Machine Learning: How Much Does It Tell about Protein Folding Rates?

Area of Science:

• Biophysics
• Computational Biology
• Machine Learning

Background:

• Protein folding rate prediction is crucial for understanding protein folding principles.
• Numerous computational models, including machine learning-based ones, have been developed.
• Some models claim over 90% accuracy in predicting protein folding rates, but these claims may be inflated.

Purpose of the Study:

• To evaluate the actual predictive power of published protein folding rate models.
• To identify common machine learning misconceptions leading to overestimated accuracy.
• To propose methods for more reliable model assessment.

Main Methods:

• Tested three selected published machine learning models using new experimental data.
• Analyzed model performance, focusing on predictive power and potential overfitting.
• Investigated common violations of machine learning principles in model development.

Main Results:

• Published models demonstrated significantly lower predictive power on new data than originally reported.
• Overly optimistic accuracy claims stemmed from violations of fundamental machine learning practices.
• Current experimental data is insufficient for robust linear predictors based on amino acid composition.

Conclusions:

• Existing protein folding rate prediction models may be unreliable due to exaggerated claims.
• Learning curves are recommended as a safeguard against common machine learning errors.
• Further research with more data is needed for accurate protein folding rate prediction.