---
title: BioPhi on the platform
description: >-
  Score antibody humanness with OASis and humanize sequences with the Sapiens
  deep-learning model or CDR grafting using BioPhi.
tags:
  - biophi
  - antibodies
---

Scoring humanness tells you a candidate is too murine; the next thing you need is a sequence that fixes it without breaking the CDRs. BioPhi does both — it scores humanness with the OASis metric and then humanizes the sequence, either with the Sapiens deep-learning model or by classic CDR grafting, and the Sapiens path can hold binding residues fixed while it rewrites the rest. It is the humanize-and-score workhorse of the [antibody engineering](/docs/antibodies) cluster.

## Try it

**Try this prompt** — [open in a guest chat](/chat/new?guest=true&prompt=I%20have%20a%20murine%20antibody%20candidate%20(use%20any%20TheraSAbDab%20clinical%20mouse%20mAb%20as%20a%20stand-in)%20%E2%80%94%20walk%20me%20through%20humanizing%20it%20without%20disrupting%20the%20CDRs.%20Score%20the%20current%20humanness%2C%20find%20the%20closest%20natural%20human%20heavy-chain%20frameworks%2C%20recommend%20specific%20framework-region%20back-mutations%20toward%20germline%2C%20and%20tell%20me%20how%20each%20mutation%20affects%20predicted%20thermostability.%20I%20want%20a%20final%20pick%20of%202%E2%80%933%20back-mutations%20that%20improve%20humanness%20AND%20Tm%20before%20lead%20nomination.).

> I have a murine antibody candidate (use any TheraSAbDab clinical mouse mAb as a stand-in) — walk me through humanizing it without disrupting the CDRs. Score the current humanness, find the closest natural human heavy-chain frameworks, recommend specific framework-region back-mutations toward germline, and tell me how each mutation affects predicted thermostability. I want a final pick of 2–3 back-mutations that improve humanness AND Tm before lead nomination.

_BioPhi scores humanness and humanizes; the platform chains numbering, repertoire mining, and stability around it._

## What BioPhi does

BioPhi is one backend behind three separate tools, one per mode — which tool you call is what selects the mode; it isn't a parameter you set. `biophi_humanness` scores a VH/VL pair against the OASis database, reports an OASis percentile, classifies the antibody as human, humanized, chimeric, or murine, and breaks the humanness down per CDR and per framework region. `biophi_humanize` uses the Sapiens deep-learning model to rewrite the sequence toward a target humanness, returning the humanized VH/VL and the mutations it made, and can hold known binding residues fixed while it does. `biophi_graft` humanizes by classic CDR grafting.

## Citations

> Prihoda D, Maamary J, Waight A, Juan V, Fayadat-Dilman L, Svozil D, Bitton DA. BioPhi: A platform for antibody design, humanization, and humanness evaluation based on natural antibody repertoires and deep learning. *mAbs*. 2022;14(1):2020203. doi: [10.1080/19420862.2021.2020203](https://doi.org/10.1080/19420862.2021.2020203)

## When to reach for it

- **Humanization** — you have a murine or chimeric candidate and need a humanized sequence.
- **OASis humanness assessment** — you want the OASis percentile and human/humanized/chimeric/murine classification.
- **Immunogenicity reduction** — you want to lower predicted immunogenicity while keeping the CDRs intact.

Reach for [Humatch](/docs/humatch) instead when you only need a fast humanness *score* and don't need the sequence rewritten. The two are complementary: Humatch and BioPhi are frequently run together for a two-metric humanness read. BioPhi chains forward into `admet_ai_predict` and `diffdock_dock`, and pairs with `anarci_number` and `antibody_developability`.

## How it fits the workflow

BioPhi is the act-on-it step after scoring. A typical arc is `anarci_number → biophi_humanize → antibody_developability`: number the sequence, humanize it, then confirm the humanized candidate is still manufacturable. Because humanization can shift charge and aggregation, following BioPhi with a developability check is the standard guardrail.

## How it runs

BioPhi runs as an asynchronous CPU job on Modal compute. You submit a VH (and optionally VL) sequence, the platform tracks and polls the job, and a result card lands on the canvas when it finishes — typically in tens of seconds to a couple of minutes.

## Read next