biohub_get_inference_run
ChatGPTRead one fold run by run_id: its status, and once it is done, its scores on a 0-to-1 scale and what the fold was billed. Only the fold tools create runs: esmfold2_fold_max_accuracy always, and esmfold2_fold_complex when it returns queued; a fold that returned done has no run. It starts no compute and costs nothing. It waits on the server for at most 5 seconds, so a run still queued or in_progress is normal and can be read again later. A run not started within 24 hours is cancelled and not charged. Failed and cancelled runs carry Biohub's reason in error_message, and calling the fold tool again folds and bills again. Reading a run more than 31 days after submission fails with expired. It returns no structure file, coordinates or PAE; those come from the authenticated Biohub API or SDK.
esm_atlas_get_cluster_info
ChatGPTGet the ESM Atlas cluster a protein belongs to: cluster size, the share of members with characterized Pfam annotations, top Pfam domains, taxonomy and top phyla, and the SAE features typical of the cluster. ESM Atlas clusters group proteins by ESMC sparse-autoencoder (SAE) feature similarity, not by sequence identity. Use it to put a protein in context or judge how well studied its neighborhood is. Takes the raw one-letter sequence of a protein already in the ESM Atlas, such as a hit from esm_atlas_search_similar_protein_clusters; other sequences return not_found, because this tool computes nothing.
esm_atlas_get_protein_details
ChatGPTLook up what the ESM Atlas knows about one protein sequence: its source accession, the confidence of its predicted structure (pTM and mean pLDDT), whether a structure exists or can be predicted (up to 700 residues), and its strongest ESMC sparse-autoencoder (SAE) features, which are learned, labeled protein properties, with the residue ranges that drive them. Use it first for a sequence you already have. Takes a raw one-letter sequence; resolve a protein name or accession with esm_atlas_search_uniprot or esm_atlas_lookup_accession. Sequences not in the ESM Atlas get features computed for up to 2,048 residues. This tool never predicts a structure; ui_show_structure with the same sequence shows one. No saved ID or download is created.
esm_atlas_get_sae_feature_detail
ChatGPTExplain one ESMC sparse-autoencoder (SAE) feature by its index: its label, summary and description, category, activation pattern, example protein families, how often it occurs in UniRef90, the reviewed UniProt (Swiss-Prot) proteins that activate it most, and the most similar features. Use it when another ESM Atlas tool returns a feature_index worth understanding. It takes an index, not a sequence.
esm_atlas_lookup_accession
ChatGPTGet the amino-acid sequence for a UniParc (UPI...) or MGnify (MGYP...) ID, to pass to the ESM Atlas tools. Use it when a protein is identified by one of those IDs. UniProt accessions and entry names are not accepted here; esm_atlas_search_uniprot looks them up. Other formats return invalid_input.
esm_atlas_search_similar_protein_clusters
ChatGPTFind ESM Atlas protein clusters whose representatives are most similar to a sequence, compared by ESMC sparse-autoencoder (SAE) feature profiles rather than by sequence alignment. Use it to find functional or structural neighbors of a protein. Takes a raw one-letter sequence up to 2,048 residues. Returns up to top_k representatives above a minimum similarity, so fewer may come back, each with a similarity score, cluster size, name, and a sequence to pass to the other ESM Atlas tools, plus the features most shared across hits. restricted_count is the number of hits withheld by safety screening. Hits without a stored structure may be predicted during the search, which can add time.
esm_atlas_search_uniprot
ChatGPTFind proteins in UniProt by name, gene, organism, function, or accession, and get their sequences for the ESM Atlas tools. Use it when a protein is named rather than given as a sequence. Accepts free text, such as "human hemoglobin alpha", or UniProt query syntax, such as "gene:TP53 AND organism_id:9606 AND reviewed:true"; a UniProt accession (P69905) or entry name (HBA_HUMAN) is looked up directly. Returns up to size entries with accession, name, gene, organism, function, and sequence, plus the total number of matches. For UniParc or MGnify IDs, use esm_atlas_lookup_accession.
esmc_scan_mutation_tolerance
ChatGPTFind where a protein is most and least tolerant to change, when no substitutions are named yet. It masks each scanned position, scores all 20 amino acids with ESMC, and returns only the most constrained and most tolerant positions, ranked by ranking_metric, with their strongest substitutions, not the full matrix. These are predictions, not measured fitness. It bills the signed-in user's Biohub account one ESMC call per scanned position, so a scan costs positions times length divided by 10,000 credits: 9 for all of a 300-residue protein, and never more than 20. Each call starts new work and can bill again; an indeterminate failure may already have been billed. At the time limit, or on a rate limit after some positions finished, it stops starting calls and returns partial, billed for the finished positions only, with the rest in coverage.unscanned. Otherwise a failure returns an error, and positions that finished were still billed. It never shortens a long protein silently, and offers the 300M, 600M and 6B models.
esmc_score_mutations
ChatGPTScore substitutions the user already named, such as W43F or a list from a paper, in one protein with ESMC. Each is scored alone on the wild-type sequence, not as part of a combined variant. The score is ln P(substituted) - ln P(wild type) with that position masked, so a positive score means ESMC finds the new residue more likely there. These are predictions, not measured fitness, stability or activity. It bills the signed-in user's Biohub account one ESMC call per distinct position, at most 8, each costing the protein's length divided by 10,000 credits, so any substitutions at 3 positions of a 300-residue protein cost 0.09 credits. Each call starts new work and can bill again; an indeterminate failure may already have been billed. esmc_scan_mutation_tolerance finds positions worth testing when none are named.
esmfold2_fold_complex
ChatGPTPredict structure confidence with ESMFold2 for one protein, several protein chains (an antibody, a homodimer, a binder and its target), or proteins with DNA, RNA or ligands. It returns scores on a 0-to-1 scale: pTM, a pLDDT summary, interface pTM for the whole complex, and ipTM per chain pair when returned. Protein folds can also return mean pLDDT per chain; a warning says when it is unavailable. A high pTM with a low interface pTM means each part is confident alone but the contact is not. These are predictions, not measured binding. It never returns a structure or coordinates. For free, esm_atlas_get_protein_details gives the ESM Atlas pTM and mean pLDDT of a protein already in the Atlas, and ui_show_structure shows one protein chain with a single protein entity, predicting it if needed for up to 700 residues. Inputs of up to 768 tokens fold within the call and return done. Longer inputs, up to 4,096, or use_batch_api=true, return queued with a run_id at one fifth of the price, readable with biohub_get_inference_run for 31 days. It takes no alignment; esmfold2_fold_max_accuracy builds one. It bills the signed-in user's Biohub account, growing with the square of the length: about 0.45 credits for 300 residues within the call and 4 for 2,000 as a run. Each call starts new work and can bill again; an indeterminate failure may already have been billed.
esmfold2_fold_max_accuracy
ChatGPTPredict structure confidence at Biohub's highest accuracy, slower and costlier than esmfold2_fold_complex, for the same entities. Biohub builds its own alignment from UniRef30 and ColabFold, and the tool cannot take one; it folds five times with different seeds and keeps the most confident. It always returns queued with a run_id, readable with biohub_get_inference_run for 31 days; a run takes minutes to hours. Its scores, on a 0-to-1 scale, are pTM, a pLDDT summary, interface pTM and ipTM per chain pair when returned. It does not return per-chain pLDDT. It never returns a structure or coordinates. It bills the signed-in user's Biohub account when the run finishes, with no batch discount: about 0.28 credits for 56 residues, 1.6 for 300, 4.4 for 768, 14 for 2,000 and 37 for 4,096. By default the signed-in user's Biohub account can start 10 of these runs a day, and a run that fails is not charged but still counts. Each call starts new work and can bill again; an indeterminate failure may already have been billed.
list_resources
ChatGPTList the complete enabled Biohub MCP resource and resource-template catalog, including hosted Skills and published metadata. Returns resources with uri and resource_templates with uriTemplate; template URIs need concrete parameter values before reading. Does not read resource contents or start compute. Results exceeding 128 KiB of serialized tool output return resource_too_large without truncation.
read_resource
ChatGPTRead one enabled Biohub MCP resource by exact URI, including hosted Skill instructions, supporting files, and linked artifacts. Returns MCP contents with text or base64 blob, MIME type, and available metadata. Only registered resources are accessible; this is not a web or local file fetcher. Results exceeding 128 KiB of serialized tool output return resource_too_large without truncation. Skill reads start no compute.
ui_show_structure
ChatGPTShow a protein, or a complex of proteins, DNA, RNA and small-molecule ligands, in an interactive 3D viewer from sequences. Use it when the user wants to see or visualize a structure or how molecules fit together, such as a protein bound to ATP, a protein on DNA, or a homodimer, or when you need to look at one to answer a question about its shape. Give one entity per distinct molecule, in order, with copies for identical chains; give each DNA strand as its own entity; a ligand takes a CCD code or a SMILES string, such as [{"type": "protein", "sequence": "MKTAYIAKQR"}, {"type": "ligand", "ccd": "ATP"}]. For a protein name or accession, get the sequence first with esm_atlas_search_uniprot or esm_atlas_lookup_accession. One protein with one copy returns its ESM Atlas structure when one exists, or a prediction for up to 700 residues, at no cost. Anything else is a new fold with ESMFold2 on the Biohub Platform billed to the signed-in user's Biohub account, so call it once per request; a complex holds up to 768 tokens, one per residue or nucleotide and one per ligand heavy atom, counted for every copy, in at most 36 chains; larger complexes need the Batch API. Structures are computational predictions, not experimental. In clients that support MCP Apps, the user gets an interactive viewer they can rotate, recolor, and select residues in without another call; the result has each entity's chain IDs, pTM, ipTM and warnings, and only the viewer gets the coordinates. Nothing is saved and no link is created. The result has no image by default. Set include_png to true only when you need to see the structure yourself or the client can't show the viewer, because it adds several seconds. Without an image, don't describe how the structure looks.