Workflows

Multi-step investigations where the agent chains several tools, each step’s output feeding the next. The agent picks the tools itself from their descriptions (i.e. you don’t name them) but seeing the mapping shows how the tools compose.

For what each tool does on its own, see the by-tool examples.

Characterise an unfamiliar domain, end to end

Prompt: “I keep running into the ‘kringle domain’ in papers. What is it, where does it sit on the tree of life, which human proteins carry it, and is there a structure I can look at?”

Chain: searchInterPro → getEntry → getTaxonomicDistribution → getEntryProteins → getEntryStructures

  1. Find the entry. There’s no accession to start from, so the agent runs searchInterPro("kringle domain") and takes the best hit: the InterPro domain IPR000001 (“Kringle”).

  2. Understand it. getEntry("IPR000001") returns the definition, GO terms, literature, and the member signatures it integrates (Pfam PF00051, SMART SM00130, PROSITE PS50070, CDD cd00108), confirming it’s a domain, not a family.

  3. Place it on the tree of life. That same accession flows into getTaxonomicDistribution("IPR000001", depth: 3), whose superkingdom breakdown shows kringles are mainly eukaryotic: a metazoan module of the clotting and fibrinolysis machinery.

  4. List concrete carriers. getEntryProteins("IPR000001", reviewedOnly: true, taxId: "9606") returns reviewed human proteins containing it: plasminogen (P00747), prothrombin, hepatocyte growth factor, apolipoprotein(a), each with the residue ranges of every kringle copy.

  5. Get something to look at. getEntryStructures("IPR000001") lists the PDB entries that include a kringle domain, with title, method and resolution, so the agent can point you at a solved structure.

One accession, discovered in step 1, drives steps 2–5. You never type an accession by hand.

Identify an unknown sequence and find its relatives

Prompt: “I pulled this protein out of an assembly and don’t recognise it. What domains does it have, and what known proteins are built the same way? <your sequence>”

Chain: matchSequences → analyzeSequence → getEntry → searchDomainArchitectures

  1. Try the instant lookup first. matchSequences("<your sequence>") hashes the sequence to an MD5 and checks InterPro’s precomputed set. For something straight out of an assembly this often comes back found: false, i.e. it isn’t in UniParc yet.

  2. Run the real analysis. Because it’s novel, the agent falls back to analyzeSequence("<your sequence>"), which runs InterProScan 6 on the EMBL-EBI Job Dispatcher. Say it reports two domains: a Kringle (IPR000001 / Pfam PF00051) and a trypsin-like serine-protease domain (IPR001254 / Pfam PF00089). If the job doesn’t finish inside the wait window it returns a jobId; the agent retrieves it a moment later with getSequenceAnalysis.

  3. Confirm the key domain. getEntry("IPR001254") names it: “Serine proteases, trypsin domain” (peptidase family S1), with its member signatures and GO terms, so you know it’s a catalytic protease module, not just a fold.

  4. Find proteins built the same way. Feeding both accessions to searchDomainArchitectures("IPR000001-IPR001254", ordered: true) returns the known domain architectures that pair a kringle with a trypsin domain, and how many proteins share each, led by plasminogen (P00747), with HGF, HGF activator and coagulation factors among the relatives. That places your unknown protein squarely in the plasminogen / fibrinolysis family.

A raw sequence with no identifier becomes two accessions in step 2, and those accessions locate its relatives in step 4.