Humatch on the platform
Score antibody humanness against the human germline repertoire with a deep-learning classifier, per V-gene and per position.
A therapeutic antibody has to look human to the immune system, or it risks provoking an anti-drug response. "How human is this sequence?" is therefore one of the first questions you ask of a candidate — and you want the answer per position, so you know exactly where to intervene. Humatch scores humanness with a deep-learning classifier trained on the human germline repertoire, flags the least-human positions, and names the closest human germline V gene. It is the fast humanness gauge in the antibody engineering cluster.
Try it
What Humatch does
Humatch takes antibody variable-region sequences and scores how closely each resembles the human germline repertoire. Under the hood it aligns each sequence and runs a CNN classifier that predicts V-gene family membership, producing an overall humanness score, per-V-gene scores, per-position scores, and a list of the positions that drag the humanness down. It names the closest human germline V gene and classifies each sequence as human or not against your threshold.
Citations
Chinery L, Jeliazkov JR, Deane CM. Humatch - fast, gene-specific joint humanisation of antibody heavy and light chains. mAbs. 2024;16(1):2434121. doi: 10.1080/19420862.2024.2434121
When to reach for it
- Fast humanness scoring — you want a quick, per-position humanness read on one or many sequences.
- Paired-chain evaluation — you want VH and VL each scored against their matching germline model (one call per chain).
- Humanization quality check — you want to confirm a humanized candidate clears a threshold before lead nomination.
Reach for BioPhi instead when you want the OASis humanness metric and an actual humanization (Sapiens or CDR grafting) in the same step — Humatch scores but does not rewrite the sequence. The two are complementary and are often run together. Humatch chains forward into ADMET-AI prediction and DiffDock docking.
How it fits the workflow
Humatch is a scoring step. A common arc is IgBLAST annotation → Humatch scoring, or OAS search → Humatch scoring: annotate or mine first, then score humanness. Its per-position and flagged-position output tells you where to humanize, which you then act on with BioPhi.
How it runs
Humatch runs as an asynchronous CPU job on Modal compute. You submit one or more sequences, the platform tracks and polls the job, and a result card lands on the canvas when it finishes — typically within two to three minutes.
Read next
How Humatch results compare to the bench
Why V-gene classification accuracy is not an immunogenicity prediction.
Humatch quickstart
Score humanness on a sequence in a few steps.
Humatch inputs
Up to 50 sequences and the chain_type that picks the model.
Humatch outputs
Humanness scores, flagged positions, and closest germline.
BioPhi humanization
Score humanness and humanize in one step.
Was this page helpful?
OAS search quickstart
Run a KA-Search over natural repertoires: set chain, species and region — full, cdr3 or cdr_all — then read ranked matches with identity, study and CDR3.
Humatch inputs and parameters
Up to 50 sequences, the chain_type that picks the heavy or light model — mislabel it and the score is meaningless; the threshold moves only the human call.