Rewiring protein sequence and structure generative models to enhance protein stability prediction.

Area of Science:

• Computational biology
• Protein engineering
• Machine learning in bioinformatics

Background:

• Predicting protein thermostability changes from amino acid substitutions is crucial for disease research and protein engineering.
• Existing protein generative models show promise but have limitations in predicting protein functions like stability.
• The potential of these models to enhance protein stability prediction remains largely unexplored.

Purpose of the Study:

• To introduce SPURS, a novel deep learning framework for predicting protein thermostability.
• To integrate protein language models (ESM) and inverse folding models (ProteinMPNN) for enhanced stability prediction.
• To evaluate SPURS's performance and versatility in protein stability and function analyses.

Main Methods:

• SPURS integrates ESM and ProteinMPNN using a neural network module to combine sequence and structure information.
• A rewiring strategy enhances sequence representation learning by incorporating structure priors.
• The framework is trained on a large-scale thermostability dataset for supervised prediction of mutation effects.

Main Results:

• SPURS consistently outperforms state-of-the-art methods in accuracy, speed, scalability, and generalizability across 12 benchmark datasets.
• The framework accurately identifies protein functional sites in an unsupervised manner when combined with a protein language model.
• SPURS improves low-N protein fitness prediction models by acting as a stability prior.

Conclusions:

• SPURS is a powerful tool for advancing protein stability prediction and machine learning-guided protein engineering.
• The framework's ability to integrate sequence and structure data offers significant advantages.
• SPURS demonstrates versatility, enhancing both stability prediction and functional site identification.