Multi-instance learning of graph neural networks for aqueous pKa prediction

Area of Science:

• Computational chemistry
• Cheminformatics

Background:

• The acid dissociation constant (pKa) is vital for understanding chemical ionization.
• Experimental pKa determination is complex and time-consuming, particularly for micro-pKa values.

Purpose of the Study:

• To develop a fast and accurate computational method for predicting pKa values.
• To enable the prediction of both macro-pKa and micro-pKa values.

Main Methods:

• Compiled a large dataset of 16,595 compounds with 17,489 pKa values.
• Developed a novel pKa prediction model, Graph-pKa, utilizing graph neural networks.
• Integrated multi-instance learning for automatic deconvolution of macro-pKa into micro-pKa values.

Main Results:

• Graph-pKa achieved high accuracy in macro-pKa prediction (MAE ~0.55, R² ~0.92) on a test dataset.
• The model successfully deconvoluted macro-pKa into discrete micro-pKa values.
• The developed model demonstrates significant potential for accelerating chemical research.

Conclusions:

• Graph-pKa offers a robust and efficient solution for pKa prediction.
• The model's ability to predict micro-pKa values provides atomic-level insights.
• The Graph-pKa tool is publicly accessible via a web interface for broader scientific use.