Integrated Quantum Mechanical and Machine Learning Methods Enable Scalable and General pKa Prediction for Pharmacokinetic Modeling

Abstract

We present the development of the BIOiSIM™ pKa model for predicting aqueous molecular ionization. Assessment and optimization of drug pKa occurs early in drug discovery; thus, models for high‐throughput pKa prediction must balance computational cost, scalability, accuracy, and applicability to diverse chemical structures. Key to our pKa prediction model is a set of low computational cost, quantum mechanical charge descriptors that incorporate not only ionizable group charges, but the additional atoms and bonds likely to modulate pKa. The finalized machine learning models produced good accuracy on the test holdout set (R2 = 0.9, mean absolute error = 0.6−0.7, root mean square error = 1.0) and show good generalizability predicting the most acidic and basic ionization centers of the SAMPL6 external benchmark validation set (mean absolute error = 0.8, root mean square error = 1.0). Performance of the model is comparable to commercial pKa predictors but is trained on less data and exhibits higher scalability needed for pharmacokinetic modeling.