ESP DNN

ESP DNN predicts high-quality electrostatic potential (ESP) surfaces for ligands and proteins using a graph-convolutional deep neural network trained on quantum mechanics (QM) ESP data.

Key Features:

• Rapid Generation: The model generates ESP surfaces in milliseconds, providing orders-of-magnitude speed improvement over high-level QM calculations.
• High-Quality Predictions: The graph-convolutional DNN was trained on approximately 100,000 molecules with ESP data from high-quality QM calculations to reproduce molecular electrostatic features.
• Compatibility and Integration: For ligands the tool produces PQR files from the DNN predictions, and for proteins it uses parametrized charges for amino acids to ensure compatibility between ligand and protein ESP representations.
• Correlation with Experimental Properties: Predicted ESP surfaces have been shown to correlate well with experimental properties relevant to medicinal chemistry.

Scientific Applications:

• Drug Discovery: Characterizing and optimizing electrostatic interactions between ligands and target proteins to inform structure-based drug design.
• Medicinal Chemistry: Guiding compound design by relating predicted ESP profiles to experimentally observed properties of small molecules.

Methodology:

The approach trains a graph-convolutional deep neural network on ESP surfaces derived from high-quality quantum mechanics (QM) calculations (approximately 100,000 molecules), uses the trained model to predict ligand ESP and output PQR files, applies parametrized amino-acid charges for protein ESP, and performs inference in milliseconds.