Protein design

What is protein design?

Protein design is the computational creation or optimization of protein sequences and structures for a defined objective. The objective may be a new fold, a sequence compatible with a backbone, catalytic activity, antigen recognition, peptide binding, target-specific protein binding, or improved sequence properties. Different design tasks require different inputs, models, scores, and experimental evidence.

Protein design is an umbrella use case rather than one algorithm. De novo methods generate new structural or sequence space; inverse folding finds sequences for a fixed backbone; enzyme, antibody, peptide, and binder design add functional and molecular-class constraints. Protein sequence design spans both structure-conditioned and sequence-generative approaches.

Choose the narrowest task that matches what you are trying to create. Every workflow should preserve constraints, random seeds, generated candidates, component scores, structures, and rejection reasons. Computational ranking reduces a search space, while synthesis, expression, structural characterization, and functional assays determine whether a design works.

When to use protein design

Choose the design task from what you intend to make and how it will be tested:

• Create a new backbone before choosing a compatible sequence.
• Redesign residues around a known fold, interface or active site.
• Compare candidate binders, antibodies or peptides under an explicit assay plan.

Strengths and limitations

Design tools can explore backbone and sequence choices that are impractical to enumerate by hand. Explicit constraints make it possible to trace a candidate back to the target, fixed residues and generation settings. Independent refolding and property review help reject obvious failures before committing to laboratory work.

A model can satisfy its design objective while missing expression, stability or function. Scores from different generators and predictors describe different tasks and cannot be combined into a universal success probability. Experimental testing must determine whether a protein folds in the intended context and performs the proposed job.

Which protein design tool to use

Start from what you already have: a backbone, a structure with a ligand or metal in it, a target surface, or only a sequence. The tools below run from a PDB file, an RCSB ID or a pasted sequence, and each one quotes its credit cost before the job starts. ESM-IF1 runs start at 20 credits, ProteinMPNN at 25 and RFdiffusion3 at 100, while BindCraft reserves 5,400 credits for a standard binder-design run.

What you want to do	Tool	What you get back
Design sequences for a fixed backbone	ProteinMPNN	Designed sequences with confidence and recovery values, backbone PDB files, and optional side-chain packing
Get a second, independent sequence set for the same backbone	ESM-IF1	Designed sequences with native sequence recovery and a mutation list
Redesign residues around a ligand, metal or cofactor	LigandMPNN	Designed sequences with overall and ligand-context confidence, FASTA and PDB files
Generate new protein backbones, or a binder backbone for a target	RFdiffusion3	Backbone-only structures, each needing a sequence assigned afterwards
Design a binder against a target structure	BindCraft	Ranked binder structures with interface confidence and a per-candidate checks table
Design an antibody or nanobody against an epitope	RFantibody	Designed antibody-antigen complexes with RF2 interface PAE and pLDDT scores
Sample new sequences without a structure	ProGen2	Generated sequences as FASTA, or likelihood scores for a sequence you supply
Check whether a designed sequence refolds	ESMFold	A predicted structure with per-residue pLDDT

ProteinMPNN is the usual first run when you have a backbone, because it is cheap and its output feeds the refolding and property checks directly. Pair it with ESM-IF1 when you want to see whether two models agree on the same positions. RFdiffusion3 and BindCraft start from a target or a length range instead of a sequence, so budget for the structure prediction and filtering that follow them. Whichever tool you pick, refold the sequences before anything goes to the lab.

Choosing a protein design method

The useful taxonomy follows researcher intent. De novo protein design creates new backbones before sequence assignment. Inverse folding starts with a fixed backbone and asks which sequences might support it. Protein sequence design includes this structure-conditioned task as well as broader sequence optimization.

Enzyme design needs catalytic geometry and a measurable reaction. Protein binder design starts from a target surface and an interface objective. Antibody design adds framework, CDR and chain-format constraints, while peptide design handles short sequences with different stability and chemistry concerns.

Ligand-conditioned design and stability design remain important methods and objectives, but they fit within enzyme, sequence, or binder workflows here rather than becoming additional maintained spokes without demonstrated search demand.

Applications of protein design

A new scaffold can test whether a structural motif works outside its natural protein context. Structure-conditioned redesign can explore substitutions around a stable fold while keeping catalytic or binding residues fixed. In both cases, independent folding predictions help prioritize what to test, but they do not establish expression or function.

Targeted design can also propose enzymes, binders, antibodies and peptides for a defined experiment. The useful output is a diverse set of candidates with stated constraints and failure reasons, so an assay can test a hypothesis rather than merely confirm the top model's appearance.

How to design proteins online

A defensible design workflow begins with a measurable objective. The featured ProteinIQ panel is specifically a fixed-backbone sequence comparison; choose a spoke workflow for backbone generation or a functional binder, enzyme, antibody or peptide task.

1. Choose a backbone. Supply a reviewed protein PDB with the intended chains and fixed residues identified.
2. State the design objective. Decide whether general backbone compatibility, solubility context or thermostability bias is the comparison of interest.
3. Run the method panel. ProteinMPNN proposes sequences from the backbone. ESM-IF1 supplies an independent geometry-conditioned branch.
4. Compare specialized branches. SolubleMPNN uses a soluble-protein training set. HyperMPNN uses hyperthermophile-trained weights; neither guarantees the named property in an experiment.
5. Review and select. Keep each model's scores and sampled sequences separate, check fixed positions and diversity, then refold selected candidates before laboratory testing.

How to interpret protein design results

No single score establishes design success. Evaluate constraint satisfaction, structural agreement, local confidence, geometry, sequence diversity, stability, solubility, aggregation, and task-specific evidence separately.

Compare candidates within the same model and settings unless scores are explicitly calibrated across methods. Preserve rejected designs and score distributions to avoid selection based only on attractive visualizations.

Experimental validation for protein design

Expression and folding are early gates, not proof of function. The decisive assay must match the objective: affinity and specificity for binders, catalytic rate and selectivity for enzymes, or structural and biophysical properties for scaffold designs.

Plan controls, replication, and candidate diversity before generation. Export sequences, structures, settings, provenance, and selection logic together for synthesis and laboratory handoff.

How the featured workflow works

The hub workflow compares four sequence-design methods from one backbone; each spoke uses a workflow matched to its specific design task.

1. Define the objective. Choose the protein-design task and acceptance criteria.
2. Prepare the backbone. Clean the PDB and identify chains and fixed residues.
3. Run four methods. Submit the same backbone to four sequence-design models.
4. Compare candidates. Review probabilities, diversity, conserved positions, and method agreement.
5. Validate and export. Refold and screen selected candidates before experimental testing.

Inputs and outputs for the featured workflow

Check formats before running, then inspect and download the result from every workflow step.

Inputs

Design evidence

PDBFASTAJSONTXT

Protein structures, sequences, targets, motifs, fixed residues, and constraints required by the selected design task.

Outputs

Design candidates

PDBFASTACSVJSON

Generated structures and sequences, model-native scores, rankings, logs, and validation files.