Protein Electrostatic Properties are Finetuned Through Evolution.

Area of Science:

• Biochemistry and Molecular Biology
• Computational Biology
• Artificial Intelligence in Science

Background:

• Protein ionization states (pKa values) are critical for protein function, but predicting them accurately remains a significant challenge.
• Existing structure-based methods have limitations in predicting protein pKa values across various conditions.
• Understanding protein electrostatics is key to deciphering biological mechanisms and engineering novel proteins.

Purpose of the Study:

• To develop a novel, sequence-based deep learning approach for accurate protein pKa prediction.
• To challenge the existing structure-based paradigm in protein electrostatics prediction.
• To create a versatile platform (KaML) for various applications in protein science and engineering.

Main Methods:

• Developed KaML-ESMs, neural networks integrating ESM protein language models.
• Trained models on a synthetically augmented experimental dataset of protein pKa values.
• Utilized a novel latent space sampling approach (GAINES) to address data scarcity.

Main Results:

• KaML-ESMs significantly outperform traditional structure-based methods for protein pKa prediction.
• Achieved high accuracy across six ionizable amino acids in native proteins, approaching experimental resolution.
• KaML-ESM2 demonstrated superior performance on challenging engineered protein datasets (OBTRUDEs) with an RMSE of 1.36.
• Application to the human proteome identified functional sites and inferred catalytic mechanisms.

Conclusions:

• Protein electrostatic properties, including pKa values, are effectively encoded within the amino acid sequence.
• Sequence-based prediction offers a powerful alternative to structure-based methods, potentially revealing evolutionary co-optimization of sequence, structure, and function.
• The KaML platform provides a robust tool for advancing biological exploration, drug design, protein engineering, and biomolecular simulations.