Unifying pK_a and Protonation Prediction with Sequence-Based Deep Learning

Predictions of pK_a values provide insights into key aspects of molecular behavior, including solubility, lipophilicity, and binding affinity. Despite their importance, experimental microscopic pK_a data remain scarce, creating a bottleneck in the training of accurate prediction models. In addition, inconsistent terminology across commonly used data sets hinders effective model development and benchmarking. While recent advances have been driven largely by graph-based neural networks, the potential of sequence-based deep learning for pK_a prediction remains underexplored. T5Chem, a sequence-based multitask chemical reaction model, offers an attractive way to cast molecular protonation/deprotonation as a language modeling task and to couple microstate generation with subsequent pK_a estimation. To pursue this direction, we introduce pKaCHU (pK_a data that are combined, honed, and updated), a curated data set comprising 9000 experimentally derived microscopic pK_a entries with ionization-state annotations. We also present T5pKa, a text-based transformer model for small-molecule pK_a prediction built on T5Chem. T5pKa leverages multitask learning to enumerate microstates, enabling both protonation and deprotonation to be predicted by a single sequence-to-sequence model, and then predicts microscopic pK_a values from the resulting microstate pairs using a separate regression model. Across benchmark data sets, T5pKa achieves performance comparable to established pK_a prediction tools and published models while offering the advantage of a unified multitasking framework for microstate enumeration and microscopic pK_a prediction.