Eight questions to ask a nanobody discovery service before signing

Written by the Immunochem scientific team and reviewed by Philippe Herman, CEO and Chief Scientific Officer.

Many providers of a nanobody discovery service describe the same models, the same timelines and the same promise, because the design tools are public and most of them run broadly the same ones. The questions below separate them anyway, because each one has a specific and checkable answer that is awkward to give if the work behind it is not being done. They are the questions we would ask, and the answers include ours.

1. What happens if nothing binds, and how will I know?

Design campaigns have a per-attempt success rate rather than a guarantee, and some targets yield nothing. A published independent evaluation of one design tool found binders against two targets and none at all against two others, with the same method competently applied.

A good answer describes the attrition: how many designs were generated, how many survived each filter, how many were tested, how many bound. A provider that reports only successes is reporting half the experiment.

Ours: it happens, you know by week twelve, and the funnel is reported in full. The reasoning is in our note on what a binding experiment proves.

2. Will anything be made and tested, or am I receiving sequences?

This is the single largest difference between offers, and it is often left ambiguous.

A design pipeline always returns candidates, ranked, whether or not any of them work. Expressing those candidates and testing them for binding is a separate stage with a real cost, and it is the stage where candidates get eliminated.

Ask what is made, in what expression system, against what controls, and what the report contains. If the answer is that sequences are delivered with scores, that is a legitimate offer and it is a different product, and the price should reflect it.

Ours: candidates are expressed in a cell-free system and tested by ELISA before delivery. A design-only engagement is possible, and we state it explicitly in the quotation rather than letting the two be confused.

3. Which epitope will you target, and how was it chosen?

Epitope control is the main thing computational design offers over selection-based routes. A provider who cannot tell you which site they are designing against is not using the capability.

A good answer names the site, explains why it suits what the binder has to do, and says what makes it designable or difficult: shape, chemistry, rigidity, conservation, glycosylation.

Ask also what they would do if the site you want turns out to be undesignable. The answer tells you whether the assessment is real or decorative. Our criteria are in choosing which site to target.

4. What does a nanobody discovery service need from you, and what does it not do?

Scope creep in reverse is the common problem: a provider who implies they cover more of the chain than they do.

Ask whether they need physical antigen or only a structure, whether they run functional assays, whether they do epitope mapping experimentally, whether they handle reformatting, and whether they manufacture.

Ours, including the uncomfortable parts: we need a structure, experimental or predicted, and we will say when confidence in it is too low. We do not run experimental epitope mapping or functional assays, we do not build VHH-Fc or albumin fusions, and we do not manufacture. We have never run an immunisation and we hold no library, so on a flexible target with no usable structure a display route is the better option and we will tell you so.

5. Who owns the sequences, and what do you keep?

Ask it directly, and ask what the provider retains.

The distinction that matters is between the molecule and the method. A provider keeping their tooling and their filtering criteria is normal. A provider retaining rights over the molecules designed for you, or reserving the right to reuse them, is a different proposition and you should know which one you are signing.

Ask also what you receive beyond the sequences: the design rationale, the structural models, the developability file, the raw binding data. Those are what let a third party, or your own team later, judge the work.

Ours: the sequences, the data and any intellectual property arising are yours. Our methods and filtering criteria remain ours. The detail is in our common questions.

6. When is developability assessed, during selection or after?

This one is diagnostic, because the answer reveals the order the work is actually done in.

Assessing developability after a lead has been selected is the conventional sequence, and it is the point at which a liability becomes expensive: you either engineer it out and risk the affinity you selected for, or go back to the panel.

Applying the same filter across the whole candidate set before the shortlist is drawn costs compute rather than laboratory time, and it changes which candidates survive.

Ask which it is, and ask to see the filter applied to the candidates that were rejected. The argument is in developability by design.

7. Which developability metrics, and were they calibrated for single domains?

This technical question separates people who run the tooling from people who merely quote it.

Descriptor ranges derived from conventional antibodies do not transfer to single domains, because the structural basis differs: no pairing interface, a different framework, a longer CDR3. An antibody profiler is not suitable for direct application to a VHH.

Ask which reference set the guideline values come from. A single-domain-specific profiler exists: the Therapeutic Nanobody Profiler, calibrated on the 36 nanobody sequences that have reached clinical stage and benchmarked against experimental data on 108 nanobodies (Gordon et al., Communications Biology 2026, DOI, reference located through PubMed). A provider who cannot name the reference set is quoting antibody numbers at a single domain.

Ask too which lines in the report are computed, which are predicted, and which were measured. The three have different standing and are often presented in the same table, which is the subject of predicting or measuring.

8. What would make you tell me not to do this?

This is the question that is hardest to answer well, and the most informative.

A provider who can describe the targets they turn down, and why, is applying a real assessment. A provider for whom every target is suitable is selling capacity.

Ask for the criteria rather than for reassurance: what makes a site undesignable, what structural confidence is too low, which target classes they would send elsewhere.

Ours: no usable structure and none predictable with confidence, a disordered or highly flexible site, a surface dominated by glycosylation, or a requirement for maximum affinity irrespective of epitope, where a display route does better. We say no during the free feasibility study, before anyone has committed anything, which is the whole reason it is free.

What not to weigh when comparing nanobody discovery services

• The model names: RFdiffusion, ProteinMPNN, RFantibody and the rest are public, and every nanobody discovery service runs much the same stack. The differences between providers lie in site assessment, filtering and validation, as set out in our note on which tools do what.
• Affinity figures quoted without context: a single campaign against one target says very little about the method behind it. In the RFantibody study, initial designs started in the tens to hundreds of nanomolar and needed an optimisation step, which is a normal pattern rather than a disappointing one.
• Timelines quoted without a deliverable: six weeks to sequences means one thing and twelve weeks to tested binders means another, while a number quoted without a defined deliverable means nothing at all. The full picture is in how long antibody development takes.

Receive a feasibility study on your antigen - for free

Frequently asked questions

What should I ask a nanobody discovery service first?

Whether anything is expressed and tested before delivery, or whether you receive ranked sequences. It is the largest difference between offers and it is often left ambiguous.

How do I know if a nanobody discovery service's success rate is real?

Ask for the attrition rather than the successes: designs generated, candidates tested, candidates that bound. A rate quoted without the denominator is not a rate.

Is it a bad sign if a provider says no to my target?

The opposite. A provider who assesses targets and declines some is applying a real filter. One for whom every target is suitable is selling capacity.

Should I expect to own the sequences?

Yes, for a design service. The provider keeping their methods and tooling is normal. Retaining rights over your molecules is a different proposition and should be explicit in the contract.

Does it matter which AI models a provider uses?

Less than it appears. The main tools are public and widely used. Site selection, filtering criteria and whether anything is measured matter considerably more.

What is a reasonable timeline?

For computational design, our working timeline is six weeks to candidate sequences and twelve weeks to validated binders. A provider quoting substantially less usually delivers less, so check the deliverable rather than the number.