Improving Small Molecule pK _a Prediction Using Transfer Learning With Graph Neural Networks

Area of Science:

• Computational chemistry
• Drug discovery
• Machine learning in chemistry

Background:

• Accurate prediction of small molecule protonation states and microstate pKa values is crucial for drug discovery and molecular modeling.
• Existing computational tools face limitations including high costs, resource demands, and complex usability.
• There is a need for accessible, accurate, and efficient methods for pKa prediction.

Purpose of the Study:

• To develop an open-source computational tool for enumerating protonation states and calculating microstate pKa values.
• To leverage graph neural networks for high-accuracy pKa prediction.
• To provide a user-friendly and cost-effective solution for researchers.

Main Methods:

• A graph neural network model was trained on over 700,000 calculated pKa predictions from the ChEMBL database.
• The model was fine-tuned using 5,994 experimental pKa values for improved accuracy.
• The graph neural network was integrated with Dimorphite-DL for protonation state enumeration.

Main Results:

• The developed pkasolver package demonstrates high accuracy in predicting microstate pKa values.
• The model shows improved performance on challenging test datasets after fine-tuning.
• pkasolver successfully combines protonation state generation with accurate pKa calculation.

Conclusions:

• pkasolver offers an accurate, open-source solution for calculating microstate pKa values and enumerating protonation states.
• This tool addresses limitations of existing commercial and non-commercial software.
• pkasolver facilitates lead optimization and molecular modeling in drug discovery.