← back

A note on de novo binder design model evaluation

Boltz Team · Blog · June 16, 2026

main takeaways

Typical binder experiment setup is a target protein attached to a surface, then running a solution containing the proposed ligand over the surface. A device measures when the flow of the solution is interrupted (ideally binding).

1. Good curves (binders) look something like initial step-function response during binding, then a slow fall off once the solution flow stops (gradual unbinding, exponential decay). Should have peak response values positively related to binder concentration.
2. Bad curves: noisy, flat, changing concentration doesn’t matter.
3. Ambiguous curves: maybe not complete unbinding after solution flow stops, or different concentrations after solution flow stops don’t converge. Have some signal, but don’t clearly reflect one-binder-to-one-protein events. Non-exponential-decay dissociation fits here.

People fit curves to these response plots to attain affinities, but doing so for ambiguous curves results in misleading values.

Flipping the assay is something you can do: attach the binder to the surface and run a solution of the target over it. Can help de-ambiguify type (3) curves.

personal thoughts

Nothing much negative. These are good standards. Maybe some shade thrown at Adaptv Bio and/or Chai lol