Use Humatch Online
Commercially Available Humatch No-Code Web Server
Humatch: Fast, Gene-Specific Joint Humanization of Antibody Heavy and Light Chains
Humatch is a fast human classification and joint humanization tool for antibody heavy and light chains. It was developed by Lewis Chinery, Jeliazko R. Jeliazkov and Charlotte M. Deane (Oxford Protein Informatics Group) and described in a 2024 bioRxiv preprint. The code is open-source on GitHub under the BSD-3-Clause license.
Therapeutic antibodies often start as non-human sequences and must be humanized to reduce immunogenicity. Humanizing the heavy and light chains separately can break their pairing, hurting expression and stability.
Humatch addresses this by humanizing both chains together, keeping expression and stability high while steering the sequences toward single human V-genes.
Property | Detail |
|---|---|
Task | Human classification and joint humanization of VH/VL pairs |
Models | Three CNNs: CNN-H (heavy), CNN-L (light), CNN-P (pairing) |
Input encoding | Kidera feature vectors over IMGT-numbered positions |
Training data | Observed Antibody Space (OAS), with artificial badly paired sequences for CNN-P |
Speed | About 36 s per pair on a 16-CPU desktop |
License | BSD-3-Clause |
Why Joint Antibody Humanization Matters
Tools that treat chains independently can produce sequences that look human but pair poorly, or that mix genes from several germlines. Humatch scores pairing explicitly and works toward a single human gene per chain, which the authors link to better stability and lower anti-drug antibody risk.
How Humatch Works
Germline-likeness matching: a fast first stage moves each chain toward a germline-likeness target of 0.40.
Single-point scoring: all single-point variants are scored with the CNNs and weighted by germline frequency.
Iteration: mutations are applied until CNN scores reach 0.95 or a maximum of 60 edits is hit.
CDR handling: CDRs are excluded from mutation by default.
The CNNs use 40 convolutional filters (kernel size 10), dropout of 20 percent, max pooling and a sigmoid output, trained with Adam and binary cross-entropy.
Humatch Performance: Classification, Stability and Humanization Results
Evaluation | Result |
|---|---|
CNN-P pairing classifier | PR AUC and ROC AUC both above 0.99 |
Stability (137 therapeutics) | High CNN-P scores (above 0.5) linked to higher melting temperatures (p < 0.001) |
ADA association (211 therapeutics) | A 0.01 cut-off removed 89% of high-ADA, 59% of medium-ADA and 23% of low-ADA therapeutics |
Overlap with experimental humanization (25 therapeutics) | 77% of heavy and 82% of light chain mutations; Pearson 0.89 for mutation counts |
Mixed-gene designs | 51% ranked non-human vs 28% AbNatiV, 8% Hu-mAb, 1% OASis |
On 1,000 OAS sequences, 12 pairs failed to humanize within the edit limit, with mean edits of 11.6 (heavy) and 13.9 (light).
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 Humatch on Tamarind Bio
Humanize leads: convert non-human VH/VL pairs into human-like sequences with few edits.
Classify humanness: get V-gene predictions and CNN scores for existing sequences.
Screen for developability risk: use pairing scores as a stability signal and humanness as an ADA-related signal.
Protect key residues: specify residues that must not be mutated, where supported.
How to Use Humatch on Tamarind Bio
Open the tool: log in to tamarind.bio and select Humatch.
Enter sequences: provide the complete VH and VL sequences; the authors state only complete variable-region sequences should be used.
Choose the mode: classification or humanization, if exposed.
Set options: target genes or residues that cannot be mutated, if exposed.
Run the job: submit it; each pair takes tens of seconds on the reference setup.
Download outputs: humanized sequences, edit counts, V-gene assignments and CNN-H, CNN-L and CNN-P scores.
Interpret scores: scores near the 0.95 target indicate a human-like, well-paired result.
Validate: check binding and expression experimentally.
Things to Keep in Mind
Sparse training data for some V-genes, such as KV7, lowers classification accuracy.
Artificial bad pairs may be noisy, which could cause false positives and negatives in CNN-P.
The ADA analysis is retrospective and uses a single dataset.
Humatch makes the second-largest number of mutations among the tools compared.
Inputs must be complete VH/VL sequences.
Source: Chinery, Jeliazkov and Deane, "Humatch - fast, gene-specific joint humanisation of antibody heavy and light chains", bioRxiv (2024), DOI 10.1101/2024.09.16.613210. Code: github.com/oxpig/Humatch.