Turn a research question into a reviewable investigation.
A guided workspace for organising evidence, tracing mechanisms and drafting candidate hypotheses and experiment proposals. Generated artefacts stay separate from approved scientific knowledge — never auto-promoted to the curated graph. Literature is a local candidate workflow (not live PubMed bulk ingest); model adapters are planned / interface-only stubs.
6WORKBENCH STEPS
5FRONTIER TOPICS
partialPHASE E
Workbench
Candidate investigations only. Export for human review; nothing is automatically published to the knowledge graph.
Research investigation: question → evidence → mechanism → hypothesis → experiment → review. Every artefact is labelled candidate · draft. Nothing auto-promotes into the curated graph or marks a claim experimentally supported.
Paste a literature project JSON, curation-candidate manifest, or a prior workbench export. No PMIDs or model scores are invented; rows stay draft candidates.
Search live article metadata from five Nature Portfolio journals. Proposing an article starts an unreviewed candidate in the literature workflow below — it never becomes approved evidence without curator review. Saving imports to the shared evidence store is not yet available (database configuration pending).
Article metadata from Crossref for Nature Immunology, Nature Reviews Immunology, Nature, Nature Medicine, Nature Microbiology. Titles, authors, dates, DOIs and licences come from the publisher's deposit; abstracts appear only when deposited. Full text stays on the publisher site. These records are source material, not reviewed evidence. No endorsement by, or affiliation with, Nature Portfolio is implied.
Literature intelligence
Phase C workflow — local metadata + candidate claims. Not live bulk PubMed ingest; export candidates or send them into the workbench evidence step (still drafts).
Literature intelligence (PRD FR-036/037): identify a source, validate metadata locally, extract candidate claims, map to canonical graph IDs, flag conflicts, then prepare records for human curation. Identifiers are optional and never invented. Accepted items stay candidates — they do not approve graph claims. This is not live PubMed bulk ingest; server-side PubMed/Crossref checks run only after you import a manifest in Curation.
Live PubMed/Crossref verification runs in Citation curation after import — not in this browser scaffold.
Curation handoff
0 claim(s) accepted · 0 exportable row(s). Export a candidate manifest for Citation curation. Import adds unreviewed rows only — scientific claim approval stays a separate human step. This UI never writes approved evidence.
Phase D interfaces for future omics, structure, pMHC/epitope, genomics, and agent adapters. Badges read planned / interface-only · not live — see also Methods §4.
Advanced analysis adapters are modular contracts for omics, structure, pMHC/epitope, genomics, and agents. None are live predictors in this build (implementedModels() is empty) — dry-run always returns typed not_available and never invents scores. Also documented on Methods & trust.
Honesty: 5 interface-only · 2 planned · 0 implemented · 2 molecular stubs (planned · not live) · badges say “not live” until a real adapter ships
OmicsInterface only · not live
Single-cell & spatial omics analysis
Map experimental profiles onto Immune Super Intelligence concepts for hypothesis generation.
ID sc-spatial-omics · no runtime version · stub only · Interface contract only — no omics runtime.
Not implemented; No structure prediction runtime in this app; UI documents the contract only; Never treat interface previews as fold/dock results
pMHC / epitopeInterface only · not live
Peptide–MHC / epitope predictors
Support vaccination and T-cell recognition exploration.
ID peptide-mhc · no runtime version · stub only · Interface contract only — no binding ranks.
Related concepts (navigation only): EG01
Inputs
Peptide sequences; HLA / MHC alleles
Outputs (documented)
Binding ranks; Candidate epitope list
Limitations
Not implemented; Predictor outputs must never be labelled as validated epitopes without lab confirmation; UI documents the contract only; Never treat interface previews as binding ranks
GenomicsInterface only · not live
Genomic variant analysis
Link coding variants to PID / HLA / pathway concepts in the graph.
ID genomic-variants · no runtime version · stub only · Interface contract only — no clinical variant caller.
Related concepts (navigation only): EG07, EG12
Inputs
Variant calls (VCF-like); Gene / HLA context
Outputs (documented)
Mapped immunogenetic concepts; Uncertainty flags
Limitations
Not implemented; No clinical interpretation pathway; UI documents the contract only; Never treat interface previews as clinical variant calls
AgentsInterface only · not live
Specialised scientific agents
Orchestrate retrieval → mechanism → hypothesis drafts under human review.
ID scientific-agents · no runtime version · stub only · Interface contract only — no agent runtime.
Related concepts (navigation only): C003, M001
Inputs
Scoped research question; Allowed tool list
Outputs (documented)
Candidate hypotheses; Experiment proposals marked as candidates
Limitations
Not implemented; Candidates are never established findings; UI documents the contract only; Never treat interface previews as agent conclusions
StructurePlanned · not live
TCR–pMHC docking (molecular stub)
Future structural context for TCR recognition hypotheses under human review.
ID tcr-pmhc-dock · no runtime version · stub only · Planned molecular stub — contract only.
Related concepts (navigation only): EG01, C014
Inputs
TCR sequence or structure handle; Peptide–MHC complex identifier
Not implemented; Planned molecular stub — no docking engine, no scores, no poses; Must not be labelled as immunogenicity or clinical binder prediction; Promotion to implemented requires rights + held-out eval + tests
StructurePlanned · not live
Cytokine–receptor MD refine (molecular stub)
Educational structural context for cytokine signalling hypotheses — never a dynamics result.
ID cytokine-receptor-md · no runtime version · stub only · Planned molecular stub — contract only.
Not implemented; Planned molecular stub — no MD engine or energy scores; UI documents a future contract only; Never treat previews as simulation trajectories
Implementation phases B–E
Status reflects shipped, testable code — not marketing claims. See also Methods & trust.
Phase Bpartial
Knowledge-layer gap fills
draft-ext-4 adds IVIG, pharmacogenetic HLA screen, herd-protection and post-infectious flare teaching drafts; domain coverage panel + gap queue on Evidence. Exemplar claim-review: TNF→RA on-path done; PD-1 partial; IFN→SLE still scaffold-ish — curator work, not silent graph completion.
New records live in extensions with draft:true at merge time.
Immunogenetics lesson + route appear in curriculum merges.
Graph release snapshot separates core vs draft counts; extensions version bumps.
Phase Cpartial
Scientific literature intelligence
NOT LIVE bulk ingest. Local workflow + Research UI + curation-candidate export only. PubMed/Crossref identifier verification remains the existing server-side curation path after human import — this scaffold never invents PMIDs or pretends bulk retrieval is wired.
Browser-safe workflow module creates versioned candidate artefacts with conflict flags.
Review queue UI maps to graph IDs, sets claim_support, accept/reject/edit with candidate labelling.
Curation handoff export matches importManifest shape without auto-approving claims.
Phase Dpartial
Advanced model registry interfaces
NOT LIVE predictors. Contracts + stub dry-run + request-shape validation + promotion gate + planned molecular stubs on Research/Methods. UI labels planned/interface-only; implementedModels() stays [].
MODEL_REGISTRY rendered with status badges, request-shape previews, inputs/outputs, limitations.
invokeModelStub / dryRunAllStubs refuse all current adapters with scientificClaimAllowed=false.
implementedModels() remains empty until a real adapter ships.
Phase Epartial
Scientific Workbench steps
NOT WIRED to KG write-back. Six-step machine with Ask/lit/mechanism imports, local draft save, versioned export (wb-1), plus human-gated curation-queue handoff download. No lab execution or auto-promotion.
Workbench UI first-class on Research layer (not a 13th nav item).
Hypothesis and experiment proposal artefacts stay candidate/draft_proposal.
Export includes schema + graph/extensions version identifiers and candidateFlags.
CD19 CAR-T produced drug-free remissions in refractory SLE, myositis and systemic sclerosis case series. Open questions: durability, toxicity, cost, and whether a reset is truly achieved.
