Good number of de novo protein design workflows in the literature today. Increasingly, many of them are also reporting assay readouts showing nanomolar and picomolar de novo designed binders against a diverse array of targets.
And yet, no two assay are necessarily the same. In Hannes Stärk's write-up "A note on de novo binder design model evaluation", the Boltz team found that two CROs' results disagreed with each other given the same target and binder sequences. How do we know if a result is a screening hit or a confirmed binder? The authors propose the following for a clean sensogram result:
How do we know if a result is a screening hit or a confirmed binder? The trick is that every sensorgram is really a clear binder, no interaction, or an ambiguous signal where something is happening but you can't say it's one design binding one target and fitting an affinity equation to that third case always returns a number, just not a real one. The Boltz criteria are a checklist for staying out of it:
A long enough association phase to resolve binding. Association and dissociation that fit the expected kinetic model, with the signal actually coming off during dissociation: if it barely decays, you can't trust the off-rate. Little or no biphasic behavior for a presumed 1:1 interaction, since two dissociation phases mean you're fitting one exponential across two events. Plausible on and off rates, without suspiciously slow kinetics, slow association paired with slow dissociation usually signals an artifact, not an unusually good molecule. And a clean reference channel that doesn't dominate the signal.
Of course, even a confirmed binder does not a drug make. Here, I'll turn to Debbie and Hetu's "Progressable Binders" framework.
They build on those five criteria with two additions. First, antigen quality: a measurement is only as good as the reagent behind it, so mammalian-expressed antigen (correct folding and PTMs) and monodispersity QC before screening set the ceiling on everything downstream. Second, and the one I'd watch hardest, valency. When a partner is multivalent (often an Fc tag dimerizing the target) you're no longer measuring 1:1 binding; the off-rate collapses and the affinity reads far tighter than it is. Their CSF1 example resolved from an apparent ~20 nM to no better than ~5 µM once the assay was flipped. Which is why any affinity against a multimeric partner should be reported as apparent and why flipping the orientation is the cheapest sanity check you have.
Plenty of tight binders never become candidates because they don't express, are immunogenic, or simply don't perform their intended function. A progressible binder is therefore one that clears binding, along with functional and biophysical constraints.
Most importantly, a candidate must clear all of these constraints, expressability, stability, low polyreactivity, not having sequence liabilities, and low predicted immunogenicity. The authors show an interesting result here, that making camelid VHHs more human (lowering immunogenicity risk), actually degrades developability.
De novo design is having its moment in the zeitgeist, most R&D orgs I speak with now take AI-designed protein seriously. With this, the field still somewhat underinvests in reporting conventions, assay quality, and comparisons between service providers.
AI-generated molecules are getting good. We just need the evidence for their success to be consistent enough that "de novo binder against a hard target" stops meaning different things to different groups.