# What is protein design?

> Protein design is the computational search for sequences and structures that do a chosen job. Inverse folding, de novo backbone generation and binder pipelines solve different parts of that search, and the structure-prediction filter after them decides much of the success rate. Learn what each stage does, how often published campaigns worked, and how many designs that means ordering.

## What is protein design?

Protein design is the computational search for amino-acid sequences, and the structures they fold into, that do a job chosen in advance: hold a fold, bind a target, position a catalytic residue. It is the reverse of structure prediction. Prediction takes a sequence and returns a shape; design takes a wanted shape or function and returns sequences worth testing.

The first fully designed protein with a new fold, Top7, appeared in 2003 and matched its design model to about 1.2 Å in the crystal structure. Two decades later the work is split into stages, each with its own software, and the stage that gets the least attention often matters most. In the RFdiffusion paper, a roughly 100-fold gain in binder success over earlier campaigns was credited half to the new generator and half to filtering designs with AlphaFold2.

Three jobs make up most of the field. Inverse folding assigns sequences to a backbone that already exists. De novo generation creates the backbone itself, sometimes around a motif or a target. Binder design chains the two and adds a check that the result sticks to the intended surface. The sections below take them in that order, and the worked example after them turns the published success rates into how many designs a campaign has to order.

## Inverse folding designs sequences for a fixed backbone

Inverse folding asks which sequences would fold into a given backbone. [ProteinMPNN](/app/protein-mpnn) is the standard answer: on native backbones it recovered 52.4% of the original residues, against 32.9% for Rosetta, and generated a sequence in about a second where Rosetta needed minutes. The task and its inputs are covered in more detail on the [inverse folding](/use-cases/inverse-folding) use-case page.

Sequence recovery is a benchmark, not a goal. A recovery of 52.4% on a 100-residue protein means about 48 residues differ from the natural sequence, and the design is still expected to fold. A high recovery shows that the model has learned what nature does at each position; it does not show that a new sequence for a new backbone will fold. That second question is answered by predicting the structure of the designed sequence and comparing it with the backbone it was designed for.

The catch is that this check uses another model. A design that refolds in [ESMFold](/app/esmfold) with a high [pLDDT](/guides/plddt) and a low [RMSD](/guides/rmsd) to the target backbone has passed one model's opinion of another's output. It is strong evidence that the sequence is self-consistent and weak evidence about expression, solubility or stability. Two sequence designers disagreeing on the same backbone, for example [ProteinMPNN](/app/protein-mpnn) and [ESM-IF1](/app/esm-if1), is a cheap way to see how narrow the compatible sequence space is.

When the structure contains a ligand, metal or nucleic acid, [LigandMPNN](/app/ligand-mpnn) designs the residues around it. [Protein language models](/guides/protein-language-models) can propose sequences with no backbone at all, but then nothing ties the sequence to a particular shape.

## De novo design generates the backbone

De novo design starts one step earlier. A diffusion model such as [RFdiffusion3](/app/rfdiffusion3) generates backbone coordinates under constraints: a length, a symmetry, a motif to scaffold or a target surface to face. The backbone has no meaningful sequence yet, so ProteinMPNN or a similar tool assigns one and a structure predictor refolds it.

RFdiffusion reported this pipeline across several tasks. For TIM barrels, the in silico success rate was 42.5%, and at least 8 of 11 designs tested in the lab were soluble and thermostable with circular dichroism spectra matching the design. The authors defined in silico success by agreement with AlphaFold2, so the 42.5% describes a model agreeing with a model, and the 8 of 11 is the experimental figure.

Constraints cut both ways. A motif or symmetry requirement can force strained geometry, and a backbone that looks tidy can have few or no sequences that adopt it. Generating several sequences per backbone and keeping the ones that refold separates backbones that are broadly designable from ones that were lucky once. The [de novo protein design](/use-cases/de-novo-protein-design) page lists the inputs for each kind of constraint.

## Binder design adds a target

A binder campaign generates backbones against a target surface, assigns sequences, then asks whether the complex still forms when it is predicted again. Bennett and colleagues described two ways this fails: the sequence does not fold into the designed monomer, or the folded monomer does not bind the target as designed. They found that AlphaFold2 or RoseTTAFold predictions caught both, and that the average predicted error across the interface (pAE_interaction) separated experimental binders from non-binders, with success rates rising sharply below 10. Adding that filter and swapping Rosetta for ProteinMPNN raised success nearly 10-fold.

RFdiffusion then reported binders to five targets with an average experimental success rate of 19%, fewer than 100 designs tested per target, and a binder found for every target. BindCraft, which runs generation, sequence design and prediction in one loop, reports experimental success rates of 10% to 100% across the targets its authors tested. [BindCraft](/app/bindcraft) runs that loop online; [BoltzGen](/app/boltzgen) covers miniproteins and antibody formats, and [RFantibody](/app/rfantibody) designs antibodies against an epitope. The [protein binder design](/use-cases/protein-binder-design) and [antibody design](/use-cases/antibody-design) pages cover the inputs.

The numbers come with a condition. Each rate belongs to a target, a hit threshold and a lab, and the PAE-style interface metrics in these pipelines are not all on one scale. What [ipTM](/guides/iptm-vs-ptm) or [ipSAE](/guides/ipsae) says about a complex is a ranking signal from one predictor, not a dissociation constant, and the [mBER metrics](/guides/mber-metrics) are a different set of scores again.

## How many designs a campaign has to order

A success rate becomes useful when it is turned into a number of designs. If each design binds independently with probability p, the chance that at least one of n works is 1 − (1 − p)^n, so the number needed for a 95% chance of at least one hit is ln(0.05) / ln(1 − p), rounded up. We calculated this for the success rates reported by the RFdiffusion paper.

  <caption><strong>Table 1. Designs needed for a 95% chance of at least one binder.</strong> We calculated n = ln(0.05) / ln(1 − p) from success rates reported by Watson and colleagues: their estimates for earlier Rosetta campaigns if AF2 filtering had been used (TrkA 1.5%, PD-L1 3.7%, InsR 5.5%) and the 19% average for RFdiffusion binders.</caption>
  <thead>
    <tr>
      <th scope="col">Pipeline</th>
      <th scope="col">Success rate p</th>
      <th scope="col">Designs for 95% chance of a hit</th>
    </tr>
  </thead>
  <tbody>
    <tr>
      <td>Rosetta with AF2 filter, TrkA</td>
      <td>1.5%</td>
      <td>199</td>
    </tr>
    <tr>
      <td>Rosetta with AF2 filter, PD-L1</td>
      <td>3.7%</td>
      <td>80</td>
    </tr>
    <tr>
      <td>Rosetta with AF2 filter, InsR</td>
      <td>5.5%</td>
      <td>53</td>
    </tr>
    <tr>
      <td>RFdiffusion, five-target average</td>
      <td>19%</td>
      <td>15</td>
    </tr>
  </tbody>

At 19%, a plate of 15 designs is a reasonable first experiment; at 1.5% the same confidence takes about 200. The estimate assumes designs are independent, which they are not. Designs from one backbone share its fate, and a target with a poor surface lowers every p at once, so treat the table as a floor on how many designs to order, not a guarantee.

It also shows why the filter matters as much as the generator. The Rosetta rows already assume the AF2 filter, as in Watson's estimate; the original campaigns tested thousands of designs per target, against fewer than 100 for RFdiffusion. A pipeline that skips the refolding step spends its budget on designs that were never going to fold.

## What each stage cannot see

  <caption><strong>Table 2. What each design stage cannot see and a cheap check.</strong></caption>
  <thead>
    <tr>
      <th scope="col">Stage</th>
      <th scope="col">What it cannot see</th>
      <th scope="col">Sign of trouble</th>
      <th scope="col">Cheap check</th>
    </tr>
  </thead>
  <tbody>
    <tr>
      <td>Inverse folding</td>
      <td>Expression, solubility, stability in a cell</td>
      <td>Low confidence at many positions; designs that differ from each other everywhere</td>
      <td>Refold the sequences and compare the RMSD to the backbone; run a second designer</td>
    </tr>
    <tr>
      <td>De novo backbone generation</td>
      <td>Whether any sequence adopts the backbone</td>
      <td>Refolding fails for every sequence of a backbone</td>
      <td>Several sequences per backbone; keep rejected backbones in the record</td>
    </tr>
    <tr>
      <td>Binder pipeline</td>
      <td>Affinity, specificity, the right conformational state of the target</td>
      <td>Interface confidence high only in one predictor; design placed on a surface that is buried in the biology</td>
      <td>Re-predict with a second model; check the target preparation and hotspots</td>
    </tr>
    <tr>
      <td>Structure-prediction filter</td>
      <td>Anything the predictor was not trained to notice</td>
      <td>Every metric agrees, because they come from one model</td>
      <td>Treat the score as a ranking, not a measurement</td>
    </tr>
  </tbody>

None of these checks replaces the experiment. Computation narrows thousands of candidates to a few dozen, and expression, folding and binding tests decide which of those are real.

## Protein design tools on ProteinIQ

A typical run starts from the structure or target you have. Fixed backbones go to the sequence designers, new backbones and binders go to the generators, and every output goes back through a structure predictor before anything is ordered.

  <caption><strong>Table 3. Protein design tools on ProteinIQ.</strong></caption>
  <thead>
    <tr>
      <th scope="col">Task</th>
      <th scope="col">Tool</th>
    </tr>
  </thead>
  <tbody>
    <tr>
      <td>Design sequences for a fixed backbone</td>
      <td>[ProteinMPNN](/app/protein-mpnn)</td>
    </tr>
    <tr>
      <td>Get a second, independent sequence set for the same backbone</td>
      <td>[ESM-IF1](/app/esm-if1)</td>
    </tr>
    <tr>
      <td>Redesign residues around a ligand, metal or cofactor</td>
      <td>[LigandMPNN](/app/ligand-mpnn)</td>
    </tr>
    <tr>
      <td>Generate new backbones or target-conditioned binder backbones</td>
      <td>[RFdiffusion3](/app/rfdiffusion3)</td>
    </tr>
    <tr>
      <td>Design and filter binders against a target in one run</td>
      <td>[BindCraft](/app/bindcraft)</td>
    </tr>
    <tr>
      <td>Generate miniprotein, nanobody or Fab candidates</td>
      <td>[BoltzGen](/app/boltzgen)</td>
    </tr>
    <tr>
      <td>Design an antibody against an epitope</td>
      <td>[RFantibody](/app/rfantibody)</td>
    </tr>
    <tr>
      <td>Check whether a designed sequence refolds</td>
      <td>[ESMFold](/app/esmfold)</td>
    </tr>
  </tbody>

## Frequently asked questions

### What is protein design?

Protein design is the computational search for sequences and structures that fold into a chosen shape or perform a chosen function. The first fully designed protein with a new fold, Top7, matched its design model to about 1.2 Å in the crystal structure.

### What is the difference between de novo design and inverse folding?

Inverse folding assigns sequences to a backbone that already exists; ProteinMPNN recovered 52.4% of native residues on test backbones. De novo design creates the backbone too, usually with a diffusion model, and then uses inverse folding to give it a sequence.

### How often do designed binders work?

It depends on the target and the pipeline. RFdiffusion reported a 19% average experimental success rate across five targets with fewer than 100 designs tested each, and BindCraft reports 10% to 100%. Earlier energy-based campaigns were roughly an order of magnitude lower before deep-learning filtering was added.

### Does a high AlphaFold or ESMFold score prove a design works?

No. It shows that one model predicts the designed sequence to fold into the designed structure. Bennett and colleagues found the interface pAE below 10 sharply raised the success rate, but it remained a filter that ranks designs, with the experiment deciding.

### How many designs should I test?

At a 19% success rate, about 15 designs give a 95% chance of at least one binder if the designs are independent, and at 1.5% it takes about 200 (our calculation). Designs from one backbone are correlated, so order more than the minimum.

### Can I design a protein without a structure?

Language models can generate or score sequences without a backbone, but nothing then ties the sequence to a known shape. Most design campaigns pair a structure-aware designer with a refolding check for that reason.
