Inverse folding

What is inverse folding?

Inverse folding is the task of finding amino-acid sequences that are compatible with a supplied three-dimensional protein backbone. Unlike structure prediction, which maps sequence to structure, inverse folding holds the backbone geometry fixed and predicts residue identities or sequence probabilities. It is widely used for fixed-backbone redesign, sequence recovery, stability-oriented diversification, and sequence assignment after backbone generation.

The backbone defines much of the structural context, but it does not uniquely determine a sequence. ProteinMPNN, ESM-IF1, and related models can return different high-probability solutions because they use different representations, training data, and decoding procedures.

Generate multiple sequences at controlled sampling temperatures, preserve per-residue probabilities, and avoid overinterpreting native-sequence recovery as proof of design quality. Refolding and developability filters can reject obvious failures, while experiments establish whether candidates actually adopt the intended structure.

When to use inverse folding

Use inverse folding when the backbone is the constraint rather than the output:

• An experimental or reviewed model provides a fold to preserve.
• Catalytic, binding or interface residues must remain fixed.
• You want several sequence solutions to compare for one backbone.

Strengths and limitations

Backbone-conditioned models produce diverse amino-acid sequences while preserving specified fixed positions. Comparing ProteinMPNN with ESM-IF1 makes model dependence visible, and refolding each branch tests whether its sequences return to the intended geometry.

A sequence can score well against a supplied backbone but fail to fold, express or perform a function. Errors in the backbone, chain context or fixed-residue mask propagate into every design. Agreement between inverse-folding scores and a refolded model remains computational evidence, not structural validation.

Inverse folding methods

Inverse-folding networks encode backbone geometry and estimate compatible residue identities. Autoregressive and masked approaches differ in how sequence positions condition one another, so model comparison can identify candidates that depend on a single scoring assumption.

Sampling temperature controls the tradeoff between high-probability residues and sequence diversity. Fixed-position masks should preserve residues whose chemistry or interactions are essential rather than asking the model to rediscover every constraint.

Applications of inverse folding

Inverse folding is useful for redesigning an existing structural scaffold while retaining catalytic or binding residues. It can explore sequence diversity around an experimentally characterized fold without asking a model to invent a new topology.

The same method can test whether a proposed backbone has many plausible sequences or only a narrow, model-dependent solution. Refolding different sequence families can identify candidates whose structural agreement is less sensitive to one inverse-folding model.

How to run inverse folding online

The featured workflow runs ProteinMPNN and ESM-IF1 from the same backbone, then refolds each branch. Keep their native sequence scores separate.

1. Prepare the backbone. Supply one reviewed PDB with the intended chains, residue numbering and missing regions resolved.
2. Set fixed positions. Record catalytic, interface, disulfide or other residues that must not change, and verify that each tool accepts the intended mask.
3. Run two design branches. Sample sequences with ProteinMPNN. Compare them with a separate ESM-IF1 branch on the same backbone.
4. Refold each branch. Use ESMFold on the designed FASTA outputs without treating a refold score as experimental stability.
5. Compare candidates. Check fixed residues, diversity and structural agreement with the original backbone, keeping model-specific sequence scores separate.
6. Choose a test set. Export the exact backbone, masks, model settings, sequences and refolded structures for expression or structural tests.

How to interpret inverse folding results

Review sequence log-probabilities, diversity, conserved positions, refold agreement, clashes, and local confidence. Compare candidates within the same model and settings because raw scores are not necessarily calibrated across methods.

A fixed-backbone design can fail when the real protein relaxes, changes oligomeric state, binds a partner, or encounters cellular constraints. Test structure, stability, and the intended function with appropriate experiments.

Experimental validation and handoff

Keep the exact backbone, fixed-position masks, sampling temperature, model version, sequence scores, and refolded models. Validate selected designs experimentally.

Export the backbone, chain mapping and fixed-position mask with every selected sequence. Expression and structure measurements are needed to learn whether the designed amino acids actually support the intended fold.

How inverse folding works

Design sequences from the same backbone with ProteinMPNN and ESM-IF1, then refold both branches independently.

1. Prepare backbone. Clean the backbone and confirm chain identities, residue numbering, missing atoms, and intended oligomeric context.
2. Define fixed residues. Specify residues that must remain fixed, including catalytic, binding, disulfide, or interface positions.
3. Sample sequences. Run complementary inverse-folding models and sample multiple sequences at documented temperatures.
4. Refold designs. Refold candidates and compare local and global agreement with the supplied backbone.
5. Rank candidates. Review diversity, stability, solubility, and experimental constraints before selecting sequences.

Inputs and outputs

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

Inputs

Design input

PDBFASTAJSONTXT

A protein backbone in PDB format plus optional fixed-position and chain-design constraints.

Outputs

Design and review outputs

PDBFASTACSVJSON

Designed FASTA sequences, model scores, sequence alignments, refolded PDB files, and comparison results.