Trace an adverse outcome pathway from a molecular initiating event to an organ finding

You have a compound and an organ-level finding, and you need a mechanistic account you can act on: which molecular initiating event plausibly starts it, which key events sit between there and the finding, and how well-evidenced each link actually is.

   
Problem class Knowledge synthesis
Subject areas Chemistry, Drug Repurposing and Discovery, Translational Medicine
Evidence level Proposed
Complexity One skill or MCP
Availability Fully open
Compute Laptop

Problem

A tox study comes back with a finding — hepatic steatosis, reduced anogenital distance, thyroid follicular hypertrophy — and the next meeting asks why. The answer determines what happens next: a receptor-binding assay if the mechanism runs through a nuclear receptor, a mitochondrial panel if it runs through bioenergetics, a read-across argument if the pathway is shared across an analog series, and nothing at all if the finding is a species-specific artifact. Getting this wrong is expensive in both directions: chasing the wrong mechanism burns a quarter, and dismissing a real one shows up in a regulatory filing.

The adverse outcome pathway framework is the structured form of this reasoning — molecular initiating event, an ordered chain of measurable key events, an adverse outcome, with a weight-of-evidence call on each link (Tanabe et al. 2025). AOP-Wiki holds 595 AOPs, 1,939 key events and 3,168 key-event relationships (fetched 2026-08-09), but the content is uneven: some pathways are OECD-endorsed and quantitatively supported, others are one group’s draft, and progress toward fully endorsed AOPs has been slow (Holmer et al. 2024). “Solved” here means a committed table where each candidate pathway is named with its OECD status, each link carries its own evidence grade, and the chemical-specific evidence connecting your compound to the pathway is a separate column from the pathway’s own biology — because those two things are almost always at different strengths.

  1. Install the Adverse Outcome Pathway skill, which needs the ToolUniverse MCP server first — the catalog pages own both install commands. The skill sets disable-model-invocation: true upstream, so invoke it explicitly (“use the adverse-outcome-pathway skill”) rather than relying on automatic dispatch.

  2. Write the query from the finding, not from the compound. AOP discovery searches by organ, mechanism, or receptor keyword (AOPWiki_list_aops) — the pathway index is organized around biology, not chemistry. Put the observed endpoint, the target organ, the species, and the study type into finding.yaml, along with the exact search terms you intend to use. Commit that file first. A pathway list generated from remembered search terms cannot be re-derived, and the term choice is the single largest determinant of which pathways come back.

  3. Commit the workflow as a slash command. Have Claude Code write .claude/commands/trace-aop.md capturing the instruction, the output schema, and the gates below, so the next finding is traced by the same text. Run it with /trace-aop.

  4. Record OECD status per candidate pathway before reading any of its content. aop_candidates.csv carries aop_id, title, oecd_status (endorsed / under review / in the workplan / under development), mie, adverse_outcome, and retrieved_utc. An AOP under development is a hypothesis one group wrote down; an endorsed one has been through the OECD review process that exists precisely to make its links usable for decisions (Tanabe et al. 2025). Sorting by status before you read titles keeps you from anchoring on a well-written draft.

  5. Emit one row per key-event relationship, never one per pathway. key_events.csv carries aop_id, upstream_ke, downstream_ke, adjacency, weight_of_evidence, and quantitative_understanding as reported by AOPWiki_get_aop. The KER is the unit of causal knowledge, and it is the unit that gets developed and reviewed on its own (Holmer et al. 2024). A pathway summarized as “moderate evidence” hides whether the weak link is the one you were planning to assay. Report the weakest link alongside the pathway, not the average.

  6. Make stressor absence an explicit value, not a blank. AOP-Wiki lists 871 prototypical stressors across 595 AOPs (fetched 2026-08-09) — most AOPs name a handful of exemplar chemicals or none, so your compound being absent is the ordinary case and carries almost no information. Record stressor_match as one of listed, not_listed, or aop_lists_no_stressors. Collapsing the last two into one blank cell is the mistake that turns a sparse curation field into a false negative.

  7. Keep chemical-specific evidence in its own table, and split direct from inferred. chemical_evidence.csv holds the CTD gene and disease hits (CTD_get_chemical_gene_interactions, CTD_get_chemical_diseases) with an association_type column that is either direct_curated or inferred. CTD’s inferred chemical-disease links are transitive constructs assembled through a shared gene, and they outnumber the directly curated ones by orders of magnitude — enough that CTD itself had to build a network-topology score to rank them (King et al. 2012). An inferred link is a hypothesis generator; presented next to a curated one without a label, it reads as corroboration.

  8. Put the hazard layer in a third table and do not let it validate the mechanism. GHS classification, IARC status and LD50 values (PubChemTox_*) tell you the compound is hazardous; they say nothing about which pathway produced your finding. Keep them in hazard_context.csv for the reader’s orientation.

  9. Grade each candidate pathway three ways, defaulting to the middle. In aop_report.md, label every candidate supported, plausible_unsupported, or not_supported, defaulting to plausible_unsupported. Reserve supported for pathways that are OECD-endorsed and have your compound as a listed stressor or a direct curated CTD interaction with an MIE gene. Every verdict must cite the specific key_events.csv and chemical_evidence.csv rows behind it; a graded pathway with no row references is a model opinion. Then state the one assay that would discriminate between the surviving candidates — that is what the reader came for.

  10. Record provenance. provenance.json carries the AOP-Wiki release version (2.8, released 2026-03-08) with the UTC query timestamp, the ToolUniverse version, the skill commit SHA, the sha256 of finding.yaml, and the model id. These are live APIs and nothing here is byte-reproducible; the release version plus timestamp is what makes divergence visible. Commit .claude/commands/trace-aop.md, finding.yaml, all four tables, aop_report.md, and provenance.json. See the reproducibility guide for the pattern.

An AOP is dose- and kinetics-free by construction. It describes the biology once the MIE is engaged; it does not tell you whether your exposure reaches the MIE at the concentration your study used. Anything in the report that reads as a potency claim is not coming from AOP-Wiki.

Why this assembly

Rung 2. The work is retrieval and structured assembly across AOP-Wiki, PubChemTox and CTD, and the Adverse Outcome Pathway skill already encodes the four-phase ordering, the tool names, and an evidence-grading frame. Rung 1 fails on grounding, and it fails invisibly: plain Claude Code will produce a confident, plausibly-ordered key-event chain from pre-training, complete with pathway-sounding names that do not correspond to any AOP-Wiki key event, and nothing about the prose reveals which links were retrieved. Rung 3 would add a second mechanistic source, but every additional retrieval layer widens the report rather than sharpening the verdict, and the verdict is the deliverable.

Availability

Fully open on licence. ToolUniverse is Apache-2.0; AOP-Wiki, PubChem/PubChemTox and CTD are free public APIs with no account or subscription.

The access bar is disclosure. The compound identity and the observed finding are sent to third-party public services, which rules this out for confidential chemistry or an unpublished safety signal. There is no offline substitute — the pathway and toxicogenomics content is the databases, so a local mode is not merely absent but not constructible. If the compound cannot leave your machine, you can search AOP-Wiki by endpoint alone (steps 2 and 4–5, which are chemical-agnostic) and simply omit steps 6–8, recording their omission in the report so that a missing chemical-evidence table does not read as a negative result.

Compute requirements

Laptop. Everything is API-bound and the wall clock is network round-trips: budget a few minutes per finding across the four phases. A broad search term (“liver”) returns dozens of candidate pathways and each AOPWiki_get_aop call is a separate request, so narrow the term before widening the batch. This recipe takes one finding at a time by design — it is not a screening tool.

Evidence

Proposed. No documented attempt at this assembly — Claude Code driving the ToolUniverse Adverse Outcome Pathway skill over an observed toxicity — is known, and the skill has not been benchmarked end-to-end. The catalog records it as verified: works (2026-07-20), which establishes that the tool calls resolve, not that the graded pathway is correct. The closest evidence is framework- and component-level, and each gate traces to a specific result:

  • Step 4’s status gate reflects the OECD AOP Programme’s own position that structured, reviewed development is what builds confidence in an AOP’s applicability to decisions, and the fact that the programme runs a coaching effort and “gardening” of redundant key events precisely because AOP-Wiki content quality is uneven (Tanabe et al. 2025).
  • Step 5’s per-KER rule follows the pragmatic development approach now used in the field, in which the key-event relationship — not the whole pathway — is the essential unit of knowledge from which causality is inferred, and is developed and systematically evidenced on its own (Holmer et al. 2024).
  • Step 6’s three-value stressor field is arithmetic from the current AOP-Wiki metrics: 871 prototypical stressors against 595 AOPs and 1,939 key events (fetched 2026-08-09) means stressor lists are exemplars, not coverage.
  • Step 7’s direct-vs-inferred split is CTD’s own distinction. Inferred chemical-disease relationships are generated by chaining a curated chemical-gene statement to a curated gene-disease statement, they grow superlinearly with curation, and CTD built a local-network-topology metric to rank them because their raw volume is not usable as evidence (King et al. 2012; Davis et al. 2009).

Alternatives considered

Profile the compound instead of the finding. If the question is “what is known about this compound’s liabilities” rather than “why did this happen”, use triage-a-compound-toxicology-and-hazard-profile, which drives the Chemical Safety skill over a shortlist and touches AOP-Wiki as one retrieval among eight. That recipe is compound-keyed and produces a dossier; this one is finding-keyed and produces a mechanism with a next assay. Run that one first if you have a shortlist and no findings yet.

Predict endpoint values instead. predict-admet-properties-for-a-compound-series gives calibrated per-endpoint predictions across a series and will rank compounds on predicted liability far faster. It cannot tell you how — reach for it upstream, when you are choosing which compounds to run, not downstream when one of them has already surprised you.

Off-target explanations for a clinical signal. If the finding came from patients rather than a tox study, scan-adverse-events-for-drug-safety-signal and profile-compound-polypharmacology address the disproportionality and target-engagement halves of that question respectively; an AOP trace is the mechanistic layer you add after one of those produces a candidate target.

See also

Sources

Tried this recipe?

Share feedback — what worked, what didn’t, what you’d change. The form opens with this recipe pre-selected and a link back to this page.