informed deep learning model for more accurate and interpretable binding affinity prediction.

Main Methods:

• Developed PIGNet, a novel deep neural network model that predicts atom-atom pairwise interactions using physics-informed equations.
• Parameterized neural networks with physics-informed equations to calculate total binding affinity as the sum of pairwise interactions.
• Augmented training data with a wider variety of binding poses and ligands to improve model generalization.

Main Results:

• PIGNet demonstrated superior docking and screening performance compared to existing methods on the CASF 2016 benchmark dataset.
• The physics-informed approach enabled the interpretation of predicted binding affinities by visualizing the contributions of ligand substructures.
• The model's ability to provide insights for ligand optimization was highlighted.

Conclusions:

• PIGNet significantly improves generalization in DTI prediction models through a physics-informed strategy.
• The model offers interpretable predictions, aiding in the rational design and optimization of drug candidates.
• This approach advances the field of in silico drug discovery by enhancing both accuracy and interpretability.