---
title: Paragraph on the platform
description: >-
  Predict which antibody residues contact the antigen — the paratope — with a
  graph neural network, using Paragraph.
tags:
  - paragraph
  - antibodies
---

When you humanize or engineer an antibody, the residues that touch the antigen are the ones you must not break. Knowing the paratope — the set of antigen-contacting residues — turns "don't disrupt binding" from a vague constraint into a concrete list of positions to protect. Paragraph predicts that paratope directly from sequence with a graph neural network, folding each antibody into a structure and scoring every residue's probability of being a binding site. It is the paratope-mapping tool in the [antibody engineering](/docs/antibodies) cluster.

## Try it

**Try this prompt** — [open in a guest chat](/chat/new?guest=true&prompt=Predict%20paratope%20residues%20on%20trastuzumab%20(PDB%3A%201N8Z)%20with%20Paragraph%2C%20then%20use%20AntiFold%20to%20redesign%20the%20top-scoring%20CDR%20loop%20%E2%80%94%20which%20residues%20are%20predicted%20as%20antigen-contacting%2C%20and%20recommend%20the%20top%203%20redesigned%20positions%20most%20likely%20to%20improve%20affinity%20for%20Kd%20measurement%20by%20SPR%3F).

> Predict paratope residues on trastuzumab (PDB: 1N8Z) with Paragraph, then use AntiFold to redesign the top-scoring CDR loop — which residues are predicted as antigen-contacting, and recommend the top 3 redesigned positions most likely to improve affinity for Kd measurement by SPR?

_Paragraph maps the paratope; a redesign tool then engineers around those binding residues._

## What Paragraph does

Paragraph takes antibody variable-region sequences, folds each into a paired Fv structure with ImmuneBuilder, and runs a graph neural network (an EGNN) over the structure to predict, per residue, the probability that it belongs to the paratope. It returns the paratope residues, their probabilities, the region each falls in (CDR-H1, FR-H2, and so on), and a mean paratope probability. The result is the map of binding residues you engineer around.

## Citations

> Chinery L, Wahome N, Moal I, Deane CM. Paragraph—antibody paratope prediction using graph neural networks with minimal feature vectors. *Bioinformatics*. 2023;39(1):btac732. doi: [10.1093/bioinformatics/btac732](https://doi.org/10.1093/bioinformatics/btac732)

> Abanades B, Wong WK, Boyles F, Georges G, Bujotzek A, Deane CM. ImmuneBuilder: Deep-Learning models for predicting the structures of immune proteins. *Communications Biology*. 2023;6:575. doi: [10.1038/s42003-023-04927-7](https://doi.org/10.1038/s42003-023-04927-7)

## When to reach for it

- **Paratope mapping** — you want the antigen-contacting residues predicted from sequence.
- **Binding residue identification** — you need to know which positions to preserve during humanization or optimization.
- **CDR engineering guidance** — you want per-residue binding probabilities to guide loop redesign.

Paragraph maps *which* residues bind; it does not measure *how strongly*. For affinity or stability changes, pair it with a redesign or ddG tool. Its output is the guardrail that humanization ([BioPhi](/docs/biophi)) and developability work read against.

## How it fits the workflow

Paragraph is a mapping step whose output constrains the others. Run it before or alongside humanization so you humanize away from the paratope, and cross-check its residues against [Humatch](/docs/humatch)'s flagged positions to avoid rewriting a binding residue. Because it folds each sequence internally, it needs no external structure.

## How it runs

Paragraph runs as an asynchronous GPU job on Modal compute (it folds each antibody before predicting). You submit one or more sequences, the platform tracks and polls the job, and a result card lands on the canvas when it finishes. Allow more time than the CPU tools in this cluster, since folding is part of the run.

## Read next