Predict Protein Binding Sites and Epitopes with ScanNet on Vecura

What is ScanNet?

ScanNet (Spatio-Chemical Arrangement of Neighbors Network) is an interpretable geometric deep-learning model designed to predict biologically relevant protein binding sites directly from 3D atomic coordinates. By encoding the local spatio-chemical environment of amino acid residues, it identifies protein-protein interfaces, intrinsically-disordered-partner binding sites, and conformational B-cell epitopes. It operates without the need for evolutionary information or multiple sequence alignments (MSAs), making it highly effective for novel or orphan proteins.

It helps users understand the functional landscape of protein structures by generating residue-level probability scores. It is especially useful for structural biologists and protein engineers looking to pinpoint critical interaction regions in 3D models.

What can users do with ScanNet on Vecura?

With ScanNet on Vecura, users can:

• Predict protein-protein binding sites to identify potential interaction partners.
• Map conformational B-cell epitopes on antigen structures for vaccine design.
• Identify binding sites for intrinsically disordered proteins (IDP).
• Extract structural features from internal model layers for downstream representation learning and clustering tasks.

What the output means

The output provides residue-level probability scores, which can be visualized by mapped directly onto the input PDB structures via the B-factor column. Additionally, users receive structured data containing feature vectors and numerical predictions for each residue.

This output should be used to support scientific decision making. It does not replace experimental validation.

Why this matters

Understanding how proteins interact at the atomic level is fundamental to drug discovery, immunology, and synthetic biology. Traditionally, predicting these sites relied heavily on evolutionary conservation data, which is not always available for every protein. ScanNet’s ability to rely purely on 3D geometry democratizes this analysis for orphan proteins and novel designs.

By providing an interpretable framework where model activations correspond to meaningful structural motifs, ScanNet bridges the gap between opaque deep-learning predictions and physical structural biology.

• Developed by: The research group associated with the original Nature Methods publication (Jérôme Tubiana et al.)
• Source: Official GitHub repository
• Reference: Nature Methods paper (2022)