StackPPI

StackPPI predicts protein-protein interactions by combining multiple biologically relevant feature encodings with a stacked ensemble classifier to improve PPI prediction accuracy.

Key Features:

• Feature Encoding: Uses pseudo amino acid composition, Moreau-Broto autocorrelation descriptor, Moran and Geary autocorrelation descriptors, amino acid composition position-specific scoring matrix (PSSM), Bi-gram PSSM, and composition, transition, and distribution encodings to represent protein sequence characteristics.
• Feature Selection and Dimensionality Reduction: Applies XGBoost for feature selection and dimensionality reduction using gradient boosting and average gain techniques to reduce feature noise.
• Stacked Ensemble Classifier: Employs a stacked ensemble integrating random forest, extremely randomized trees, and logistic regression for final PPI prediction.

Scientific Applications:

• Predictive Accuracy: Demonstrates performance under five-fold cross-validation with ACC 89.27%, MCC 0.7859, AUC 0.9561 on Helicobacter pylori datasets and ACC 94.64%, MCC 0.8934, AUC 0.9810 on Saccharomyces cerevisiae datasets.
• Comparison to State-of-the-Art Models: Outperforms existing state-of-the-art models on independent test sets.
• Biological Significance: Facilitates inference of biologically significant PPI networks for elucidating functional pathways and mechanisms relevant to drug design and biomedical research.

Methodology:

Feature encodings (pseudo amino acid composition; Moreau-Broto, Moran, and Geary autocorrelation descriptors; amino acid composition PSSM; Bi-gram PSSM; composition/transition/distribution) are generated, XGBoost is used for feature selection/dimensionality reduction via gradient boosting and average gain, a stacked ensemble of random forest, extremely randomized trees, and logistic regression performs classification, and performance is evaluated by five-fold cross-validation reporting ACC, MCC, and AUC on Helicobacter pylori and Saccharomyces cerevisiae datasets.