ScanNet

What is ScanNet?

ScanNet (Spatio-Chemical Arrangement of Neighbors Network) is a geometric deep learning model for predicting protein binding sites directly from 3D structures. It identifies where proteins interact with other proteins, antibodies, or intrinsically disordered proteins (IDPs).

Unlike traditional methods that rely on handcrafted features or structural homology, ScanNet learns spatio-chemical patterns end-to-end from atomic coordinates. The model constructs representations by examining the spatial and chemical arrangement of neighboring atoms, enabling it to detect binding sites even on novel protein folds not seen during training.

Published in Nature Methods in 2022, ScanNet demonstrated state-of-the-art accuracy on multiple benchmarks while remaining interpretable through visualization of learned filters.

Applications

Recommended ScanNet applications:

• Protein-protein interface mapping: Identifying which residues mediate interactions between protein binding partners
• Epitope prediction: Locating antibody binding sites on antigens for vaccine design and therapeutic development
• IDP binding site detection: Finding regions where structured proteins interact with intrinsically disordered proteins
• Drug target analysis: Characterizing binding surfaces for structure-based drug design
• Structural biology: Guiding mutagenesis experiments to validate predicted interaction sites

Limitations

Known ScanNet limitations:

• Static structure assumption: ScanNet analyzes a single conformational state. Proteins with multiple binding-competent conformations or significant induced-fit changes may require ensemble analysis.
• Structure-only execution: ProteinIQ currently runs ScanNet in no-MSA mode, which avoids local HHblits and sequence database requirements. Natural proteins with rich homolog coverage may benefit from MSA-enabled ScanNet runs in a separately configured environment.
• Binding partner agnostic: ScanNet predicts general binding propensity, not partner-specific interfaces. A residue predicted as high probability may participate in any protein-protein interaction, not necessarily the one of interest.
• Resolution sensitivity: Very low-resolution structures or models with significant coordinate errors may produce unreliable predictions.

How to use ScanNet online

ProteinIQ provides a web interface for running ScanNet without command-line installation or local database setup. Upload a protein structure, select a prediction mode, and receive per-residue binding probabilities.

Inputs

Settings

Prediction settings

ProteinIQ runs ScanNet in no-MSA mode for this online job type, so predictions use the submitted structure without building multiple sequence alignments.

Results

ScanNet returns a table of per-residue binding site predictions and downloadable native result files, including the prediction CSV and annotated structure or visualization scripts when ScanNet writes them.

Interpreting binding probabilities

Higher probabilities indicate residues ScanNet ranks as more likely to participate in a binding interface. Use the per-residue values together with the annotated structure outputs to inspect contiguous surface patches rather than relying on a single residue in isolation. ScanNet does not define universal high/medium/low probability cutoffs, so interpret the score distribution in the context of the submitted structure.

How does ScanNet work?

ScanNet employs a hierarchical architecture that processes protein structures at both atomic and amino acid scales, learning to recognize spatio-chemical patterns associated with binding interfaces.

Spatio-chemical filters

At the core of ScanNet are trainable filters that detect specific spatial arrangements of atoms with particular chemical properties. For each atom in the structure, neighboring atoms within a local coordinate frame are extracted. These point clouds pass through linear filters that respond to arrangements such as "hydrophobic atoms surrounding a polar center" or "aromatic ring flanked by charged residues."

The filters are parameterized to be interpretable—each can be visualized to understand what molecular pattern it detects. This interpretability distinguishes ScanNet from black-box deep learning approaches.

Hierarchical representation

ScanNet builds representations at two scales:

1. Atomic level: Spatio-chemical filters process local atomic neighborhoods, producing per-atom features that capture fine structural details.
2. Amino acid level: Atomic features are aggregated within each residue and combined with amino acid attributes (type, secondary structure). A second round of spatio-chemical filtering operates on amino acid neighborhoods.

This multi-scale approach allows ScanNet to integrate information from individual atomic contacts up to residue-level surface geometry.

Evolutionary information

When MSA data is available in a configured ScanNet installation, the method can incorporate position-weight matrices derived from sequence alignments. Conservation patterns provide complementary evidence because residues conserved across evolution often participate in functionally important interfaces. The ProteinIQ online job runs the no-MSA mode, which relies on structural features.

Training and generalization

ScanNet was trained on protein-protein binding sites from the MaSIF-site benchmark dataset. Testing revealed strong generalization: even for proteins with no sequence or fold similarity to training examples, ScanNet maintained high accuracy. This contrasts with homology-based methods that fail on novel folds.

Performance

ScanNet achieves state-of-the-art accuracy on protein-protein binding site prediction benchmarks.

Performance remains robust across different levels of similarity to training data. While structural homology methods excel when close homologs exist, their accuracy degrades rapidly for distant or novel folds. ScanNet maintains consistent performance across all homology levels, demonstrating true generalization rather than memorization.

Predictions are also robust to conformational changes between bound and unbound structures, with only minor accuracy drops (88.3% to 86.6% on simulated data, 91.9% to 91.3% on experimental structures).