Using predicted shape string to enhance the accuracy of γ-turn prediction

Area of Science:

• Structural bioinformatics
• Computational biology
• Machine learning in protein science

Background:

• Predicting gamma-turns in proteins is crucial for understanding protein folding and function.
• Existing prediction methods have limited accuracy, with a Matthews correlation coefficient (MCC) typically below 0.18.

Purpose of the Study:

• To develop an improved method for accurate prediction of gamma-turns in proteins.
• To introduce protein shape string as a novel feature for enhancing prediction accuracy.

Main Methods:

• Utilized the geometric mean metric to optimize support vector machine performance on imbalanced datasets.
• Developed a predictor to generate protein shape strings via structure alignment against a protein structure database.
• Employed a fivefold cross-validation technique on a benchmark dataset of 320 non-homologous protein chains.

Main Results:

• Achieved an overall prediction accuracy (Qtotal) of 92.2% and an MCC of 0.38.
• Demonstrated superior performance compared to existing gamma-turn prediction methods.
• Validated the utility of protein shape strings and dihedral angle information for predicting protein tight turns.

Conclusions:

• The developed method significantly enhances gamma-turn prediction accuracy.
• Protein shape string is a valuable feature for predicting protein tight turns.
• Dihedral angle information is a reasonable variable for machine learning in protein folding prediction.