Interpretable deep-learning pKa prediction for small molecule drugs via atomic sensitivity analysis

Machine learning (ML) models play a crucial role in predicting properties essential to drug development, such as a drug’s logscale acid-dissociation constant (pKa). Despite recent architectural advances, these models often generalize poorly to novel compounds due to a scarcity of ground-truth data. Further, these models lack interpretability, in part due to a dependence on explicit encodings of input molecules’ molecular substructures. To this end, atomic-resolution information is accessible in chemical structures by observing model response to atomic perturbations of an input molecule; however, no methods exist that systematically utilize this information for model and molecular analysis. Here, we present BCL-XpKa, a substructure-independent, deep neural network (DNN)-based pKa predictor that generalizes well to novel small molecules. BCL-XpKa discretizes pKa prediction from a regression problem into a multitask-classification problem, which accumulates data for prediction at biologically relevant pH values and records the model’s uncertainty in its prediction as a discrete distribution for each pKa prediction. BCL-XpKa outperforms modern ML pKa predictors and accurately models the effects of common molecular modifications on a molecule’s ionizability. We then leverage BCL-XpKa’s substructure independence to introduce atomic sensitivity analysis (ASA), which quickly decomposes a molecule’s predicted pKa value into its respective atomic contributions without model retraining. When paired with BCL-XpKa, ASA informs that BCL-XpKa has implicitly learned high-resolution information about molecular substructures. We further demonstrate ASA’s utility in structure preparation for protein-ligand docking by identifying ionization sites in 97.8% and 83.4% of complex small molecule acids and bases. We then apply ASA with BCL-XpKa to understand the physicochemical liabilities and guide optimization of a recently published KRAS-degrading PROTAC.