ME-pK_a: A Deep Learning Method with Multimodal Learning for Protein pK_a Prediction.

Area of Science:

• Biochemistry and Molecular Biology
• Computational Biology and Cheminformatics

Background:

• Protein pKa values dictate amino acid protonation states, crucial for protein structure, function, and drug interactions.
• Experimental pKa determination is laborious, and existing computational methods struggle with data limitations and complex protein attributes, particularly for buried residues.

Purpose of the Study:

• To develop a novel, accurate, and efficient multimodal protein pKa prediction model.
• To improve predictions for buried residues and enhance generalization across different residue types.

Main Methods:

• Developed ME-pKa (Multimodal ESM pKa), a model integrating local amino acid environmental attributes with FASTA sequence characteristics.
• Employed a multifidelity learning strategy to augment data and address limited data availability.
• Validated performance against state-of-the-art models on benchmark datasets.

Main Results:

• ME-pKa achieved superior prediction accuracy, outperforming existing models with low RMSE (0.845 ± 0.09) and MAE (0.641 ± 0.07) on the PE-pKa dataset.
• Demonstrated robust performance across major ionizable residues (Asp, Glu, His, Lys).
• Showcased exceptional accuracy for buried residues (RSA < 0.2), achieving low MAE values on multiple datasets.
• Confirmed the pH-dependent binding of PD-L1 antibodies, highlighting the model's practical implications in drug design.

Conclusions:

• ME-pKa offers a significant advancement in predicting protein pKa values, particularly for challenging buried residues.
• The model's ability to integrate multimodal data and employ multifidelity learning enhances accuracy and generalization.
• Accurate pKa prediction is vital for understanding protein mechanisms and advancing rational drug design.