Unifying pK_a and Protonation Prediction with Sequence-Based Deep Learning.

Area of Science:

• Computational Chemistry
• Machine Learning in Drug Discovery
• Chemical Informatics

Background:

• Accurate prediction of acidity constants (pKa) is vital for understanding molecular properties like solubility and binding affinity.
• Scarcity of experimental microscopic pKa data and inconsistent terminology in existing datasets impede the development of reliable prediction models.
• While graph-based neural networks dominate, sequence-based deep learning for pKa prediction remains an underexplored area.

Purpose of the Study:

• To introduce a curated dataset, pKaCHU, containing 9000 experimentally derived microscopic pKa entries with ionization-state annotations.
• To develop and present T5pKa, a novel text-based transformer model for small-molecule pKa prediction utilizing sequence-based deep learning.
• To leverage a unified multitasking framework for both microstate enumeration and microscopic pKa prediction.

Main Methods:

• Developed pKaCHU, a comprehensive dataset by combining, honing, and updating experimental microscopic pKa data.
• Built T5pKa, a sequence-to-sequence model based on the T5Chem architecture, to predict molecular protonation/deprotonation.
• Employed multitask learning within T5pKa to enumerate microstates and a separate regression model for pKa value prediction.

Main Results:

• T5pKa successfully predicts microscopic pKa values by treating protonation/deprotonation as a language modeling task.
• The model demonstrates performance comparable to existing state-of-the-art pKa prediction tools across benchmark datasets.
• T5pKa offers a unified framework for microstate enumeration and pKa prediction, enhancing model development efficiency.

Conclusions:

• T5pKa represents a significant advancement in small-molecule pKa prediction using sequence-based deep learning.
• The pKaCHU dataset provides a valuable resource for training and benchmarking pKa prediction models.
• This approach offers a promising direction for improving the accuracy and efficiency of predicting key molecular properties.