SuperAtomicCharge

SuperAtomicCharge predicts quantum-mechanical partial charges of atoms in molecular structures using deep graph neural networks to provide RESP, DDEC4, and DDEC78 charge estimates for computational chemistry applications.

Key Features:

• Deep Graph Learning (GNNs): Implements deep graph neural networks that capture both 2D and 3D structural information of molecules to improve charge prediction accuracy.
• Predicted Charge Types: Produces RESP, DDEC4, and DDEC78 partial charges derived from high-level quantum-mechanical calculations.
• Transfer Learning Strategy: Incorporates a simple transfer learning approach to improve generalization across different molecular datasets.
• Multitask Self-Supervised Learning: Employs multitask learning with self-supervised descriptors to enable simultaneous prediction of multiple charge types.
• Performance: Outperforms existing baselines, including other GNN-based and machine learning predictors, across three external test sets.

Scientific Applications:

• Drug design and virtual screening: Provides QM-derived partial charges for use in structure-based virtual screening workflows.
• Scoring and screening robustness: RESP and DDEC4 charges predicted by the method demonstrate enhanced robustness in scoring and screening compared to commonly used partial charges.
• Computational chemistry applications: Supplies partial charge estimates relevant to a range of computational chemistry tasks that rely on QM charge information.

Methodology:

Uses deep graph neural networks with molecules represented as graphs (atoms as nodes, bonds as edges) that incorporate 2D and 3D structural information; applies transfer learning to adapt pre-trained models to new datasets; and employs multitask learning with self-supervised descriptors to simultaneously predict RESP, DDEC4, and DDEC78 charges.