VoroIF-GNN

VoroIF-GNN evaluates inter-subunit interfaces in protein–protein complexes by extracting contacts from Voronoi tessellation of atomic balls and using an attention-based Graph Neural Network to predict per-contact accuracy and aggregate these predictions into interface scores.

Key Features:

• Voronoi tessellation: Derives interface contacts from the Voronoi tessellation of atomic balls within a protein complex model.
• Graph representation: Represents interface contacts as graph nodes with edges capturing interactions between contacts.
• Attention-based GNN: Employs an attention-based Graph Neural Network to predict the accuracy of individual contacts.
• Per-contact to interface scoring: Produces per-contact accuracy predictions that are summarized into an overall inter-subunit interface score.
• Structure-only input: Operates solely on structural models of multimeric proteins without requiring supplementary data beyond the input structure.
• Benchmarking: Demonstrated performance in CASP15 by selecting the most accurate multimeric model among candidates.

Scientific Applications:

• Interface assessment: Quantitative evaluation of inter-subunit interfaces in protein–protein complexes at both contact and interface levels.
• Model selection: Selection of the most accurate multimeric model from multiple candidate structural models.
• Model refinement and analysis: Informing refinement and evaluation of protein complex models to support studies of molecular interactions.

Methodology:

Derives interface contacts from Voronoi tessellation of atomic balls, represents contacts as graph nodes with edges for their interactions, applies an attention-based Graph Neural Network to predict per-contact accuracy, and summarizes these predictions into an overall interface score; operates on input structural models without additional data.