Protein stability models fail to capture epistatic interactions of double point mutations

Area of Science:

• Biochemistry and Molecular Biology
• Computational Biology
• Protein Engineering

Background:

• Accurate prediction of protein stability changes from mutations is crucial for therapeutics and understanding disease.
• Recent advances focus on single-point mutations, with less attention on double point mutations.
• Understanding double mutations is vital as they can significantly alter protein function and stability.

Purpose of the Study:

• To analyze the largest available dataset of double point mutation stability.
• To benchmark existing protein stability prediction models on double mutant data.
• To investigate the performance of additive versus non-additive models, including AI and physics-based approaches, in capturing epistatic interactions.

Main Methods:

• Analysis of the largest available double point mutation stability dataset.
• Benchmarking of multiple protein stability prediction models (AI-based, physics-based, additive, non-additive).
• Development of an extension to the ThermoMPNN framework and a novel data augmentation scheme.

Main Results:

• Additive models demonstrate surprisingly strong performance, comparable to non-additive models for predicting double mutant stability.
• Current AI and physics-based models do not consistently capture epistatic interactions between mutations.
• Epistasis-aware models show marginal improvement over additive models for predicting stabilizing double mutations.

Conclusions:

• Current protein stability models have limitations in predicting effects of concurrent mutations due to dataset constraints and model sensitivity.
• Additive models offer a robust baseline for double mutant stability prediction.
• Further development is needed to accurately model complex epistatic interactions in protein engineering and disease research.