Biohub MCP ESM Atlas tools

The ESM Atlas is is a free resource that provides predicted 3D structures from ESMFold2 and interpretable feature annotations, such as function, from the ESMC language model for over a billion proteins. Ask your AI agent questions about proteins through the ESM Atlas, and Biohub MCP looks it up and finds the proteins most like it. To see the structure of a selected protein, use the structure viewer, and to obtain structure files and more detailed structure information, see Get more from the ESM Atlas API.

Understanding SAE features

The ESM Atlas describes each protein by its SAE features: patterns that the ESMC protein language model picks up in sequences, such as a catalytic motif or a transmembrane helix. A sparse autoencoder (SAE) pulled 16,384 of these patterns out of ESMC, and each one has a number, a label, and a strength for your protein, as Understanding SAE features explains.

What you can ask

Ask in your own words, and your agent picks the tools and passes each result to the next step in the process. For example:

What does the ESM Atlas know about PETase from Ideonella sakaiensis? Explain its three strongest SAE features and where in the sequence each one is most active.

Find the ESM Atlas cluster closest to human hemoglobin alpha and summarize it: how big it is, how well characterized it is, its Pfam domains, and its most common phyla.

Find proteins in the ESM Atlas that resemble PETase from Ideonella sakaiensis but come from clusters with no characterized Pfam domains, and name three SAE features they all share.

You can use a sequence or an accession number instead of a protein name in any of these prompts.

Which tools your assistant uses

When you ask to	Your assistant uses
Look up a protein you named, or its UniProt accession	`esm_atlas_search_uniprot`, which returns matching UniProt entries with their sequences.
Look up a UniParc or MGnify accession	`esm_atlas_lookup_accession`, which returns its sequence.
See what the ESM Atlas knows about a protein	`esm_atlas_get_protein_details`, which returns its ESM Atlas record, the confidence of its stored structure when there is one, and up to ten of its strongest SAE features, each with its label and description.
Find similar proteins	`esm_atlas_search_similar_protein_clusters`, which returns up to 12 cluster representatives with the closest SAE features, optionally only from clusters with no characterized Pfam domains.
Learn about a similar protein’s cluster	`esm_atlas_get_cluster_info`, which summarizes the cluster of a similarity search hit: its size, Pfam domains, and taxonomy.
Go deeper on one SAE feature	`esm_atlas_get_sae_feature_detail`, which explains the feature and lists the SwissProt proteins where it is strongest.
See the structure in 3D	`ui_show_structure`, which opens an interactive viewer. Viewing a single protein structure is free, while viewing complexes requires the use of folding tools and is billed to your Biohub account.
Fold a protein or score mutations	The folding tools or the mutation scoring tools, which are billed to your Biohub account.

How a lookup works

Your AI agent gets the sequence

When you name a protein or give a UniProt accession, your AI agent looks up its sequence with esm_atlas_search_uniprot. For a UniParc or MGnify accession, it uses esm_atlas_lookup_accession. You can also directly input a protein sequence.

The ESM Atlas looks for the exact sequence first

If the ESM Atlas has the exact sequence, you see protein details from the stored record. A variant or a trimmed sequence counts as a new sequence.

A new sequence gets its features computed

If the ESM Atlas doesn’t have it, protein details and the similarity search compute its SAE features for this request, for sequences of up to 2,048 residues.

Similar proteins come from clusters

The ESM Atlas groups proteins into clusters by their SAE features, not by sequence identity. One protein, the representative, stands for each cluster. The similarity search compares your protein’s features with each representative and returns the closest ones, each with a sequence your agent can pass to the other tools.

Cluster details work only for a cluster representative, such as a similarity search hit, and compute nothing. Any other sequence, including many well-known UniProt proteins, returns not_found.

A name can match several UniProt entries. Add the organism, as in “PETase from Ideonella sakaiensis”, or give the UniProt accession. Your assistant can tell you which entry it used.

Read the results

Result	What it means	Returned by
SAE feature `value`	How strongly the feature is active in the protein, scaled per feature so values compare across proteins. Its `label_reliability` follows from `value` alone: `high` at 1.0 or more, `moderate` from 0.5, and `low` below 0.5. A low value means the feature is barely active, so its label says little about your protein.	Protein details
`residue_regions`	Up to ten stretches where a feature is most active, strongest first. Positions start at 0, so add 1 to match UniProt and the structure viewer.	Protein details
`similarity_score`	How alike two proteins’ SAE feature profiles are, from 0 to 1. Hits below 0.5 are left out. It does not compare similarity between sequences.	Similarity search
`cluster_pct_characterized`	The percentage of a cluster’s members that match a Pfam domain with a known function. 0 means that there is no member which matches a domain.	Cluster details
`ptm`, `mean_plddt`	Confidence of the stored predicted structure, from 0 to 1, as Read the scores explains. Often `null`, even for well-known proteins, because many have no stored structure.	Protein details

For example, PETase’s strongest feature is feature 11080, labeled “AB hydrolase H-G-x-D motif”, with a label_reliability of high. Its strongest region has start 198 and end 205, which is residues 199 to 206 in UniProt numbering, ending at Asp206, one of PETase’s catalytic residues. Feature labels describe what a feature tends to capture, not an experimental annotation of your protein. ESM Atlas structures are predictions, not experimental structures.

Get more from the ESM Atlas API

The ESM Atlas tools return summaries, so the results stay small enough for your assistant to read. For full data, call the ESM Atlas API, which does not require an API key:

You need	Use
A stored structure’s PDB file, per-residue pLDDT, or per-residue feature activations	Get protein
Up to 100 SAE features, or chosen features by index	Get protein, with `topk_features` or `feature_indices`
Up to 100 similar proteins, or a different similarity cutoff	Similarity search, with `topk_results` and `min_similarity`
The hashes of a cluster’s members, to pass to batch lookup	Get cluster
Data for many proteins at once	Batch lookup
Every SAE feature’s label and description	List SAE features

Most endpoints identify a protein by its protein_hash instead of its sequence, as the API overview explains.

Sequences

Tools that take a sequence expect a raw amino-acid sequence in one-letter codes.

Use the 20 standard amino acids or X, B, U, Z, or O.
Letters are case-insensitive, and whitespace before or after the sequence, including line breaks, is removed.
Don’t include a FASTA header, or spaces or line breaks inside the sequence.

: and | are also accepted, but they stay in the sequence and change which ESM Atlas record matches, so leave them out of a single protein. Any other character returns invalid_input.

Search UniProt

esm_atlas_search_uniprot searches UniProt and returns protein records with sequences you can pass to the other tools.

Inputs

Input	Description
`query` (required)	Free text, such as `PETase Ideonella sakaiensis`, or UniProt query syntax, such as `gene:CDK2 AND organism_id:9606 AND reviewed:true`, up to 500 characters. A UniProt accession, such as `P42212`, or an entry name, such as `HBA_HUMAN`, is looked up directly.
`size`	How many entries to return, from 1 to 6. Defaults to `6`.

A search that matches nothing returns an empty proteins list, not an error.
A direct match on an accession or entry name returns that entry, whatever size is. An accession or entry name that UniProt doesn’t have runs as a text search instead.
An accession that UniProt has deleted returns one entry with a null name and sequence.

Results

Field	Description
`query_label`	A label of up to 80 characters made from the query with UniProt syntax removed, or `null`.
`proteins[]`	Matching records, at most `size`.
`proteins[].accession`	UniProt accession.
`proteins[].protein_name`, `gene`, `organism`	Recommended name, first gene name, and scientific name of the organism.
`proteins[].function`	UniProt function text, which can be empty.
`proteins[].sequence`, `sequence_length`	Amino-acid sequence and length. The sequence is `null` when it is longer than 4,000 residues.
`num_results`	Number of records returned.
`total_results`	Total matches UniProt reported.

Resolve an accession

esm_atlas_lookup_accession fetches the sequence for a UniParc or MGnify accession.

Inputs

Input	Description
`accession` (required)	One accession of up to 128 characters, in one of the formats below.

Database	Format
UniParc	`UPI` followed by 10 hexadecimal characters, such as `UPI0000002FB4`, the UniParc entry for GFP.
MGnify	`MGYP` followed by 12 digits.

Other formats, including UniProt accessions such as P42212, return invalid_input; esm_atlas_search_uniprot looks up UniProt accessions.

Looking up an MGnify accession often returns not_found, even for accessions the ESM Atlas returns. MGnify only returns sequences for the representative protein of each cluster in its latest release, and its clusters are different from ESM Atlas clusters. Paste the sequence instead. A similarity search hit already includes its sequence, so there is no need to look up its accession.

Results

Field	Description
`accession`	The accession you sent, with surrounding spaces removed.
`database`	`uniparc` or `mgnify`.
`sequence`, `sequence_length`	Amino-acid sequence and length.

An accession the database doesn’t have, or one without a usable amino-acid sequence, returns not_found. A sequence longer than 4,000 residues returns invalid_input.

Get protein details

esm_atlas_get_protein_details returns a protein’s ESM Atlas record and up to ten of its strongest SAE features.

Inputs

Input	Description
`sequence` (required)	One protein of 1 to 4,000 residues, as Sequences describes.

If the ESM Atlas has the sequence, the tool returns the stored record. If it doesn’t, the ESM Atlas computes SAE features for sequences of up to 2,048 residues, and features_computed_on_miss is true. A longer sequence that isn’t in the ESM Atlas returns sequence_too_long_for_features. This tool never predicts a structure; to predict and view one, use ui_show_structure with the same sequence, for up to 700 residues.

Results

Field	Description
`sequence`, `sequence_length`	The normalized sequence and its length.
`accession`, `header`, `source`	The protein’s ESM Atlas accession, header, and source collection, when the sequence is in the ESM Atlas. The collection isn’t always UniProt; `source` names it, such as `uniparc` or `img_m`.
`sae_features[]`	Up to ten features, strongest first, with the fields below.
`ptm`, `mean_plddt`	Confidence of the stored structure, as Read the results explains.
`structure_available`	Whether the ESM Atlas has a stored structure.
`structure_can_be_computed`	`true` when the sequence is 700 residues or fewer, the longest the structure viewer predicts when the ESM Atlas has no stored structure. It checks length only.
`features_computed_on_miss`	Whether the features were computed for this request.
`structure_size_bytes`, `residue_confidence_count`	Size of the stored PDB file and number of per-residue confidence values.
`structure_source`, `structure_model`, `folded_on_miss`	Fixed for this tool: `structure_source` is `atlas` with a stored structure and `null` without one, `structure_model` is `null`, and `folded_on_miss` is `false`.

Each entry in sae_features has:

Field	Description
`feature_index`	Feature number, 0 to 16,383. Pass it to `esm_atlas_get_sae_feature_detail`.
`label`, `description`	What the feature is thought to represent.
`value`	Activation strength, as Read the results explains.
`label_reliability`	`high`, `moderate`, or `low`, set by `value`.
`residue_regions[]`	Up to ten stretches where the feature is most active, strongest first: `start`, `end`, `peak_residue`, and `mean_activation`. Positions start at 0, and `end` is included. `mean_activation` is a raw activation, on a different scale from `value`.

Find similar protein clusters

esm_atlas_search_similar_protein_clusters compares a sequence’s SAE features with ESM Atlas cluster representatives and returns the closest ones.

Inputs

Input	Description
`sequence` (required)	One protein of 1 to 2,048 residues, as Sequences describes.
`top_k`	The most representatives to return, from 1 to 12. Defaults to `10`.
`uncharacterized_only`	`true` returns only clusters with no characterized Pfam domains. Defaults to `false`.

You can get fewer than top_k hits, because the ESM Atlas leaves out matches scoring below 0.5 and withholds hits are restricted by its guardrails .
The ESM Atlas can also predict structures for up to six hits that don’t have one yet, which adds time.

Results

Field	Description
`query_sequence_length`	Length of your sequence.
`hits[]`	Cluster representatives, most similar first.
`hits[].protein_accession`	Representative’s accession.
`hits[].sequence`, `sequence_length`	Representative’s sequence and length.
`hits[].similarity_score`	Cosine similarity of the SAE feature profiles, from 0 to 1.
`hits[].cluster_size`, `protein_name`	Number of proteins in the representative’s cluster, and the representative’s name, when available.
`top_features_across_results[]`	Up to 20 SAE features found in the hits, ranked by mean activation: `feature_index`, `occurrence_count`, and `mean_activation`. A top feature can come from a single hit, so use `occurrence_count`, the number of hits that have it, to see which features the hits share.
`restricted_count`	Number of hits withheld by ESM Atlas guardrails.

Get cluster details

esm_atlas_get_cluster_info summarizes the cluster of a cluster representative.

Inputs

Input	Description
`sequence` (required)	A cluster representative of 1 to 4,000 residues, such as a hit from `esm_atlas_search_similar_protein_clusters`. Other sequences return `not_found`.

Results

Field	Description
`accession`, `source`, `protein_name`	Representative’s identity.
`cluster_size`	Number of proteins in the cluster.
`cluster_pct_characterized`	Percentage of members with a characterized Pfam domain, 0 to 100.
`cluster_mean_domain_coverage`	Mean Pfam domain coverage across members, from 0 to 1.
`top_pfam_domains`	Map of Pfam ID to `count` and `name`, or `null`.
`representative_features[]`	Up to ten of the representative’s strongest SAE features, with `feature_index`, `label`, `description`, and `value`.
`taxonomy_info`	Lowest common ancestor `rank` and `name`, or `null`.
`top_phyla`	Map of phylum to member count, which can be empty or `null`.

Get SAE feature details

esm_atlas_get_sae_feature_detail explains one of the 16,384 SAE features learned from ESMC.

Inputs

Input	Description
`feature_index` (required)	A feature number from 0 to 16,383, as the other ESM Atlas tools return it.

Results

Field	Description
`feature_index`	Feature number.
`label`, `summary`, `description`	Short label and longer interpretations.
`category`	Broad category, such as `Catalytic function`, when available.
`activation_pattern`	Where and how the feature tends to be active, when available.
`exemplar_protein_families`	Example families, when available.
`uniref90_frequency`	Number of UniRef90 sequences where the feature appears.
`top_swissprot_activations[]`	Up to ten SwissProt proteins with the strongest activation: `uniprot_id` and `activation`.
`decoder_nearest_neighbors[]`	Up to five related feature numbers.

Limits

Limit	Value	When exceeded
UniProt query	500 characters	`invalid_input`
UniProt entries per search	6	`invalid_input`
Sequence in a UniProt result	4,000 residues	The sequence is `null`
Accession	128 characters	`invalid_input`
Sequence for protein details or cluster details	4,000 residues	`invalid_input`
Sequence for the similarity search	2,048 residues	`invalid_input`
Features computed for a sequence not in the ESM Atlas	2,048 residues	`sequence_too_long_for_features`
Similar proteins per search	12	`invalid_input`
Stored structure read by protein details	2 MiB	`dependency_unavailable`
Complete result	6 MiB	`resource_too_large`
Tool deadline	120 seconds	`indeterminate`

Errors

Code	What to do
`invalid_input`	Fix the input the message names, such as a FASTA header, a character that isn’t an amino acid, a UniProt accession sent to the accession lookup, or a value over a limit.
`not_found`	Check the accession or feature number. For an MGnify accession, paste the sequence instead. For cluster details, use a similarity search hit, not a UniProt sequence.
`sequence_too_long_for_features`	Send one domain of up to 2,048 residues instead, which the ESM Atlas treats as a new sequence. The error gives `actual_length` and `max_length`.
`restricted`	ESM Atlas guardrails blocked the sequence. Don’t try to work around it by trimming, mutating, or rephrasing.
`rate_limited`	The ESM Atlas asked callers to slow down. Wait at least `retry_after_seconds`, then make fewer calls at once.
`dependency_unavailable`	The ESM Atlas, UniProt, or MGnify was unavailable, too slow, or sent an unusable response. Rate limits at UniProt and MGnify also show up here. Wait and try again.
`indeterminate`	The deadline passed, or a feature computation or similarity search ended with an unknown outcome. Don’t retry automatically; wait, then ask once more.
`resource_too_large`	The complete result is over 6 MiB. Ask for fewer results.
`internal_error`	Try again and tell us if it continues.

See Troubleshooting for details about codes.

Biohub MCP

Tools

Skills

Security and support

Understanding SAE features

What you can ask

Which tools your assistant uses

How a lookup works

Read the results

Get more from the ESM Atlas API

Sequences

Search UniProt

Inputs

Results

Resolve an accession

Inputs

Results

Get protein details

Inputs

Results

Find similar protein clusters

Inputs

Results

Get cluster details

Inputs

Results

Get SAE feature details

Inputs

Results

Limits

Errors