Antibody design software that ends at a construct you can clone
Antibody Studio numbers your variable domains, annotates the CDRs in the scheme you choose, finds the sequence liabilities, scores humanness, humanizes by grafting onto the closest human framework, folds the Fv and redesigns a loop, and then back-translates the result into an expression construct that opens in Talindrew's own sequence editor, annotated and ready for the bench.
Numbering, CDRs, liabilities, humanness and humanization are free. Fv structure and CDR redesign run on GPUs and are on the Pro and Max plans. See pricing
The step the annotation tools stop short of
Every antibody tool can tell you where your CDRs are. None of them can hand you a cloning strategy, because none of them have a cloning engine. Talindrew back-translates the antibody you designed into DNA for your expression host, adds the signal peptide that makes it secrete, designs your restriction sites out of it, and drops the result into the same workspace you would use for any other construct.
Every step, and the model that fills it
- 1
Number
ANARCI · IMGT, Kabat, Chothia, Martin, Aho, Wolfguy
Paste one chain or a heavy/light pair, or a whole antibody, and every variable domain is found and numbered. Kappa and lambda are told apart, and a TCR is recognised and refused rather than silently mishandled.
- 2
Annotate CDRs
IMGT · Kabat · Chothia · Contact · AbM definitions
The six loops that do the binding, with framework regions either side. The definition is yours to choose and it is not cosmetic: change it and every boundary, and every count that depends on one, changes with it.
- 3
Call the germline
V-gene identity against the human repertoire
The closest human germline V gene and how far your sequence sits from it: the baseline both humanness and humanization are measured against.
- 4
Sequence liabilities
15 motif rules, each with its evidence
Deamidation, aspartate isomerisation, Asp-Pro and Thr-Ser fragmentation, methionine and tryptophan oxidation, N-linked glycosylation sequons, unpaired and missing cysteines, N-terminal pyroglutamate. Each hit says whether it falls in a CDR and whether the germline already carries it.
- 5
Humanness
Sapiens v1 (BioPhi), run in numpy
A per-position score from a masked language model trained on observed human antibody repertoires. A model's opinion, labelled as one, and never drawn in the colour a liability gets.
- 6
Humanize
CDR graft onto the closest human framework
Graft the loops onto the nearest human framework, see every substitution it made, and keep or revert each back-mutation yourself. If the graft introduces a new liability, the panel says so instead of burying it.
- 7
Fv structure
ABodyBuilder2 (ImmuneBuilder), with OpenMM refinement
Fold the paired variable domains in well under a minute and pin every liability onto the 3D model, so you can see which ones are actually on the surface. Needs the compute service: a deployment without it says so.
- 8
CDR redesign
AntiFold (inverse folding)
Score every residue against that backbone and sample variants of a chosen loop. A high score means structurally plausible; it is not a prediction that binding survives, and the panel says that plainly. Needs the compute service.
- 9
Expression construct
CHO · HEK293 · E. coli · Pichia / S. cerevisiae
Back-translate to DNA with the host's codon preference, add a human Ig heavy-chain or kappa signal peptide so the thing is actually secreted, design out the restriction sites you want to keep free, and cap homopolymer runs. Every codon changed from the host's favourite is reported with its reason.
- 10
Clone it
Talindrew's own cloning engine
The construct opens in the sequence editor as a real sequence in your library, with the leader and all three CDRs annotated on it, ready for a digest, an assembly or a notebook entry. This is the step the annotation tools stop short of.
A CDR is a property of the scheme, not of the molecule
This is the quietest way an antibody pipeline goes wrong: a boundary table applied under the wrong numbering, and every count downstream of it, from which liabilities are in a CDR to which residues a graft keeps and which loop gets redesigned, inherits the error while looking perfectly plausible. Talindrew makes the definition an explicit choice, stores each one in its own native numbering, and re-derives everything from it.
Liabilities, with the context that makes them readable
A list of motifs is easy. What matters is which ones sit in a CDR, where binding lives, and which ones the closest human germline already carries. A site the repertoire tolerates is a different risk from one your engineering introduced. Every hit carries both, plus the evidence for the rule that found it.
Watch it run on the real app
Two and a half minutes, one antibody, trastuzumab's two variable domains, all the way to a construct. Nothing in the film is replayed: every model runs live in the recording. Only the waiting is sped up, and the narration states the real durations.
What the film proves on screen
- ✓Both variable domains found and numbered, with the closest human germline called as IGHV3-66
- ✓The CDR boundaries genuinely move when IMGT is swapped for Kabat
- ✓14 possible liability sites, 7 of them in a CDR, 5 already carried by the germline
- ✓An 8-substitution humanizing graft that reports the one new liability it created
- ✓A 420 bp construct for CHO, designed around 13 restriction sites
- ✓The construct open in the sequence editor with the signal peptide and all three CDRs annotated
Chapters
Antibody design questions
Numbering assigns a scheme's canonical position to every residue of an antibody variable domain, so that the same position can be compared between two different antibodies. Talindrew numbers with ANARCI and supports IMGT, Kabat, Chothia, Martin (extended Chothia), Aho and Wolfguy. Kappa and lambda light chains are distinguished rather than both treated as 'L', and a T-cell receptor is recognised and refused rather than numbered as if it were an antibody.
Because a CDR is a property of the definition, not of the molecule. IMGT, Kabat, Chothia, Contact and AbM each draw the loop differently: on trastuzumab's heavy chain, IMGT's CDR1 spans its native positions 27 to 38, Kabat's 31 to 35 and Chothia's 26 to 32: three different bands over the same residues. Talindrew makes the definition an explicit choice and re-derives everything downstream from it, including which liabilities count as being in a CDR.
A motif that can degrade during expression, purification or storage: asparagine deamidation, aspartate isomerisation, Asp-Pro and Thr-Ser fragmentation, methionine and tryptophan oxidation, N-linked glycosylation sequons, unpaired or missing canonical cysteines, and N-terminal pyroglutamate. Talindrew reports each hit with the motif it matched, whether it falls inside a CDR, and whether the closest germline V gene already carries it. A site the human repertoire already tolerates is a different risk from one your engineering introduced. These are motifs that can degrade, not measurements that they will.
No, and Talindrew keeps them visually separate on purpose. Humanness is scored per position by Sapiens v1, the masked language model published with BioPhi (Prihoda et al. 2022, mAbs), which Talindrew runs in numpy with no GPU. It answers 'how like a human antibody repertoire does this position look', which is a model's opinion. A liability is a chemical motif with a literature basis. Only liabilities are drawn in the alarm colour.
Yes. It grafts the CDRs onto the closest human germline framework, reports every substitution it made, and lets you keep or revert each back-mutation yourself. It also re-runs the liability scan on the result and tells you if the graft introduced a new liability. Sapiens itself offers no humanize function, so the graft is Talindrew's, measured against the Sapiens score.
Fv structures are folded with ABodyBuilder2 from ImmuneBuilder (Abanades et al. 2023, Communications Biology 6:575), with OpenMM refinement, and CDR redesign uses AntiFold (Høie et al. 2025, Bioinformatics Advances 5(1):vbae202). Both run on Talindrew's compute service rather than in your browser; a deployment without that service configured says so rather than returning something fake.
BioPhi and OPIG's SAbPred are excellent at the parts they cover, and Talindrew uses the same published Sapiens, ABodyBuilder2 and AntiFold models where they are the right tool. The difference is the last step: neither can hand you a cloning strategy, because neither has a cloning engine. Talindrew back-translates the designed antibody into an expression construct for your host, designs your restriction sites out of it, and opens it in the same sequence editor and cloning workspace you would use for any other construct.
Numbering, CDR annotation, the germline call, sequence liabilities, humanness and humanization are CPU work and are free. Fv structure prediction and AntiFold CDR redesign run on GPUs and count against the monthly prediction allowance on the Pro and Max plans. Building the expression construct and cloning it are part of the free cloning suite.
Liability motifs, humanness scores, predicted structures and inverse-folding scores rank and flag candidates; none of them is a measurement, and a structurally plausible CDR variant is not a prediction that binding survives. Results are for research and do not constitute medical advice.
Paste a sequence and see where it stands
Numbering, CDRs, liabilities, humanness and humanization are free. No install.
Last updated November 4, 2026