Graph-pKa

Graph-pKa predicts macro- and micro-pKa values of chemical compounds using graph neural networks to characterize ionization properties at atomic and molecular levels.

Key Features:

• Graph Neural Networks (GNNs): Graph neural networks are used to model molecular structure and capture intramolecular relationships that influence ionization.
• Large-Scale Dataset: The model is trained on a dataset comprising 16,595 compounds with 17,489 pKa values.
• Macro-pKa Prediction Accuracy: On testing datasets, Graph-pKa reports a mean absolute error (MAE) of approximately 0.55 and a coefficient of determination (R²) of approximately 0.92 for macro-pKa predictions.
• Multi-Instance Learning for Micro-pKa Deconvolution: Multi-instance learning is applied to deconvolute predicted macro-pKa values into discrete micro-pKa values at the atomic level.

Scientific Applications:

• Drug discovery: Predicts ionization behavior of pharmaceutical candidates to inform solubility, absorption, and distribution properties.
• Environmental chemistry: Predicts ionization behavior of pollutants to inform their fate and interactions in natural systems.
• Materials science: Informs design of compounds with target chemical reactivity or stability through accurate pKa estimation.

Methodology:

Graph-pKa implements graph neural networks trained on a dataset of 16,595 compounds with 17,489 pKa values and employs multi-instance learning to deconvolute macro-pKa predictions into micro-pKa assignments.