For most of the history of protein engineering, improving a protein meant starting from one that already existed: mutate it, screen the variants, keep what worked. De novo design inverts that. You specify the function you want — bind this epitope, catalyse this reaction, adopt this fold — and generate a sequence that has never existed in any organism.
Since roughly 2022 this has stopped being a curiosity. Designed binders now reach picomolar affinities, and designed proteins have entered preclinical development. It has also acquired a reputation for overpromising, so this article covers what the methods actually do, how well they work, and where they still fail.
The two-stage architecture
Almost every modern pipeline separates the problem into shape and sequence.
Stage 1: generate a backbone
RFdiffusion is the dominant method. It applies a denoising diffusion model — the same family of generative models behind image generation — to protein backbones, starting from noise and iteratively refining it into a plausible three-dimensional structure.
Critically, it can be conditioned. You can specify a target surface for a binder to engage, hold a catalytic motif fixed while everything around it is generated, scaffold an epitope for vaccine design, or impose symmetry for oligomeric assemblies. The output is a backbone: coordinates without a sequence.
Stage 2: find a sequence that folds into it
This is the inverse folding problem, and ProteinMPNN solved it well enough to change the field. Given backbone coordinates, it predicts an amino acid sequence likely to adopt that shape, using a message-passing neural network over the residue graph.
Its practical impact is easy to understand: earlier physics-based design produced sequences that expressed and folded perhaps a few percent of the time. ProteinMPNN raised that dramatically, and it runs in seconds rather than CPU-days.
Stage 3: check your own work
The step that separates credible campaigns from wishful ones is in silico refolding. Take each designed sequence, predict its structure independently with AlphaFold2 or ESMFold, and ask whether the prediction matches the backbone you designed.
Designs where the prediction agrees closely with the intended structure — typically under 2 Å RMSD with high predicted confidence — succeed far more often experimentally. Designs where it disagrees almost never do. This filter is cheap and discards most failures before anyone orders a gene.
What the success rates actually look like
Honest numbers matter here, because the gap between the headline result and typical practice is wide.
For binder design against a well-behaved target with a defined, reasonably rigid epitope, published campaigns often report low single-digit percentage hit rates from tens of thousands of designs, screened down to a handful of validated binders. That sounds poor until you compare it to the alternative: it is a genuinely novel binder, obtained without immunisation or library screening, in weeks.
Success drops sharply when the target is flexible, when the epitope is flat and polar rather than a defined pocket, or when the interface has to be large. Targets that are hard for antibodies are usually hard for designed binders too, for the same physical reasons.
Affinity maturation is normal. First-round designs frequently land in the micromolar range and need optimisation — computational or experimental — to become useful. Reports of picomolar designs straight from the generator are the exception, not the expectation.
Where de novo design is genuinely the right tool
- Targets antibodies struggle with. Small, rigid designed binders can access clefts and grooves that an antibody paratope cannot, and they are far easier to express.
- Thermostable formats. Designed mini-binders are often extremely stable, which matters for diagnostics, reagents and anything shipped without a cold chain.
- Epitope scaffolding for vaccines. Presenting a neutralising epitope in a rigid, correctly conformed scaffold is a design problem that de novo methods handle directly.
- Symmetric assemblies and nanoparticles. Self-assembling designed cages for vaccine or delivery applications are among the field’s clearest successes.
- Biosensors. Designing a binder plus a conformational switch into one molecule is difficult by evolution and natural for design.
Where it still fails
Catalysis remains hard. Designing a binding site is now routine; designing a site that stabilises a transition state and turns over efficiently is not. De novo enzymes generally show activity orders of magnitude below natural counterparts and usually require directed evolution afterwards.
Conformational change is poorly handled. Most methods design toward a single target structure. Proteins that must switch between states — allosteric regulation, transporters — are a much harder problem.
Membrane proteins lag. The training data is dominated by soluble proteins, and design methods inherit that bias.
Developability is not modelled. A design can be predicted to fold beautifully and still aggregate, express poorly, or prove impossible to purify. Aggregation propensity, expression liability and solubility filtering are separate steps, and skipping them wastes a lot of bench time.
What a realistic campaign looks like
- Assess the target honestly. Structure quality, epitope definition, rigidity, and whether the site is genuinely druggable by a protein binder. This conversation is where most projects should be reshaped or stopped.
- Generate broadly. Thousands to tens of thousands of backbones with RFdiffusion, conditioned on the target site.
- Design sequences with ProteinMPNN, several per backbone.
- Filter hard by in silico refolding agreement and predicted confidence. Most candidates die here, which is the point.
- Simulate the survivors. Molecular dynamics on the designed complex shows which interfaces persist in explicit solvent and which dissolve within nanoseconds — a filter that static structure prediction cannot provide.
- Apply developability filters for aggregation, solubility and expression.
- Order a panel, not a candidate. Typically dozens to a few hundred designs for expression and binding screens.
The strategic point is that computation is cheap and synthesis is not. Every filter that removes a design before it is ordered pays for itself. Anyone proposing to design a single molecule and take it straight to the bench has misunderstood how these methods behave.
Design is a triage method, not an oracle
De novo design does not tell you what a protein will do. It produces a focused, well-reasoned set of candidates far faster and more cheaply than screening a library, with the obviously implausible options already removed. The experiment still decides.
Used that way, it compresses months of discovery into weeks. Presented as computational certainty, it produces exactly the disappointment that has made some experimentalists sceptical of the whole field.
