Protein-Protein Interaction Sites Prediction Based on an Under-Sampling Strategy and Random Forest Algorithm

Area of Science:

• Computational Biology
• Bioinformatics
• Structural Biology

Background:

• Protein-protein interactions (PPIs) are crucial for cellular functions.
• Predicting PPI sites computationally offers an alternative to costly experimental methods.
• Class imbalance between interface and non-interface residues hinders prediction accuracy.

Purpose of the Study:

• To develop a novel computational strategy for predicting protein-protein interaction sites.
• To address the challenge of class imbalance in protein sequence datasets.
• To enhance the accuracy of differentiating interface from non-interface residues.

Main Methods:

• Utilized Position-Specific Scoring Matrix (PSSM)-derived features, hydropathy index (HI), and relative solvent accessibility (RSA) for residue representation.
• Implemented NearMiss-based under-sampling to balance the dataset by reducing non-interface residues.
• Employed the Random Forest (RF) algorithm for binary classification of residues.

Main Results:

• The proposed NearMiss-based under-sampling for unbalancing datasets and Random Forest classification (NM-RF) model achieved high prediction accuracy.
• Achieved 87.6% accuracy on the Dtestset72 dataset.
• Achieved 84.3% accuracy on the PDBtestset164 dataset.

Conclusions:

• The NM-RF method effectively overcomes class imbalance issues in predicting protein interaction sites.
• The developed method demonstrates significant improvement in differentiating interface and non-interface residues.
• NM-RF offers a promising computational tool for advancing PPI site prediction research.