Artificial intelligence challenges for predicting the impact of mutations on protein stability

Area of Science:

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

Background:

• Protein stability is essential for protein function and is influenced by mutations.
• Accurate prediction of mutation effects on protein stability is vital for protein engineering, drug design, and understanding genetic variants.
• Numerous computational methods, particularly those leveraging artificial intelligence (AI), have been developed to predict these effects.

Purpose of the Study:

• To critically analyze the features, algorithms, computational efficiency, and accuracy of AI-based mutation effect predictors.
• To identify limitations, biases, generalizability issues, and interpretability challenges in current prediction methods.
• To assess the progress and current state of predicting protein stability changes due to mutations.

Main Methods:

• Review and critical analysis of existing AI-based methods for predicting protein stability changes upon mutation.
• Evaluation of predictor performance using an independent test set.
• Assessment of computational efficiency, generalizability, and interpretability.

Main Results:

• The accuracy of protein stability change prediction methods has plateaued at approximately 1 kcal/mol for over 15 years.
• Common limitations include biases toward training data, and challenges in generalizability and interpretability.
• Despite advancements in AI, significant improvements in prediction accuracy have not been realized recently.

Conclusions:

• Current AI predictors for mutation effects on protein stability have reached a performance ceiling.
• Addressing limitations in generalizability, interpretability, and training data biases is crucial for future progress.
• Further research is needed to overcome existing challenges and achieve substantial improvements in prediction accuracy.