iLoops

iLoops predicts protein-protein interactions (PPIs) by leveraging local structural features—specifically loops and domains—to assess interaction propensity between queried protein pairs.

Key Features:

• Input Requirements: Accepts protein sequences and user-specified pairs of proteins to test for potential interactions.
• Structural Feature Assignment: Assigns structural features to query proteins based on sequence similarity using ArchDB for loops and SCOP for domains.
• Structural Feature Classification: Classifies pairs of structural features (loops or domains) by their propensity to favor or disfavor interactions derived from observations in known interacting and non-interacting protein pairs.
• Prediction Methodology: Employs a random forest classifier to evaluate the potential interaction between protein pairs.
• Surface-wide Analysis: Considers the entire protein surface and integrates both favoring and disfavoring structural features when scoring interaction likelihoods.
• Scoring for Unbalanced Datasets: Produces interaction likelihood scores applicable to datasets with unbalanced ratios of interacting and non-interacting pairs.

Scientific Applications:

• Elucidation of PPI mechanisms: Predicts interactions based on local structural features to reveal molecular determinants of protein binding.
• Interactome construction: Supports generation of protein interaction networks by identifying potential PPIs across protein pairs.
• Energy landscape analysis: Provides scores that support the funnel-like intermolecular energy landscape theory by identifying likely docking interfaces.
• Handling challenging datasets: Aids identification of potential PPIs even in conditions with unbalanced interacting versus non-interacting pair ratios.

Methodology:

Assigns structural features to query sequences by sequence similarity against ArchDB (loops) and SCOP (domains); classifies pairs of structural features using observations from known interacting and non-interacting pairs; evaluates protein-pair interaction likelihoods using a random forest classifier while considering the entire protein surface and both favoring and disfavoring features; reported predictions can exceed 25% accuracy in challenging scenarios.