Basis for Accurate Protein pK_a Prediction with Machine Learning

Area of Science:

• Biochemistry
• Computational Biology
• Protein Science

Background:

• Protein function is modulated by pH through the protonation state of ionizable amino acid side chains, determined by their pKa values.
• Accurate prediction of these pKa values is critical for advancing research in life sciences and for protein and drug design.

Purpose of the Study:

• To present a new theoretical pKa dataset, PHMD549, and evaluate the performance of machine learning models for pKa prediction.
• To demonstrate the efficacy of the DeepKa model in accurately predicting protein pKa values, including challenging cases.

Main Methods:

• Development of the PHMD549 theoretical pKa dataset.
• Application and comparison of four machine learning methods, including DeepKa, using PHMD549 and the EXP67S test set.
• Evaluation of DeepKa's performance on enzyme catalytic sites and intrinsically disordered peptides.

Main Results:

• DeepKa demonstrated significantly improved performance, outperforming other state-of-the-art methods on the EXP67S test set.
• The model accurately reproduced experimental pKa values for acidic dyads in enzyme active sites and showed applicability to intrinsically disordered peptides.
• DeepKa provided highly accurate predictions even for buried side chains where interactions are compensated by desolvation.

Conclusions:

• The PHMD549 and EXP67S datasets serve as benchmarks for future AI-driven protein pKa prediction tools.
• DeepKa, trained on PHMD549, is an efficient predictor applicable to pKa database construction, protein design, and drug discovery.