PepBonPred

PepBonPred predicts peptide binding regions within protein sequences to identify peptides in disordered regions that bind structured domains, using Vina docking scores and a bidirectional recurrent neural network trained on SLiM-containing proteins from the ELM database.

Key Features:

• Training Dataset: Uses a non-redundant dataset derived from the ELM database containing proteins with Short Linear Motifs (SLiMs).
• Docking Analysis: Employs AutoDock Vina docking scores to assess overlapping tripeptides within disordered regions for binding to structured domains in associated PDB structures.
• Machine Learning Integration: Implements a bidirectional recurrent neural network (BRNN) trained on protein sequence, predicted secondary structure, Vina docking scores, and disorder predictions.
• Performance Metrics: Reports that Vina docking scores alone yield limited discrimination (AUC 0.58) while the BRNN integration improves prediction accuracy (AUC 0.72).

Scientific Applications:

• Mapping peptide-binding sites: Identification of peptides in disordered regions that interact with structured domains to map peptide-binding sites in proteins.
• Studying disordered-region function: Investigation of the functional roles of intrinsically disordered regions that contain SLiMs and mediate interactions.
• Protein–protein interaction networks: Informing analyses of protein–protein interaction networks by prioritizing disordered-region peptides that mediate binding to structured domains.

Methodology:

Uses Vina docking scores calculated for overlapping tripeptides in predicted disordered regions against associated PDB structures as an initial discriminator (AUC 0.58), and trains a BRNN on inputs including protein sequence, predicted secondary structure, Vina scores, and disorder predictions using a non-redundant ELM-derived SLiM dataset to achieve improved performance (AUC 0.72).