Graph-Based Deep Learning Models for Predicting pK_a Values of Protein-Ionizable Residues via Physically Inspired...

Area of Science:

• Biochemistry
• Computational Biology
• Structural Biology

Background:

• Protein pKa values are critical for biological functions like enzymatic activity and ligand binding.
• Accurate pKa prediction is essential for drug discovery, but current methods like PROPKA have limitations.

Purpose of the Study:

• To develop an improved computational framework for predicting protein residue pKa values.
• To enhance the accuracy and generalizability of pKa predictions using integrated molecular dynamics and deep learning.

Main Methods:

• Utilized high-throughput molecular modeling with the AMOEBA polarized force field to generate a protein structure dataset.
• Engineered physics-inspired features, including atomic electrostatics.
• Trained three graph-based neural network models on experimentally determined pKa values from the PKAD-2 dataset.

Main Results:

• All developed models showed significant improvements in pKa prediction accuracy for aspartic acid, glutamic acid, lysine, and histidine compared to PROPKA3.5.1.
• The graph attention networks model demonstrated superior accuracy and generalizability against recent machine learning benchmarks.
• Feature importance analysis revealed physically meaningful patterns related to the protein microenvironment and atomic geometry.

Conclusions:

• The integrated framework offers a substantial advancement in predicting protein residue pKa values.
• The developed deep learning models and curated dataset provide a valuable resource for researchers.
• Potential applications include early-stage drug target identification and protein engineering.