SCRIBER

SCRIBER predicts protein-binding residues (PBRs) from protein sequences while minimizing cross-predictions with RNA-, DNA- and small ligand-binding residues.

Key Features:

• Comprehensive Dataset: Leverages a dataset built by transferring binding annotations from multiple protein–protein complexes to provide diverse training examples of binding residues.
• Input Features: Incorporates sequence-derived features specifically selected to distinguish protein-binding residues from RNA-, DNA- and small ligand-binding residues.
• Two-layer Architecture: Uses a two-layer prediction architecture in which the first layer predicts protein-, RNA-, DNA- and small ligand-binding residues and the second layer refines these outputs to reduce cross-predictions and focus on PBRs.
• Performance: Empirical evaluation on independent datasets shows reduction of cross-predictions by 41%–69% relative to existing predictors while maintaining high accuracy and achieving at least a 3× speed increase enabling genome-scale applications.
• Proteome-scale Prediction: Applied across the human proteome to predict putative PBRs and estimates that approximately 14% of known human protein domains are involved in protein binding.

Scientific Applications:

• Protein–protein interaction mapping: Provides residue-level PBR predictions to inform interaction network analyses.
• Protein–protein docking: Supplies candidate interface residues to constrain and validate docking models.
• Functional annotation: Aids annotation of protein function by identifying likely interaction interfaces.
• Structural biology: Supports interpretation of structural data by highlighting potential binding sites.
• Drug discovery: Identifies protein interface residues that may serve as therapeutic targets.
• Systems biology: Enables large-scale studies of interaction topology by providing genome-scale PBR predictions.

Methodology:

Training data were assembled by transferring binding annotations from protein–protein complexes; a two-layer predictor first forecasts protein-, RNA-, DNA- and small ligand-binding residues and then refines PBR predictions; performance was validated on independent datasets showing 41%–69% reduced cross-predictions and ≥3× speed, and the method was applied to the human proteome to predict putative PBRs (~14% of domains).