Why do antibody programs fail before they reach the clinic?

Written by the Immunochem scientific team and reviewed by Philippe Herman, CEO and Chief Scientific Officer. Published 13 October 2026, updated 13 October 2026.

What are the main causes of failure?

It depends entirely on where in the programme you are standing, and conflating the stages is the most common mistake in this discussion.

In the clinic, candidates fail mostly for lack of efficacy. The molecule works as designed, engages its target, and does not change the disease. That is a biology problem, and no amount of protein engineering fixes it, because the hypothesis about the target was wrong.

Before the clinic, the picture is different. Almost nothing is lost to efficacy, because efficacy has not been tested. What removes candidates is binding first and manufacturability second.

Keeping these apart matters, because the fixes are unrelated. Choosing a better target addresses clinical failure. Better molecular design addresses preclinical failure. Doing the second well does not protect you from the first.

The first filter: it does not bind well enough

This is where the volume is lost, and it is worth being plain about it.

A discovery campaign produces a large number of candidate molecules and most of them are not good enough. They bind weakly, or they bind the wrong epitope, or they cross-react with something they should not, or they bind the recombinant antigen and not the protein in its native context. The ratio between molecules generated and molecules worth keeping is unflattering in every discovery method, including ours.

What changes between methods is how early and how cheaply you find out. Screening a library tells you after the screen. A computational campaign tells you at the scoring stage, which is faster, but a score is a prediction and not a measurement, and a pipeline will always return something that scores well. This is why we express the shortlisted candidates in a cell-free system and test them by ELISA before delivering anything: the question "does it actually bind" deserves an answer from an experiment, not from a model.

The second filter: it binds but cannot be made

A molecule can have excellent affinity and still be undevelopable. The usual causes are these.

Aggregation. The molecule clumps at the concentrations needed for formulation, or during freeze-thaw, or on storage. This is the single most common developability liability and it can kill a candidate late, after real money has been spent.

Poor expression. The sequence does not produce enough protein in the chosen system to be viable at scale, which is a cost problem that becomes a feasibility problem.

Instability. Low thermal stability, sensitivity to pH, or chemical degradation hotspots such as deamidation and oxidation sites in the CDRs.

Immunogenicity risk. Sequence features likely to provoke an anti-drug antibody response, which can neutralise the therapeutic and create safety issues.

Specificity problems that only appear later, such as polyreactivity or binding to unrelated human proteins, which may not show up against a purified antigen.

These are the expensive failures, because they are usually discovered after a candidate has been selected and work has been built on top of it.

Why developability is the avoidable one

Developability liabilities are, to a useful extent, visible in the sequence and the structure before anything is made. That is what makes them different from the other failure modes.

For conventional antibodies this has been computable for some years. For nanobodies it has not, and the reason matters: the structural differences between a single domain and a paired format make the antibody methods unsuitable for direct transfer. Running an antibody tool on a VHH returns a number that looks credible and is not.

The Oxford group behind the Therapeutic Antibody Profiler has since published the Therapeutic Nanobody Profiler, which follows the same principles but is calibrated for single domains. Its guideline values come from the 36 sequences of nanobody-based drugs that have reached clinical stage, and the authors tested them against experimental developability data collected on 108 nanobodies. It is open source and available as a web application (Gordon et al., Communications Biology, 2026, DOI, reference located through PubMed).

Two things are worth noting honestly. It is a flagging system rather than a verdict: a candidate outside the guideline range is not disqualified, it is one to look at harder. And it is a profiling method, not a study of why programmes fail, so it should not be cited as evidence that developability is the leading cause of attrition. What it establishes is narrower and still useful: for the format we actually design, a meaningful part of developability risk can be computed from sequence and structure, before synthesis rather than after.

What you can check before committing anything

Three things, and all of them before a molecule exists.

Whether the site is bindable at all. Not every epitope is a realistic target. Flat, featureless or highly polar surfaces are hard, and knowing that in advance is worth more than discovering it after a campaign. We score this for a specific site in a free feasibility study before any work is committed.

Whether the candidates carry obvious liabilities. Run the developability assessment on the designed sequences and rank them on it, rather than ranking on affinity alone and checking developability at the end. Designing against both at once costs very little; retrofitting a fix into a selected candidate costs a great deal.

Whether they bind in an experiment. Before you build anything on a sequence, have something measure it.

What this changes about how a programme is run

The practical conclusion is about ordering, not about effort.

The conventional sequence is to optimise for affinity, select a candidate, then assess developability and discover the problem. The alternative is to carry both criteria from the start, so that the candidate you select has already been filtered on both. This is what "developability by design" means, and it is less a technique than a decision about when to apply a filter you were going to apply anyway.

It does not protect you from the first filter, and it does not protect you from clinical failure. What it removes is the category of loss where a good binder turns out to be unusable, which is the one category in this list that was avoidable from the beginning.

Frequently asked questions

Is developability the main reason antibody programmes fail?

No, and the claim is often stated too strongly. Most attrition before the clinic happens at the binding stage, and most clinical failure is lack of efficacy. Developability is a smaller category than either, but it is the one most often avoidable, which is why it receives attention out of proportion to its share.

Can developability really be predicted from sequence?

Partly, and partly is enough to be useful. Computational metrics flag candidates that merit closer examination, and they reliably catch the common liabilities such as hydrophobic patches and extreme charge distributions. They do not replace experimental characterisation, and nobody claims they do.

Does this apply to nanobodies as well as full antibodies?

Yes, and that is exactly why the choice of tool matters. Single-domain antibodies tend to be more stable and less aggregation-prone than paired formats, since there is no light chain to mispair, but they carry their own liabilities and the antibody profilers do not transfer to them. For a nanobody, the nanobody-specific profiler described above is the one to use.

When should developability be assessed?

During design, not after candidate selection. Once a candidate is chosen and work is built on it, a developability problem becomes a redesign rather than a filter.

Why do you validate in the laboratory if the predictions are good?

Because a prediction is not a measurement. A scoring function will always rank something first. Whether the molecule binds is a fact about the world, and it costs far less to establish it at the shortlist stage than after a programme has been built around it.