---
title: Paragraph quickstart
description: Paste a sequence and set the chain type; Paragraph folds the antibody first, then returns paratope residues with a probability and region for each.
tags:
  - paragraph
  - quickstart
---

You have an antibody sequence and you want to know which residues touch the antigen — the paratope — so you can engineer around them. [Paragraph](/docs/paragraph) predicts it from sequence: paste the sequence, set the chain type, submit, and read the paratope residues and their probabilities off the result card.

## Predict a paratope

### Provide sequences

Paste one or more antibody variable-region sequences (up to 20 per job). Name them if you want the labels carried through.

### Set the chain type

Choose `H` for heavy chains or `L` for light. Keep the default `imgt` numbering scheme — `chothia` is rejected with an error, not silently applied — and set the probability threshold that decides which residues count as paratope.

### Submit and wait

Paragraph runs as an asynchronous GPU job on Modal — it folds each antibody before predicting, so allow more time than the CPU tools. A result card appears on the canvas when the job finishes.

### Read the paratope

The card shows the predicted paratope residues, each residue's probability and region, the count of paratope residues, and the mean paratope probability.

## In chat

Describe the antibody and ask for the binding residues — the model routes to Paragraph and chains a redesign tool when your prompt asks to engineer the loop.

**Try this prompt** — [open in a guest chat](/chat/new?guest=true&prompt=Run%20Paragraph%20paratope%20prediction%20on%20the%20bevacizumab%20Fab%20(PDB%3A%201BJ1)%2C%20then%20redesign%20CDR-H3%20with%20AntiFold%20%E2%80%94%20does%20the%20redesigned%20loop%20preserve%20the%20predicted%20binding%20residues%2C%20and%20select%20the%20top%20AntiFold%20variant%20for%20experimental%20expression%20based%20on%20sequence%20recovery%20at%20the%20paratope%20positions%3F).

> Run Paragraph paratope prediction on the bevacizumab Fab (PDB: 1BJ1), then redesign CDR-H3 with AntiFold — does the redesigned loop preserve the predicted binding residues, and select the top AntiFold variant for experimental expression based on sequence recovery at the paratope positions?

## What to do next

- **Protect the paratope during humanization** — keep [BioPhi](/docs/biophi)'s `biophi_humanize` binding-residue preservation on (the default) so it doesn't rewrite the residues Paragraph flagged.
- **Cross-check humanness** — compare the paratope against [Humatch](/docs/humatch)'s flagged positions.
- **Number the sequence** — run [ANARCI](/docs/anarci) if you need canonical position labels.

## Read next