A correlation coefficient-based feature selection approach for virus-host protein-protein interaction prediction.

Area of Science:

• Computational biology
• Bioinformatics
• Machine learning in virology

Background:

• Virus-host protein-protein interactions (PPI) are crucial for viral pathogenesis.
• Accurate prediction of virus-host PPIs aids in understanding viral mechanisms and developing antiviral strategies.
• Machine learning models require effective feature representation of biological data for accurate PPI prediction.

Purpose of the Study:

• To develop and evaluate a correlation coefficient-based feature selection method for virus-host PPI prediction.
• To reduce the dimensionality of biological data while preserving predictive performance.
• To improve the computational efficiency of virus-host PPI prediction tools.

Main Methods:

• Utilized a virus-host PPI dataset and a reduced amino acid alphabet to generate tripeptide features.
• Applied correlation coefficient metrics for feature selection, including statistical relevance testing.
• Compared feature selection models against baseline models and existing prediction tools using AUPR (Area Under the Precision-Recall Curve).

Main Results:

• The Pearson coefficient demonstrated the best performance, with a minimal drop in AUPR (0.003) for the random forest model.
• Achieved a 73.3% reduction in tripeptide features (from 686 to 183).
• Feature selection decreased computation time and space complexity with limited impact on prediction performance.

Conclusions:

• Correlation coefficient-based feature selection is an effective strategy for optimizing virus-host PPI prediction models.
• This approach enhances computational efficiency without substantially compromising predictive accuracy.
• The findings contribute to the development of more practical and scalable tools for studying virus-host interactions.