Use BioPhi Online
Commercially Available BioPhi No-Code Web Server
BioPhi: Antibody Humanization and Humanness Evaluation with Sapiens and OASis
BioPhi is an open-source platform for antibody design, automated humanization and humanness evaluation built on natural antibody repertoires and deep learning. It was developed by David Prihoda, Jad Maamary, Andrew Waight, Veronica Juan, Laurence Fayadat-Dilman, Daniel Svozil and Danny A. Bitton, with affiliations at the University of Chemistry and Technology in Prague, MSD Czech Republic, Merck and the Institute of Molecular Genetics, and was published in mAbs in 2022.
Humanizing mouse-derived sequences remains one of the main routes to therapeutic antibodies, because sequences from non-human sources are likely to trigger an anti-drug antibody response. Traditionally, humanization is manual, laborious and dependent on expert knowledge. Automated methods have either been shown only at small scale or are proprietary, and existing humanness scores often lack diversity, granularity or interpretability because they rely on small reference libraries or give one score per chain.
BioPhi addresses this with two new methods trained or searched against the Observed Antibody Space (OAS), a large collection of immune repertoire sequencing data: Sapiens for humanization and OASis for humanness evaluation.
Property | Detail |
|---|---|
Purpose | Antibody humanization, humanness evaluation and sequence design |
Humanization engine | Sapiens, a Transformer language model trained on human antibody repertoires of 266 subjects from the OAS |
Humanness score | OASis, an interpretable score based on exact 9-mer peptide search in the OAS |
Interfaces | Interactive web interface plus a command-line interface for bulk processing |
Throughput | Hundreds of sequences per minute through the command line |
Publication | mAbs 2022, 14(1), e2020203; DOI 10.1080/19420862.2021.2020203 |
Why Antibody Humanization and Humanness Scoring Matter for Immunogenicity
Immunogenicity is largely driven by how foreign peptides are processed and displayed to the immune system. BioPhi builds on that idea by breaking an antibody into overlapping 9-mer peptides and asking how common each peptide is across human repertoires. The result is a humanness report that points to specific non-human peptides and residues instead of returning only one opaque number, so protein engineers can see where a sequence is risky and what mutation could fix it.
How BioPhi Works: Sapiens Humanization and OASis Humanness Evaluation
Sapiens humanization
Masked language modeling: Sapiens was trained only on human variable region sequences, in which residues were masked or randomly mutated and the model learned to repair them. No labeled data were needed.
Scale of training: separate heavy chain and light chain models were trained on subsets of about 20 million heavy chain and 19 million light chain sequences from the OAS.
Humanizing a sequence: a mouse sequence is passed through the model, which proposes residue changes toward the human repertoire.
OASis humanness evaluation
9-mer search: every overlapping 9-mer of the antibody is searched in the OAS and its prevalence across human subjects is recorded.
Adjustable thresholds: a peptide counts as human if it appears in enough subjects, and BioPhi offers loose, relaxed, medium (at least 50 percent of subjects) and strict (at least 90 percent) presets in addition to custom values.
Percentile scale: OASis identity can be converted to a 0 to 100 percent percentile relative to therapeutic antibodies from IMGT mAb DB.
BioPhi Benchmark Results: Humanization Quality and Immunogenicity Prediction
Sapiens was benchmarked on 25 humanized antibodies with known parental sequences and 152 with putative parental sequences, 177 in total. With Sapiens run in a single iteration, humanness (OASis medium identity) rose by 30 percent on average and parental sequence identity stayed at 89 percent, comparable to experimental humanization. In eight cases Sapiens differed from the experimental sequence by one mutation, and in one case it was identical. Humanizing the 152 antibodies through the web interface on an 8-core personal computer took 5.2 minutes.
Method (25 known pairs) | Humanness gain (OASis) | Parental preservation | Mutation precision |
|---|---|---|---|
Experimental humanization | +34% | 80% | not applicable |
Sapiens, 1 iteration | +30% | 89% | 77% |
Sapiens, 4 iterations | +34% | 86% | 72% |
Hu-mAb | +14% | 89% | 72% |
Straight CDR graft | +36% | 83% | 65% |
For humanness classification of therapeutic antibodies, OASis medium identity reached 0.966 ROC AUC, ahead of Z-score (0.837) and T20 (0.896), and close to Hu-mAb (0.977). On reported anti-drug antibody responses for 217 therapeutics, OASis medium identity correlated with immunogenicity at R = -0.53 (R2 = 0.28), second only to Hu-mAb (R = -0.58) among the methods compared. Unlike Hu-mAb, OASis is granular and interpretable.
What is Tamarind Bio?
Tamarind Bio is a no-code bioinformatics platform built to give life scientists and researchers access to powerful computational tools. Many cutting-edge machine learning models are hard to deploy and use. Tamarind provides an intuitive, web-based environment that removes the complexity of high-performance computing, software dependencies and command-line interfaces.
The platform is designed for biologists, chemists and other researchers who may not have a background in programming or cloud infrastructure but want to run models on their own data. Key features include:
A user-friendly graphical interface for setting up and launching experiments
A robust API for integration into existing research pipelines
An automated system for managing and scaling computational resources
Tamarind treats information and data security as a top priority, as detailed in its Trust Center and Terms of Service.
Accelerating Discovery with BioPhi on Tamarind Bio
Humanizing murine leads: convert antibodies from immunized mice into human-like candidates with Sapiens while keeping most parental residues.
Humanness screening: score antibody panels with OASis to rank candidates and flag peptides with low human prevalence.
Bulk processing: humanize or evaluate large sets of sequences in one run, as the BioPhi pipeline is designed for.
Manual refinement: use the humanness report to guide point mutations, for example when starting from CDR-grafted sequences.
How to Use BioPhi on Tamarind Bio
Log in: sign in at tamarind.bio and open the BioPhi tool.
Provide antibody sequences: enter or upload variable region heavy and light chain sequences to humanize or evaluate.
Choose the task: select humanization with Sapiens or humanness evaluation with OASis, if exposed.
Set parameters: pick options such as the OASis prevalence threshold or the target germline, if exposed.
Run the job: submit it and wait for the results.
Download outputs: save the humanized sequences and the humanness report.
Interpret the report: look at which 9-mer peptides and positions are non-human and compare candidates.
Use downstream: carry the best candidates into binding, expression and developability testing.
Things to Keep in Mind
The authors note that performance varies substantially from sequence to sequence, and different humanization methods win in different cases.
Humanness is a proxy for immunogenicity. A peptide-based MHC-II metric tested by the authors did not improve prediction and was not pursued.
Sapiens does not optimize developability. The authors suggest enriching training data or pairing it with structure-based developability tools.
Humanized sequences are computational proposals and still need experimental validation of binding and function.
Source: 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):e2020203. DOI 10.1080/19420862.2021.2020203. Code: https://github.com/Merck/BioPhi