There is a version of AI protein design you get from conference talks, in which you describe a target and a model hands you a binder. And there is the version you get from the methods sections, which is more interesting and considerably more useful if you are deciding whether to try it.
The short answer is that de novo binder design genuinely works now, at a hit rate that makes small campaigns viable, and it fails often enough that you should plan for failure explicitly.
Structure prediction is not design
Most confusion here comes from collapsing two different things.
Structure prediction takes a sequence you already have and predicts how it folds. That is AlphaFold and its successors. It changed structural biology, and it is now routine enough that most people treat a predicted structure as a starting hypothesis rather than a result.
Design goes the other way. You specify a target and ask for a sequence that did not previously exist and that binds it. That is a far harder problem, it is much newer, and it is where the interesting numbers are.
The tools split along the same line. RFdiffusion generates backbone geometry. ProteinMPNN designs a sequence to fold into a given backbone. BindCraft and BoltzGen wrap the whole pipeline into something closer to one step.
The number that matters
Ask about experimental success rate, meaning designs that were actually expressed and measured, not designs that passed an in silico filter. The second number is much larger and tells you almost nothing.
BindCraft is often quoted as achieving 10 to 100 percent success. The useful version of that statistic is underneath it: across 212 designs the overall hit rate was 30.7 percent, and per-target rates ran from 10 percent against HER2 to 60 percent against SpCas9. Same pipeline, same authors, a six-fold spread depending only on what you point it at.
BoltzGen reports something similar from a different angle: testing 15 nanobody designs per target produced nanomolar binders for 6 of 9 targets. That is a genuinely striking result, because 15 designs is a scale an ordinary lab can express and screen without a robot.
And the failures are documented too. RFdiffusion has been reported performing poorly at designing functional binders for biochemical detection, which is a reminder that binding in an assay and working in your application are different bars.
Why the spread is so wide
Target-dependence is the single most important thing to internalise, and it is why asking "what is the success rate?" without naming a target gets you a useless answer.
Designs do better against targets with a well-ordered, hydrophobic, concave surface to grip. They do worse against flat interfaces, highly polar surfaces, glycosylated epitopes, and disordered regions. If your target is a floppy loop on a heavily glycosylated receptor, expect to be at the wrong end of that range, and budget accordingly.
Epitope specificity is the other catch. Most pipelines will happily give you something that binds your protein somewhere. Getting a binder to the functional site you care about, rather than to whichever patch is easiest, is a meaningfully harder request.
What this means practically
If you are in a non-computational lab and want to try this, the realistic shape of a first campaign looks like:
- A structure of your target, experimental or predicted, and a decision about which surface you want bound.
- A design run. The tools are open source, and there are now hosted services that will run them if you do not want to manage GPUs.
- Gene synthesis for 15 to 50 designs, which is cheap enough that this is not the constraint.
- Expression, purification, and a binding assay. This is the constraint, and it is entirely wet lab.
Which leads to the part that gets lost: the computational step is now the fast, cheap part. The bottleneck has moved to expressing and characterising the designs. If your lab cannot do that, the model output is not yet a result. Commissioning the wet-lab validation is usually more consequential to your timeline than choosing between design tools.
What it still cannot do
A designed binder is not a drug, an antibody replacement, or a reagent you can publish on without characterisation. Affinity is the easy property. Specificity against everything else in a lysate, stability, expression yield, immunogenicity, and behaviour in your actual assay are all separate questions that no current model reliably predicts.
Treat these tools as a very good hypothesis generator that has replaced months of screening with a week of compute. That is a real change, and it is a smaller claim than the one you will hear at a conference.