Computational Modeling of Protein Stability: Quantitative Analysis Reveals Solutions to Pervasive Problems

Area of Science:

• Protein engineering
• Computational biology
• Biophysics

Background:

• Accurate modeling of mutation effects on protein stability is vital for protein engineering and understanding biological processes.
• Current prediction tools are widely used but their performance and limitations are not fully understood.

Purpose of the Study:

• To quantitatively analyze the performance of protein stability prediction tools.
• To identify biases and limitations in current prediction methodologies.
• To recommend improved performance metrics for evaluating these tools.

Main Methods:

• Rigorous quantitative analysis of mutation effects on protein stability and solubility.
• Evaluation of commonly used performance metrics, including classification accuracy.
• Assessment of a simple two-neuron neural network for stability prediction.
• Comparison of predicted versus experimentally determined mutation effects.

Main Results:

• Stability prediction tools often favor mutations increasing stability at the expense of solubility.
• Mutations predicted to stabilize proteins show near-neutral experimental effects on average.
• High classification accuracy can mask poor performance, as shown by a simple neural network with a low Matthews correlation coefficient (MCC).

Conclusions:

• The Matthews correlation coefficient (MCC) is a more informative metric than classification accuracy for evaluating protein stability prediction tools.
• Multiple mutations can enhance the likelihood of achieving stabilization targets.
• Future improvements in prediction tool precision could lead to significant advancements in protein engineering.