MolGpka: A Web Server for Small Molecule pK_a Prediction Using a Graph-Convolutional Neural Network.

Area of Science:

• Computational chemistry
• Medicinal chemistry
• Machine learning

Background:

• The ionization state of a molecule, determined by its pKa, significantly impacts biological activity, solubility, membrane permeability, metabolism, and toxicity.
• Accurate and rapid pKa estimation is essential for efficient drug discovery and lead optimization.

Purpose of the Study:

• To develop and present MolGpKa, a novel web server for predicting the pKa of small molecules.
• To leverage a graph-convolutional neural network (GCNN) for automated learning of chemical patterns relevant to pKa prediction.

Main Methods:

• A GCNN model was trained using a large dataset of 1.6 million compounds from the ChEMBL database, with pKa values sourced from ACD/pKa.
• The model learns chemical features directly from molecular structures, eliminating the need for human-engineered descriptors.

Main Results:

• The MolGpKa model demonstrated superior performance compared to traditional machine learning models that rely on human-engineered fingerprints.
• The GCNN effectively captured and learned the complex substitution effects influencing molecular pKa values.

Conclusions:

• MolGpKa provides a valuable and efficient tool for the rapid estimation of pKa during the ligand design process.
• The MolGpKa web server is freely accessible to researchers, facilitating its use in drug discovery workflows.