Antibody aggregation and developability: what can you predict, what must you measure?

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

Antibody aggregation, thermal stability and immunogenicity all appear in the same developability report, usually in the same table, usually with a number beside each one. They do not have the same standing. Some entries are read directly from the sequence and are as certain as arithmetic. Some are predictions that correlate with experimental behaviour and are sometimes wrong. And some cannot be computed at all, so a number beside them came from an assay or it came from nowhere. Reading a developability report well means knowing which of the three you are looking at.

What is certain, what is estimated, and what needs an experiment

• Read from the sequence: no model and no prediction are involved, because the property is a pattern that inspection either finds or does not. Where the pattern is present, the risk is worth checking, although whether it becomes a liability depends on how exposed and how flexible the site is.
• Predicted: a model estimates a physical behaviour from sequence and structure. The estimate correlates with reality well enough to be useful for ranking, and it is wrong often enough that it cannot be reported as a result.
• Measured only: the property depends on conditions, concentration and time in ways that no current model captures, so either an experiment was run or the number is decorative.

Read directly from the sequence: motifs and counts

These are certain, in the sense that the motif is either present or absent:

• Deamidation motifs: asparagine followed by glycine is the classic case, with serine and threonine close behind. The motif degrades over time, converting the residue and altering the molecule, and finding it requires nothing more than looking.
• Isomerisation motifs: aspartate followed by glycine, on the same logic.
• Oxidation-prone residues: exposed methionine and tryptophan. Their presence is read from the sequence, while establishing their exposure needs the structure.
• Unpaired cysteines: a matter of counting. An odd cysteine carries a risk of disulfide scrambling and dimerisation.
• N-linked glycosylation sequons: asparagine, any residue except proline, then serine or threonine. The pattern is there or it is not.

What makes these worth attention is position rather than presence. A deamidation motif in a framework loop is a minor liability, whereas the same motif inside CDR3 means the binding site itself degrades on the shelf, so that the molecule you characterised and the molecule in the vial six months later are no longer the same thing.

For a designed sequence these are also the cheapest liabilities to remove, since they can be excluded during generation rather than engineered out afterwards, which is the argument in developability by design.

Predicted, antibody aggregation propensity among them

This is the largest category and the one most often read as though it belonged to the first.

• Antibody aggregation propensity: computed from predicted aggregation-prone regions, exposed hydrophobic patch area and charge distribution on the modelled surface. These descriptors correlate with experimental aggregation without predicting the concentration at which a given molecule aggregates in a given buffer, and aggregation is a behaviour that emerges at concentration rather than a property visible in dilute solution.
• Surface hydrophobicity and patch analysis: entirely dependent on the structural model. For a designed molecule that model is itself a prediction of a structure nobody has solved, which makes this a prediction resting on a prediction.
• Charge and isoelectric point: the theoretical isoelectric point is arithmetic, whereas how the molecule behaves at a given pH is not, since the local environment shifts the behaviour of individual residues.
• Immunogenicity risk: estimated from framework composition and predicted peptide presentation on MHC class II alleles. It is useful for comparing candidates against each other, and it tells you nothing about whether a given patient will mount a response, which depends on dose, route, duration and the individual.
• Polyreactivity: predicted from the distribution of positive charge and hydrophobic patches. The correlation is real, but the instrument is crude.

Published profiling methods make this explicit. They position a candidate against therapeutic antibodies that had reached post-phase-I clinical stage, so that an outlier is flagged for attention rather than condemned. The guideline ranges derived for conventional antibodies come from clinical-stage molecules (Raybould et al., PNAS 2019, DOI, reference located through PubMed), and the equivalent for single domains is calibrated on the 36 nanobody-based drugs that have reached clinical stage, tested against experimental data on 108 nanobodies (Gordon et al., Communications Biology 2026, DOI, reference located through PubMed).

These methods operate as flagging systems, and the two ways of misreading them are treating a flag as a verdict and discarding a good candidate, or treating a clean profile as a clearance and skipping the assay.

What only an experiment settles: antibody aggregation at concentration, titre, melting point

• Expression yield: no model predicts titre reliably. A sequence either produces protein in your system at a usable level or it does not, and this is the first thing worth establishing, which is why candidates are expressed before they are delivered.
• Melting temperature: differential scanning fluorimetry or calorimetry. Predicted stability scores will rank candidates against one another without yielding a figure in degrees that can go into a development plan.
• Aggregation at concentration, over time: accelerated stability studies, size exclusion chromatography and light scattering. This is the property that matters most for formulation and the one that prediction is furthest from reaching.
• Real binding affinity: surface plasmon resonance or biolayer interferometry. A design score ranks candidates and does not yield a dissociation constant, as set out in what a binding experiment proves.
• Specificity in a realistic context: counter-screening against a panel or against cell lysate. Predicted polyreactivity narrows the field without establishing the result.
• Behaviour in your formulation: excipients, pH, surfactant and freeze-thaw, none of which any model takes as input.

How to read a developability report without being misled

Four questions separate a report that means something from a table of plausible numbers:

• Category of each line: is it a measurement or a prediction? A report that presents melting temperature and deamidation motifs in the same column with the same formatting is concealing the distinction that matters most.
• Reference used: were the descriptor ranges derived from single domains or from conventional antibodies? Antibody-derived ranges are not suitable for direct application to a single domain, because the structural basis differs: no pairing interface, a different framework and a longer CDR3. That point is developed in our note on why antibody programs fail.
• Structure used: what model was it computed on, and how confident is it? A predicted model of modest confidence propagates into every surface descriptor downstream of it.
• Measurement: was anything measured at all? A report that is entirely computational is a filter, and it should be described as one.

A report worth reading keeps these three categories apart on the page. At Immunochem the computable part is applied to the whole candidate set during selection, so that it filters rather than judges; expression and binding are measured on the shortlist and reported as measurements; and the full experimental panel, covering melting temperature, accelerated stability and counter-screening, belongs to the stage that follows, and we say so rather than letting the filter appear to replace it.

Receive a feasibility study on your antigen - for free

Frequently asked questions

Can antibody aggregation be predicted from sequence?

Partly. Aggregation-prone regions, hydrophobic patch area and charge distribution are computed from sequence and a structural model, and they correlate with experimental behaviour. They do not tell you the concentration at which a given molecule aggregates in a given buffer. That needs an experiment.

Which developability properties are certain from the sequence?

The chemical liability motifs: deamidation, isomerisation, oxidation-prone residues, unpaired cysteines and glycosylation sequons. These are patterns rather than predictions. What matters is where they sit, since a motif inside a CDR is a different problem from the same motif in a framework loop.

Can immunogenicity be predicted?

Only as a risk estimate, from framework composition and predicted MHC class II presentation. It is useful for comparing candidates and it is not a prediction of clinical outcome, which depends on dose, route, duration and the individual.

Are the predictions different for nanobodies?

Yes, and this is a common error. Guideline ranges derived from conventional antibodies do not apply to single domains. A nanobody-specific profiler exists, and the antibody version is not suitable for direct application to a single domain.

If so much has to be measured, what is the computational filter for?

Ordering the work. It removes candidates that would fail for reasons visible before synthesis, so that the expensive assays run on a filtered set. It replaces nothing, and a report that presents it as a clearance is overstating it.

What do you measure, and what do you predict?

We compute the sequence-level liabilities and the predicted surface descriptors across the candidate set during selection, and we measure expression and binding on the shortlist. Melting temperature, accelerated stability and counter-screening are the next stage and belong to whoever takes the lead forward.