InterPepRank

InterPepRank scores and ranks peptide-protein complex decoys by predicting Local Root Mean Square Deviation (LRMSD) using edge-conditioned graph convolutions on graph-encoded structural and evolutionary features.

Key Features:

• Graph encoding: Encodes peptide-protein complex structures as graphs with nodes representing evolutionary and sequence features and edges capturing physical pairwise interactions.
• Edge-conditioned convolutions: Applies edge-conditioned graph convolutions to learn from the graph representation of complexes.
• LRMSD prediction: Predicts Local Root Mean Square Deviation (LRMSD) of decoys as a quantitative measure of conformational accuracy.
• Training data: Trained on a large dataset of peptide-protein complex decoys.
• Evaluation rigor: Validated on an independent test set curated to avoid overlap in CATH annotation or sequence identity, achieving a median AUC of 0.86 for identifying complexes with LRMSD < 4 Å compared to ~0.69 for other methods.
• Docking impact: Integration into peptide docking pipelines increases production of medium-quality models by 80% and high-quality models by 40%.

Scientific Applications:

• Scoring and ranking: Prioritizes near-native peptide-protein conformations by scoring decoys using predicted LRMSD.
• Docking pipeline enhancement: Improves model selection within peptide docking pipelines, increasing yields of medium and high-quality models.
• Coarse model identification: Discriminates coarse peptide-protein complexes with LRMSD < 4 Å using AUC-based performance metrics (median AUC 0.86).

Methodology:

Encodes complexes as graphs (nodes: evolutionary and sequence features; edges: physical pairwise interactions), uses edge-conditioned graph convolutions to predict LRMSD, and is trained and evaluated on large decoy datasets with an independent test set curated to avoid CATH annotation or sequence identity overlap.