Where does computational design fit into the antibody discovery workflow?

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

The antibody discovery workflow is usually described in eight stages: target validation, antigen design, hit generation, screening, engineering and optimisation, functional characterisation, developability assessment, and preclinical translation. Computational design does not add a ninth stage but replaces the third one, and in doing so it changes three of the others, because several steps in the standard workflow exist to compensate for the fact that you could not specify what you wanted in the first place. Knowing which stages change, and which do not, is the practical way to work out what design is worth to a given programme.

The eight stages of a standard antibody discovery workflow

• Target validation establishes that the antigen matters, through expression profiling, knockdown or genetic evidence. This stage is biology and involves no molecules of your own.
• Antigen design and production supplies the protein you will work against, as a recombinant construct, a virus-like particle or a peptide, in a form resembling the native target.
• Hit generation produces candidate binders, conventionally by immunising an animal and recovering hybridomas or B cells, or by in vitro display from a phage, yeast or mammalian library.
• Screening and hit selection reduces that output to a workable set, using ELISA, flow cytometry and kinetic measurement.
• Engineering and optimisation supplies what the hits lack, through humanisation, affinity maturation, Fc engineering and reformatting into bispecifics or conjugates.
• Functional characterisation tests whether the binding does anything, in cell-based assays matched to the intended mechanism.
• Developability assessment profiles stability, aggregation, solubility and sequence liabilities to establish manufacturability.
• Preclinical translation takes the lead into animal models, pharmacokinetics and regulatory preparation.

That sequence is sound, and computational design leaves most of it in place.

The one stage of the antibody discovery workflow that design replaces

Hit generation is the stage that changes, and the change is in where the candidates come from.

Immunisation and display are both selection processes, in which diversity is produced and then enriched for binders. Design works by specification instead: a structure of the target and a chosen site go in, and sequences predicted to engage that site come out, with no repertoire, no library and no selection step in between.

The practical consequence is that you name the epitope rather than discovering which one you got. Everything else that changes downstream follows from that single difference, which is set out in detail in our comparison of phage display and computational design.

Three stages it changes without replacing: antigen, engineering, developability

• Antigen design and production becomes partly optional, while structure becomes mandatory: in the standard workflow you need physical antigen to immunise with or to pan against, and producing a difficult antigen in a native-like form is often the hardest part of the whole project. Membrane proteins and complex glycoproteins have sunk more campaigns at this stage than at any other. Generating candidates by design requires a structure, experimental or predicted, in place of the protein itself, and that requirement is a real one, since without a structure there is no design. Antigen is still needed for validation, later and in smaller quantity.
• Engineering and optimisation loses part of its job: humanisation exists because the binder came from a mouse or a llama, and a designed sequence can be built on a human or humanised framework from the outset, so the step becomes unnecessary rather than easier. Affinity maturation does not disappear, since the designs in the RFantibody study started in the tens to hundreds of nanomolar range and needed optimisation to reach single digits (Bennett et al., Nature 2026, 649:183-193, DOI, reference located through PubMed). Reformatting is unchanged.
• Developability assessment moves earlier: in the standard workflow it sits at stage seven, after a lead has been selected, which is exactly the point at which a liability becomes expensive. When candidates exist as sequences and structures before anything is made, the computable part of the assessment can be applied during selection instead, as a filter rather than a verdict. This is the change with the clearest economic value, and it is the subject of developability by design.

What stays identical, whichever route produced the binder

This list matters as much as the previous one.

• Target validation is untouched: no design method tells you whether engaging your antigen produces the biological effect you are after. That remains experimental biology, upstream of any binder.
• Screening changes shape without disappearing: you still have to establish which candidates work. What changes is that you are testing tens or hundreds of designed sequences instead of panning a library of billions, which makes the experiment smaller and more defined rather than unnecessary. Shortlisted candidates are expressed in a cell-free system and tested by ELISA, and the reasoning is set out in what a binding experiment proves.
• Functional characterisation is unchanged: binding and function remain different things, whatever produced the binder.
• Preclinical translation is unchanged: the regulatory path is indifferent to how the sequence was obtained.

Design therefore compresses one stage, improves the timing of another, removes part of a third, and leaves half the workflow exactly as it was. It is a useful change rather than a new paradigm.

When substituting the generation stage is worth it, and when it is not

• Worth it where the epitope has to be specified rather than accepted, where the antigen is hard to produce but its structure is known, where the target is a self-protein or a toxin that immunisation handles badly, where sequence distance from existing patents is a constraint, where developability should act as a selection criterion rather than a later check, or where animal work is excluded.
• Not worth it where no structure is available and none can be predicted with confidence, where the target is disordered, where the requirement is maximum affinity irrespective of epitope, or where a validated binder already exists and the real work is reformatting it.

Immunochem works entirely within the substituted stage and the two it touches: structure in, site chosen, sequences designed, developability filtered during selection, candidates expressed and tested. Six weeks to candidate sequences, twelve weeks to binders validated in the laboratory. The stages upstream and downstream of that remain yours.

Receive a feasibility study on your antigen - for free

Frequently asked questions

What are the stages of the antibody discovery workflow?

Target validation, antigen design and production, hit generation, screening and hit selection, engineering and optimisation, functional characterisation, developability assessment, and preclinical translation. Different groups split or merge them, but that is the common sequence.

Does computational design replace the whole workflow?

No. It replaces hit generation, and it changes the antigen requirement, the engineering stage and the timing of developability assessment. Target validation, functional characterisation and preclinical work are unaffected.

Do you still need screening if the antibodies are designed?

Yes, and it is smaller. Rather than enriching a library of billions, you test a defined set of designed candidates for expression and binding. The experiment is narrower but still necessary.

Do you need the physical antigen for computational design?

Not to generate candidates, which need a structure rather than a protein. You need antigen for the validation assay. This is the main practical advantage on targets that are difficult to express in native form.

Where does developability fit in?

Conventionally after lead selection. With designed candidates the computable part can be applied during selection, which turns it into a filter instead of a late discovery.

Is a designed antibody discovery workflow faster overall?

The generation stage is substantially faster, which is where our six and twelve week timelines come from. The stages design does not touch, particularly functional characterisation and preclinical work, take what they always took. Compressing one stage of eight shortens a programme without transforming it.