pK_a Predicting Models Trained with a Tautomer-Compatible Graph Dataset with Quantum Chemical Features.

Area of Science:

• Computational chemistry
• Chemical informatics
• Drug discovery

Background:

• Accurate prediction of acidity constants (pKa) is crucial for various chemical and biological applications.
• Existing datasets and models for pKa prediction have limitations in scope and accuracy.
• Handling tautomeric forms in pKa prediction presents a significant challenge.

Purpose of the Study:

• To present G-pKa, a comprehensive database of experimental pKa values and associated quantum mechanical properties.
• To develop and evaluate novel pKa prediction models utilizing graph-based representations.
• To make the database and prediction scripts publicly available for the scientific community.

Main Methods:

• Compilation of 6379 experimental pKa values and QM properties for over 39000 structures.
• Extraction of 309 molecular, atomic, and interatomic features, compatible with conformers and tautomers.
• Training and testing of two pKa prediction models: an ensemble of trees and a graph isomorphic layers model.
• Evaluation on four SAMPL pKa challenge datasets.

Main Results:

• The G-pKa database includes diverse chemical structures and accounts for tautomeric forms in 22% of the data.
• The developed models demonstrate excellent performance on SAMPL datasets.
• The graph neural network-based model achieves state-of-the-art results, outperforming existing algorithms on three of the four datasets.

Conclusions:

• The G-pKa database provides a valuable resource for advancing pKa prediction.
• Graph-based machine learning approaches show significant promise for accurate pKa prediction.
• The developed models and data contribute to improved computational chemistry tools for drug discovery and chemical research.