KaMLs for Predicting Protein pK_a Values and Ionization States: Are Trees All You Need?

Area of Science:

• Biochemistry
• Computational Biology
• Structural Biology

Background:

• Accurate prediction of protein ionization states is crucial for understanding biological processes and drug discovery.
• Existing physics-based and machine learning (ML) methods face limitations due to complex protein environments and data scarcity, respectively.

Purpose of the Study:

• To develop advanced ML models for predicting protein pKa values and ionization states.
• To improve upon the state-of-the-art in protein electrostatics prediction, addressing limitations of previous approaches.

Main Methods:

• Developed pKa ML (KaML) models using decision trees (KaML-CBtree) and graph attention networks (GAT).
• Utilized a new experimental pKa database (PKAD-3) with highly shifted pKa values and incorporated physicochemical understanding.
• Employed data augmentation with AlphaFold structures and model pretraining on theoretical pKa data.

Main Results:

• KaML-CBtree significantly outperformed current state-of-the-art methods in predicting pKa values and ionization states for all six titratable amino acids.
• Achieved accurate predictions for deprotonated cysteines and lysines, overcoming a previous limitation in the field.
• Introduced classification of protonation states as a novel evaluation metric for pKa prediction models.

Conclusions:

• The developed KaML models, particularly KaML-CBtree, represent a significant advancement in predicting protein ionization states.
• The findings highlight the effectiveness of combining physicochemical insights with ML, novel data strategies, and model architectures.
• An end-to-end pKa predictor based on KaML-CBtree and PKAD-3 is released to facilitate further research in protein electrostatics.