Structure-based self-supervised learning enables ultrafast protein stability prediction upon mutation.

Area of Science:

• Computational Biology
• Protein Engineering
• Biophysics

Background:

• Accurate prediction of protein free energy changes (ΔΔG) is crucial for understanding protein evolution, engineering novel proteins, and developing pharmaceuticals.
• Traditional ΔΔG prediction methods face limitations in computational speed and can be hindered by biased training datasets, especially for diverse protein sequences.

Purpose of the Study:

• To introduce Pythia, a self-supervised graph neural network designed for efficient and accurate zero-shot ΔΔG predictions.
• To benchmark Pythia against existing methods and demonstrate its superior performance and computational speed.

Main Methods:

• Development of Pythia, a self-supervised graph neural network model.
• Comparative benchmarking against self-supervised pretraining models, force field-based approaches, and fully supervised models.
• Validation of Pythia's performance in predicting thermostabilizing mutations for limonene epoxide hydrolase.

Main Results:

• Pythia outperforms existing self-supervised and force field-based methods for ΔΔG prediction.
• Pythia achieves competitive performance with fully supervised models while offering a computational speed increase of up to 10⁵-fold.
• Pythia successfully identified thermostabilizing mutations, leading to higher experimental success rates and enabling the exploration of 26 million protein structures.

Conclusions:

• Pythia represents a significant advancement in predicting ΔΔG, offering a powerful and efficient tool for protein engineering and drug discovery.
• The model's speed and accuracy facilitate large-scale exploration of protein sequence-phenotype relationships.
• A web server is available at https://pythia.wulab.xyz for user-friendly ΔΔG predictions.