Interpretable Deep-Learning\np<i>K</i><sub>a</sub> Prediction for Small Molecule Drugs\nvia Atomic Sensitivity\nAnalysis

Machine learning (ML) models now play a crucial role\nin predicting\nproperties essential to drug development, such as a drug’s\nlogscale acid-dissociation constant (p<i>K</i><sub>a</sub>). Despite recent architectural advances, these models often generalize\npoorly to novel compounds due to a scarcity of ground-truth data.\nFurther, these models lack interpretability. To this end, with deliberate\nmolecular embeddings, atomic-resolution information is accessible\nin chemical structures by observing the model response to atomic perturbations\nof an input molecule. Here, we present BCL-XpKa, a deep neural network\n(DNN)-based multitask classifier for p<i>K</i><sub>a</sub> prediction that encodes local atomic environments through Mol2D\ndescriptors. BCL-XpKa outputs a discrete distribution for each molecule,\nwhich stores the p<i>K</i><sub>a</sub> prediction and the\nmodel’s uncertainty for that molecule. BCL-XpKa generalizes\nwell to novel small molecules. BCL-XpKa performs competitively with\nmodern ML p<i>K</i><sub>a</sub> predictors, outperforms\nseveral models in generalization tasks, and accurately models the\neffects of common molecular modifications on a molecule’s ionizability.\nWe then leverage BCL-XpKa’s granular descriptor set and distribution-centered\noutput through atomic sensitivity analysis (ASA), which decomposes\na molecule’s predicted p<i>K</i><sub>a</sub> value\ninto its respective atomic contributions without model retraining.\nASA reveals that BCL-XpKa has implicitly learned high-resolution information\nabout molecular substructures. We further demonstrate ASA’s\nutility in structure preparation for protein–ligand docking\nby identifying ionization sites in 93.2% and 87.8% of complex small\nmolecule acids and bases. We then applied ASA with BCL-XpKa to identify\nand optimize the physicochemical liabilities of a recently published\nKRAS-degrading PROTAC.