Comparative Study on Feature Selection in Protein Structure and Function Prediction.

Area of Science:

• * Bioinformatics and Computational Biology
• * Protein Science and Engineering

Background:

• * Effective methods for predicting protein structure and function rely on extracting and fusing diverse protein features.
• * Existing feature extraction methods exhibit varying strengths and weaknesses, necessitating optimized feature selection for improved predictive performance.

Purpose of the Study:

• * To systematically compare and analyze the efficiency of different feature selection methods in predicting protein structure and function.
• * To identify optimal feature selection strategies for enhancing the accuracy of protein structure, disorder, molecular chaperone, and solubility predictions.

Main Methods:

• * Evaluation of various feature selection techniques applied to protein sequence and structure data.
• * Comparative analysis of method performance across multiple prediction tasks, including protein structure, disorder, molecular chaperone, and solubility prediction.
• * Utilizing nonlinear Support Vector Machine (SVM) as a key component in the feature selection process.

Main Results:

• * The nonlinear SVM-based feature selection method demonstrated superior performance across all evaluated prediction tasks.
• * Significant accuracy improvements ranging from 13.16% to 71% were observed after feature selection.
• * Kmer and Position-Specific Scoring Matrix (PSSM) features showed particular benefit from the selected feature subsets.

Conclusions:

• * Nonlinear SVM-based feature selection is a highly effective strategy for improving protein function and structure prediction.
• * The judicious selection of features, especially Kmer and PSSM, substantially enhances predictive model accuracy.
• * Optimized feature selection is critical for advancing the development of more precise protein prediction methodologies.