De novo protein design

What is de novo protein design?

De novo protein design is the process of creating new protein structures and sequences from scratch rather than modifying a natural protein that already performs the intended role. Generative models can propose backbones under geometric or functional constraints, after which sequence-design methods assign residues expected to stabilize those structures. The result is a set of computational hypotheses that still requires structural and experimental validation.

A complete campaign separates backbone generation from sequence assignment and validation. This matters because a visually plausible backbone does not guarantee that any amino-acid sequence will fold into it, and a high model score does not establish soluble expression, monodispersity, stability, or function.

Use unconditional generation for new structural space, or add symmetry, motif, target, and functional constraints when the design must satisfy a specific geometry. Preserve rejected candidates and score distributions so selection is based on stated gates rather than one attractive model.

When to use de novo protein design

Start from a new backbone only when the desired structural space or geometry calls for it:

• You need a fold or assembly beyond a suitable natural template.
• A motif or symmetry constraint must be built into a new scaffold.
• You can refold and test several sequence candidates per accepted backbone.

Strengths and limitations

Backbone generation can sample structures that a point-mutation campaign would never reach. Geometric conditioning allows a motif or symmetry requirement to shape the proposal, while independent sequence assignment reveals whether a backbone supports more than one plausible sequence.

A generated backbone is not evidence that a sequence will fold into it. Constraint satisfaction can narrow diversity or yield strained geometry, and favorable model scores say little about expression or function. Keep rejected structures and refolding disagreements visible so selection does not depend on one attractive image.

De novo protein design methods

Diffusion and other generative models sample structures from learned protein geometry. Conditioning can steer that sampling toward a motif, target, symmetry, or shape, but stronger constraints can reduce diversity or create incompatible requirements.

Sequence assignment is a separate inverse problem. Multiple sequences per backbone help reveal whether a structural proposal has a broad compatible sequence space or depends on a narrow, fragile solution.

Applications of de novo protein design

A new scaffold can carry a catalytic motif, present a binding surface or assemble into a designed oligomer. The generator proposes backbone geometry; sequence assignment and refolding then test whether that geometry has plausible molecular support.

For a structural materials project, symmetry and assembly constraints may be central. For a functional scaffold, the motif must remain accessible and correctly oriented after sequence design. In either case, testing several distinct backbones is more informative than optimizing one model indefinitely.

How to run de novo protein design online

The featured template generates unconditional backbones before sequence assignment and refolding. Use a conditioned setup when a motif, target or symmetry must be enforced.

1. Set the generation task. Choose unconditional backbone generation in the featured workflow; a motif- or target-conditioned question needs a different setup.
2. Generate backbones. Run RFdiffusion3 and retain its PDB outputs, settings and random seeds, including rejected geometries.
3. Assign sequences. Send accepted backbones to ProteinMPNN and sample several sequences per backbone rather than selecting one by score alone.
4. Refold independently. Pass designed sequences to ESMFold and compare the predicted folds with their generating backbones.
5. Check geometry. Inspect refolded structures with MolProbity, paying attention to clashes and backbone or side-chain outliers.
6. Select experiments. Keep backbone–sequence pairings and failure reasons; test folding and the intended property on a diverse subset.

How to interpret de novo protein design results

Compare designed and independently predicted structures using global and local agreement, then inspect clashes, secondary structure, buried polar atoms, exposed hydrophobics, aggregation risk, and sequence diversity.

Prospective validation should measure the property the design was intended to create. Expression and folding checks are necessary, but they do not replace binding, catalytic, assembly, or other functional assays.

Experimental validation and handoff

Retain generation settings, seeds, every sequence–backbone pairing, refolded structures, and rejection criteria. Confirm folding and intended function experimentally.

For laboratory work, keep each backbone with all sampled sequences and the independent refolded structures. A failed expression or fold check should be traceable to the specific geometry and sequence that produced it.

How de novo protein design works

Generate unconditional backbones with RFdiffusion3, assign sequences with ProteinMPNN, refold with ESMFold, and inspect geometry with MolProbity.

1. Specify intent. Define the desired size, topology, symmetry, motif, or functional constraints before generation.
2. Generate backbones. Generate a sufficiently diverse backbone set and retain model settings and random seeds.
3. Assign sequences. Assign several sequences to each accepted backbone instead of treating one sequence as definitive.
4. Refold candidates. Refold sequences independently and compare predicted structures with their design backbones.
5. Select for testing. Apply geometry, stability, solubility, and experiment-specific gates before synthesis.

Inputs and outputs

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

Inputs

Design input

PDBFASTAJSONTXT

Design constraints and, when applicable, a target, motif, or symmetry definition.

Outputs

Design and review outputs

PDBFASTACSVJSON

Designed backbone PDB files, candidate FASTA sequences, refolded models, scores, and review files.