Epik: pK_a and Protonation State Prediction through Machine Learning

Area of Science:

• Computational chemistry
• Machine learning in drug discovery
• Molecular modeling

Background:

• Accurate prediction of molecular properties like pKa is crucial for drug design.
• Existing methods may lack accuracy or speed for complex molecules.
• Protonation state influences a molecule's behavior and efficacy.

Purpose of the Study:

• To introduce Epik version 7, an advanced software for predicting pKa values and protonation states.
• To enhance the speed and accuracy of computational drug discovery workflows.
• To enable the analysis of ultra-large compound libraries.

Main Methods:

• Utilized an ensemble of atomic graph convolutional neural networks (GCNNs).
• Trained the model on over 42,000 experimental and computed pKa values.
• Validated performance across seven diverse test sets.

Main Results:

• Achieved median absolute error of 0.42 pKa units and root mean square error of 0.72 pKa units.
• Successfully recovered 95% of the most populated protonation states.
• Demonstrated rapid processing time of 47 ms per ligand on average.

Conclusions:

• Epik version 7 offers a significant advancement in predicting pKa and protonation states.
• The software's speed and accuracy support efficient exploration of chemical space.
• Customizable models can be generated for specific chemical contexts.