SaProt
Score protein variants using both structure and sequence. Zero-shot scores require no additional training on variant data and are not stability ΔΔG values.
API/CLI tool ID: saprot
Understanding inputs and outputs
SaProt reads protein sequence and structure together to score how plausible the model considers specified mutations. V3A in variantsFile means replacing V with A at the third amino acid of the selected chain sequence. Positions are counted from 1 in the sequence; do not assume they match residue numbers in the original structure file.
When structureSource is predicted, structural information at residues with low pLDDT is masked. This is not done for experimental structures. An experimental B-factor relates to variation in atomic positions; it is not the same quantity as predicted confidence pLDDT.
Masking compares the structure file's B-factor column, read as a 0–100 pLDDT score, against plddtThreshold. A predicted structure file does not always carry a pLDDT score in that column. ESMFold's predicted_structure.pdb writes the same quantity as a 0–1 fraction, and ImmuneBuilder writes a per-residue predicted error in ångström (Å) rather than pLDDT. Passing either file as is with the default of 70 puts every residue below the threshold, masks all structural information, and leaves a result close to sequence only. In that case masked_fraction comes out near 1, so check that value before reading the result.
A higher score means the model considers the mutant amino acid more plausible relative to the original. Its units and meaning differ from ThermoMPNN ddG, so the two scores should not be compared directly. A high masked_fraction means a larger fraction of structural information was masked; it is not the probability that a mutation will fail.
Input settings
| Setting | Format | Required | Default | Description and constraints |
|---|---|---|---|---|
pdbFile | pdb | Required | — | Protein structure to score. foldseek turns it into structure tokens — Extensions pdb · cif |
chain | chain | Required | — | The single chain to score in the structure — Example "A" |
structureSource | dropdown | Required | — | For a predicted structure, mask structure tokens at low-confidence (pLDDT) residues. The decision reads the structure file's B-factor column as a 0–100 pLDDT score. An experimental structure is not masked because its B-factor is not pLDDT — Allowed values predicted · experimental, Example "experimental" |
plddtThreshold | number | Optional | 70 | Masks structural information for residues with pLDDT below this value. The model default is 70. If the B-factor column of the structure file holds a 0–1 fraction or an error in ångström, every residue is masked, so check masked_fraction. — Minimum 0, maximum 100 — Applies when: structureSource = "predicted" |
variantsFile | csv | Required | — | Variant list with a mutation column, for example V3A. Join multiple sites with a colon (V3A:K6M) — Extensions csv · tsv |
Output metrics
| Column | Display name | Range | Description |
|---|---|---|---|
mutation | Mutation | — | Exactly as written in the input, for example V3A |
score | Mutation score | — | Sum of log(mutant probability / wild-type probability). Higher means the model finds it more plausible. This is not a stability ΔΔG |
masked_fraction | Masked structure-token fraction | — | Fraction of residues masked as low confidence. A high value warns that less structure information was used |
See Reading results for metric definitions and interpretation. A range shows possible values; whether higher or lower is better depends on the metric.
Execution availability
Open an execution plan in your project to see available tool versions and costs. Review and approve the inputs, settings, and terms before running.
License
Code, model weights, and accompanying data have their own terms. The table summarizes terms from official sources; check your execution plan for the configuration you will use. License texts are available in Third-party notices.
| Component | License | Commercial use | Source | Notes |
|---|---|---|---|---|
| Code | MIT | Allowed, subject to notice requirements | License text | — |
| Model weights | MIT | Allowed, subject to notice requirements | Weight distribution metadata | The model card lists the currently used SaProt_650M_AF2 weights under the MIT license. |
| Code | GPL-3.0-only | Conditional; see notes | License text | The foldseek binary (release 10-941cd33), which converts structures into SaProt input tokens, is included in the execution image. Because it is GPL-3.0, it runs only on our servers and the image is not sent to customer servers |
SaProt is a protein language model that takes sequence together with foldseek structure tokens. Its zero-shot mutation score reflects evolutionary plausibility, not stability ΔΔG
Checked on 2026-11-04.