ESM2 35M

Represent a protein sequence as numerical features for downstream analysis.

API/CLI tool ID: esm2-35m-embedding

Input and output

The only setting is inputFile. Supply a UTF-8 JSON file with only the sequence key. The example below is a synthetic sequence illustrating the format.

{"sequence":"ACDEFGHIKLMNPQRSTVWY"}

Accepts 1–256 uppercase characters from ACDEFGHIKLMNPQRSTVWY, at most 4096 bytes per file. Extra or duplicate keys, empty sequences, lowercase, whitespace, FASTA headers, X, *, - and batch arrays are rejected. No normalization or truncation is performed. This length limit is a service policy, not the model capacity.

embedding.json is a bare array of 480 finite numbers. result.json records schemaVersion, modelId, modelRevision, inputSha256, sequenceLength, layer, pooling, dtype, device, shape, manifestSha256, assetSha256 and embeddingSha256. The raw sequence is not copied into the results.

Only the pinned facebook/esm2_t12_35M_UR50D revision 6fbf070e65b0b7291e7bbcd451118c216cff79d8 is used. Layer 12 is averaged over actual amino acid positions, excluding beginning, end and padding tokens. pooling is mean-residues-excluding-special-and-pad and shape is [480].

inputSha256 identifies the uploaded raw bytes. Even whitespace around the JSON changes the hash; retain the original file with model and weight provenance.

Interpretation and execution scope

An embedding represents a sequence as numerical features. Larger individual values are not inherently better. Comparison distances and downstream classifiers require separate validation. The tool does not return structural coordinates, binding affinity, function, efficacy or success probabilities; numerical reproducibility is not experimental validation.

Runs on managed servers with 2 CPUs, 2 GiB and a maximum of 120 seconds.

The conservative planning estimate is 30 seconds. Whole isolated paid native runs with normal delivery and response loss took 11–14 seconds; this is not a production distribution, time guarantee or minimum billing duration.

Documentation and public metadata do not authorize execution. Check the version, resources and rate in the execution plan, then review and approve the inputs, settings and terms.

Rights and delivery scope

Evidence covers the selected code MIT text and the MIT declaration in the pinned 35M model card, not other ESM models. Nadoo Bio runs it on managed servers and provides results; images, weights and binaries are not distributed to customers.

Retain notices, conditions, disclaimers and the supplemental notices associated with the image. Dependency review retains limits concerning native build resolution and Unicode source-data provenance; it does not warrant rights over the entire image, training data or patents.

Official sources

On this page