ParaSurf

What is ParaSurf?

ParaSurf is a surface-based deep learning model that predicts paratope binding sites on antibodies. A paratope is the region of an antibody that physically contacts and binds to an antigen—identifying these sites is essential for understanding immune recognition, designing therapeutic antibodies, and developing vaccines.

The model is antigen-agnostic: it only requires the antibody structure, not the antigen. This makes ParaSurf useful for predicting binding sites before experimental characterization or when the target antigen is unknown.

ParaSurf achieves state-of-the-art accuracy on multiple benchmarks, with particularly strong performance on the highly variable CDR3 loops that are critical for antigen specificity. If you don't have an antibody structure yet, you can generate one using ESMFold, Chai-1, or Boltz-2.

How does ParaSurf work?

Surface representation

Rather than operating on raw atomic coordinates, ParaSurf extracts the solvent-accessible surface of the antibody and samples points across this surface. Each surface point becomes a local prediction target.

For each surface point, the model constructs a $41 \times 41 \times 41$ voxel grid centered on that point with 1 Å resolution, covering approximately a 20 Å radius. The grid is aligned perpendicular to the surface normal, which reduces sensitivity to arbitrary rotations.

Input features

ParaSurf encodes 22 features per voxel, grouped into three categories:

Chemical features (18 channels): Nine atom type classes (C, N, O, S, etc.), hybridization state, valence metrics, partial charge, and SMARTS-based descriptors for hydrophobicity, aromaticity, hydrogen bond donor/acceptor properties, and ring membership.

Electrostatic features (4 channels): Force field values from AMBER and CHARMM, plus atomic radii calculated via PDB2PQR.

Geometric features: Van der Waals surface representation with outward-pointing normal vectors.

Neural network architecture

The extracted features pass through a hybrid architecture:

1. 3D CNN blocks extract initial volumetric features
2. Four residual layers with dilated convolutions capture multi-scale spatial patterns
3. Compression layer reduces 2048 channels to 256
4. Transformer block applies self-attention to model long-range dependencies between surface points
5. Global average pooling and a fully connected layer produce a binary binding score (0–1) for each surface point

From surface points to residues

Individual surface point predictions are aggregated to residue-level scores using the maximum:

$$\text{Res}_{\text{score}} = \max(P(sp_1), P(sp_2), \ldots, P(sp_n))$$

where $P(sp_i)$ is the predicted binding probability for surface point $i$ belonging to that residue. Residues with scores above 0.5 are classified as binding sites.

Inputs & settings

Antibody structure

Upload a PDB file containing your antibody structure, or fetch one directly from the RCSB PDB by entering the 4-letter code. The structure should include exactly two protein chains for the antibody heavy and light chains.

Before running ParaSurf, ensure your structure is properly prepared. Use PDB Fixer to add missing atoms, fix non-standard residues, or remove water molecules and ions that may interfere with surface generation.

Model selection

ParaSurf offers nine checkpoints trained on different benchmark datasets. Every checkpoint runs on the submitted two-chain antibody structure:

• Paragraph Expanded: Trained on 1,086 antibody-antigen complexes. We recommend this for general use—it has the largest and most diverse training set.
• PECAN: Trained on 460 complexes from the PECAN benchmark. Use this if you want predictions consistent with PECAN-based literature.
• Paragraph - Heavy Chains Only: Trained on the heavy-chain subset of the Paragraph benchmark. This choice does not restrict inference to the heavy chain.
• Paragraph - Light Chains Only: Trained on the light-chain subset of the Paragraph benchmark. This choice does not restrict inference to the light chain.
• MIPE folds 1-5: Five cross-validation models from the MIPE benchmark. Use these when you need MIPE-specific reproducibility.

Surface mesh density

Controls the resolution of the molecular surface sampling. Lower values (toward 0.1) generate a coarser mesh with fewer surface points, resulting in faster predictions but less detail. Higher values (toward 1.0) produce a denser mesh with more precise predictions at the cost of longer computation time.

For most antibody structures, the default value provides a good balance between accuracy and speed. Increase the density for detailed analysis of specific binding interfaces.

Understanding the results

ParaSurf returns separate residue-level and atom-level PDB files with predicted binding scores from 0 to 1 encoded in the B-factor column. Both scored structures can be used in workflows. Pocket PDB files and pocket centers are also available when binding pockets are found.

The viewer initially colors the structure by chain. Its "Uncertainty / displacement" color option reads the B-factor column. For these scored files, that column contains binding predictions from 0 to 1. Inspect the PDB files when you need exact values, and do not interpret their scores as experimental temperature factors or pLDDT confidence.

The thresholded residue summary lists residues with scores greater than 0.5 in their PDB order. It includes every qualifying residue. Older summary fields retain limited prefixes of this list for compatibility; the field named top_binding_residues is an encounter-order prefix, not a ranking by score.

Evaluation regions

The model predicts across three antibody regions with varying accuracy:

Region	Description	Typical performance
CDR ± 2	Complementarity-determining regions plus 2 flanking residues	Highest accuracy
Fv	Full variable region (all CDRs and framework regions)	High accuracy
Fab	Entire antigen-binding fragment	Good accuracy

Predictions in the CDR regions, especially CDR-H3 (AUC-ROC ~0.96), tend to be the most reliable since these loops are directly responsible for antigen recognition.

Interpreting confidence

A residue score of 0.7 indicates the model is fairly confident that residue participates in antigen binding. Scores near 0.5 represent uncertainty—these residues may warrant experimental validation.

For therapeutic antibody development, focus on residues with scores above 0.6 as candidates for mutagenesis studies or epitope mapping experiments.

Use cases

ParaSurf is particularly valuable in several scenarios:

• Therapeutic antibody engineering: Identify which residues to preserve or modify when humanizing or affinity-maturing antibodies
• Epitope mapping: Predict the antibody side of the interface before conducting experimental mapping
• Antibody selection: Screen computationally predicted antibody structures for favorable binding site characteristics
• Structural biology: Guide experimental design by predicting which regions are most likely to form the binding interface

Limitations

ParaSurf requires a two-chain antibody structure for the heavy and light chains. Single-chain nanobodies, isolated chains, and larger antibody complexes should be prepared or split before submission.

The model was trained on conventional antibody-antigen complexes. Performance on unusual binding modes (e.g., antibodies that bind through framework regions) has not been extensively validated.

Surface generation requires atomic coordinates—ParaSurf cannot process sequence-only inputs. For antibodies with only sequence available, generate a structure first with ABodyBuilder3 or another antibody structure predictor.