XGBoost-Based Prediction of pKa Values from Molecular Descriptors

Abstract

The acid dissociation constant (pKa) is a fundamental physicochemical property that governs molecular ionization, solubility, and bioavailability, rendering its accurate prediction essential in drug design and environmental toxicology. This study entailed a systematic comparison of molecular descriptor representations for pKa prediction using eXtreme Gradient Boosting (XGBoost) regression. Four descriptor categories were evaluated: Morgan circular fingerprints, known as extended-connectivity fingerprints (ECFP), with varying radii (one to four) and bit lengths (12–16,384), Molecular ACCess System (MACCS) structural keys, physicochemical descriptors (molecular weight, logP, rotatable bond count, and fragment counts), and combined representations. A comprehensive grid search optimized the XGBoost hyperparameters across all configurations using five-fold cross-validation on a curated dataset of 7,912 compounds with experimental pKa values obtained from DataWarrior. Morgan fingerprints with radius two and 16,384 bits achieved the highest standalone performance (test R2 = 0.795, root mean squared error, RMSE = 1.539), whereas combining Morgan fingerprints (radius one, 2,048 bits) with MACCS keys and physicochemical descriptors yielded the optimal overall result (test R2 = 0.817; RMSE = 1.455). An extended analysis employing 1,135 Mordred and RDKit cheminformatic descriptors with XGBoost feature selection further improved the prediction accuracy (test R2 = 0.829; RMSE = 1.404). Moreover, feature importance analysis identified the number of acidic groups, spectral descriptors, and aromatic amine fragments as the most influential predictors. These results demonstrate that multisource descriptor fusion consistently outperforms individual representations and that XGBoost provides a computationally efficient and interpretable framework for pKa modeling.