Designed, not screened: why we compose antibodies instead of hunting for them

Most antibody programs start by looking. Ours start by building. That distinction sounds like marketing until you trace what it changes about cost, control, and what you can actually promise a partner.

Screening is a search problem

The dominant paradigm in antibody discovery is search. You generate enormous diversity — an immunized animal, a phage or yeast library, a B-cell repertoire — and then you pan, sort, and enrich until something sticks to your target. The molecule that emerges is a discovery, not a decision. You found it; you did not specify it.

Search has real strengths. Nature's diversity is vast, and selection is brutally good at surfacing binders you would never have drawn on a whiteboard. But search has structural costs that get worse, not better, as the molecule gets more complex:

• You inherit what you find. Developability liabilities, awkward epitopes, and manufacturing headaches come bundled with the binder. You discover them late.
• Multi-arm formats break the assay. Screening optimizes one binding event at a time. A bispecific or trispecific has to satisfy several constraints simultaneously — geometry, valency, chain pairing, avidity — and a panning assay does not see the whole molecule.
• The search is opaque. You can rarely explain why the winner won, which makes the next program start near zero.

For a single conventional antibody against an easy target, search works well enough that design feels like over-engineering. For multi-specific molecules, the search space stops being a help and becomes the problem.

Composition is a different problem

Design treats the molecule as something you assemble from characterized parts against an explicit specification. You decide the targets, the geometry, the valency, and the trade-offs first, then construct candidates that satisfy them.

This is what NOVA-3 is built to do. It is a computer-aided design suite for multi-specific antibodies — bispecifics, trispecifics, and VHH shuttle formats — and its unit of work is composition, not enrichment. You specify arms and architecture; the suite assembles candidate molecules from parts and evaluates them.

To be precise about where the intelligence lives: NOVA-3 does not out-fold the best structure-prediction models. It wraps them. We use best-of-breed folders — AlphaFold 3, Chai, ESMFold — to evaluate the structures we compose. Our differentiation is not a better folder. It is everything around the fold:

• Multi-specific composition — assembling and scoring multi-arm molecules as whole objects, not isolated binders.
• Cryptographic provenance — every design and prediction is bundled with verifiable evidence of what was run and when.
• A wet-lab DBTL flywheel — Design–Build–Test–Learn cycles that feed measured outcomes back into the design logic.
• Targeted quantum active-space corrections — narrow quantum-chemistry corrections on the hard chemistry, with the hardware fraction reported per job. We make no claim of demonstrated quantum advantage.

What composition changes downstream

The interesting part is not the philosophy. It is what flips once your molecule is specified rather than found.

Liabilities move upstream. When you compose, you can apply developability and manufacturability constraints at design time instead of triaging them after a hit. The trade-offs are visible while you still have cheap choices.

The whole molecule is the object. Geometry, valency, and chain pairing are first-class design variables, not emergent accidents of a library. For multi-specific formats, where the entire mechanism depends on spatial arrangement, that is the difference between intent and luck.

Provenance becomes possible. A discovered molecule has a murky lineage. A composed one has a record: these parts, this architecture, these models, this version, this result. That record is what a diligence team can actually check.

Learning compounds. Each DBTL cycle sharpens the design logic for the next program. Search restarts; composition accumulates.

The honest boundary

Design does not repeal biology. The affinity, ADCC, and brain-exposure figures attached to our internal programs are predicted values or design targets, not measured wet-lab results — and we label them that way every time. Our three internal programs are pre-wet-lab: an in-silico proof of concept, an in-vitro proof of concept, and a lead-candidate-ranked stage. Composition gets you better-specified, better-documented candidates to put into the lab. It does not let you skip the lab.

That is the whole thesis. Screening asks the world to hand you a molecule and hope it behaves. Composition asks you to state what you want, build to it, and keep the receipts. For multi-specifics, where the molecule only works if several things are true at once, "designed, not screened" is not a slogan. It is the only approach that lets you say why a candidate should work before you spend the bench time finding out.

See how composition works in practice on the platform page, or look at how we document our internal programs on the pipeline page.