---
title: Antibody engineering
description: >-
  Number, humanize, profile developability, and mine repertoires for
  therapeutic antibodies and AAV gene-therapy vectors on the platform.
tags:
  - antibodies
  - biologics
  - use-case
---

Your program is a biologic, not a small molecule — an antibody lead you need to humanize before it can go to the clinic, or an AAV vector you need to design for a target tissue. The platform's antibody cluster takes a raw VH/VL sequence through the standard therapeutic-development arc: number it to a canonical scheme, score and reduce its immunogenicity, profile whether it can actually be manufactured, map the residues that touch the antigen, and mine billions of natural sequences for closer human frameworks. It is the second-most-used tool cluster on the platform. This page frames how the pieces fit; each tool has its own reference section.

## Try it

**Try this prompt** — [open in a guest chat](/chat/new?guest=true&prompt=Look%20up%20pembrolizumab%20in%20TheraSAbDab%2C%20annotate%20its%20heavy%20chain%20with%20IgBLAST%20for%20V%2FD%2FJ%20germline%20assignment%2C%20run%20ANARCI%20for%20CDR%20numbering%2C%20score%20joint%20VH%2FVL%20humanness%20with%20BioPhi%20and%20Humatch%2C%20search%20OAS%20for%20natural%20human%20heavy%20chains%20within%2090%25%20sequence%20identity%20to%20the%20framework%20regions%2C%20and%20identify%20the%20closest%20human%20germline%20IGHV%20gene%20%E2%80%94%20which%20framework%20positions%20deviate%20most%20from%20germline%2C%20is%20the%20humanness%20above%20the%2080%25%20therapeutic%20threshold%2C%20and%20which%202-3%20back-mutations%20would%20most%20improve%20humanness%20for%20a%20biosimilar%20strategy%3F).

> Look up pembrolizumab in TheraSAbDab, annotate its heavy chain with IgBLAST for V/D/J germline assignment, run ANARCI for CDR numbering, score joint VH/VL humanness with BioPhi and Humatch, search OAS for natural human heavy chains within 90% sequence identity to the framework regions, and identify the closest human germline IGHV gene — which framework positions deviate most from germline, is the humanness above the 80% therapeutic threshold, and which 2-3 back-mutations would most improve humanness for a biosimilar strategy?

_Chains IgBLAST → ANARCI → BioPhi/Humatch → OAS from one prompt — no tool names required._

## The typical workflow

Antibody engineering on the platform follows a recurring order. You rarely run one tool in isolation — the model chains them, but knowing the arc helps you read the results and decide what to run next.

 N["1. NUMBER & ANNOTATE<br/>ANARCI numbering: canonical numbering (IMGT/Kabat/Chothia), CDR & framework regions<br/>IgBLAST annotation: V(D)J germline genes, CDR3, somatic hypermutation"]
    N --> H["2. HUMANIZE & SCORE<br/>Humatch scoring: deep-learning humanness vs human germline repertoire<br/>BioPhi humanization: OASis humanness + Sapiens humanization / CDR grafting"]
    H --> P["3. PROFILE & MAP<br/>Antibody developability: TAP, charge, pI, aggregation, viscosity, liabilities<br/>Paragraph: paratope (antigen-contacting) residue prediction"]
    P --> O["4. MINE THE REPERTOIRE<br/>OAS search: find natural human homologs in Observed Antibody Space"]`}
  caption="The typical antibody engineering workflow, from sequence to repertoire mining."
/>

The numbering step feeds everything after it: both ANARCI numbering and IgBLAST annotation chain into humanness scoring and developability assessment. OAS search is an entry point — you can start there when you want to find natural human frameworks, and it chains forward into IgBLAST annotation and Humatch scoring.

## What each tool answers

- **[ANARCI](/docs/anarci)** — "What are the canonical residue numbers, CDRs, and germline genes?" Standard numbering in IMGT, Kabat, Chothia, AHo, or Wolfguy.
- **[IgBLAST](/docs/igblast)** — "Which V, D, and J germline genes built this sequence, and how much has it hypermutated?"
- **[Humatch](/docs/humatch)** — "How human is this sequence, per V-gene and per position?" Fast deep-learning humanness scoring.
- **[BioPhi](/docs/biophi)** — "How human is it by OASis, and can you humanize it while preserving binding?" Humanness plus Sapiens humanization.
- **[Antibody developability](/docs/antibody-developability)** — "Can this candidate actually be manufactured?" TAP score, charge, pI, aggregation, viscosity, sequence liabilities.
- **[Paragraph](/docs/paragraph)** — "Which residues contact the antigen?" Graph-neural-network paratope prediction.
- **[OAS search](/docs/oas-search)** — "Are there natural human antibodies close to this one?" Search across billions of Observed Antibody Space sequences.
- **[AAV design](/docs/aav)** — the gene-therapy sibling: select a capsid serotype for a target tissue, then optimize the transgene cassette for AAV packaging.

## How they run

Every tool in this cluster runs as an asynchronous job on Modal compute — you submit, the platform tracks the job and polls it, and a result card lands on the canvas when it finishes. Most antibody tools are CPU jobs that return in a couple of minutes; [Paragraph](/docs/paragraph) and the [AAV](/docs/aav) tools use a GPU because they fold or codon-optimize, and [OAS search](/docs/oas-search) is the outlier on the CPU side — it can take tens of minutes at the larger database tiers. Sequence-based tools take one or more variable-region sequences (heavy and/or light), and several also accept a `therapeutic_name` that resolves VH/VL automatically from Thera-SAbDab.

## Explore the tools

## Read next