Skin Sensitization AOP: Assays, Key Events, and Where dismech Disagrees

Scope

This page carries the AOP 40 use case out of AOP_EMOD_ALIGNMENT, where it began as a third pilot alongside the lead and liver-fibrosis cases. It outgrew that page for a reason worth stating: the other two pilots ask whether dismech content can be aligned to an AOP. This one found a place where dismech content bears on whether an AOP's own Key Event assignment is right, which is a different and more useful relationship. The alignment project keeps the schema questions; this page keeps the worked case.

The target is AOP 40, Covalent Protein binding leading to Skin Sensitisation. Unlike the other two pilots it enters from a review article, and it is the only case where the assay-to-Key-Event mapping is already published: Aleksic M, Rajagopal R, de-Ávila R, Spriggs S, Gilmour N. The skin sensitization adverse outcome pathway: exploring the role of mechanistic understanding for higher tier risk assessment. Crit Rev Toxicol 2024;54(2):69–91 (PMID:38385441, doi:10.1080/10408444.2024.2308816). Open access under CC BY-NC-ND.

Structurally it is the inverse of the liver fibrosis case: there, one microphysiological system instrumented several Key Events of one AOP; here, a whole OECD test-guideline battery exists in which each guideline is aligned to a single Key Event, and the integration happens in a layer above the assays.

How to read the snippets

Snippets are verbatim substrings of the article's PDF text layer, with three normalizations and no others: line-break hyphenation rejoined, the superscript TM left inline as the text layer emits it (KeratinoSens TM, GARD TM skin), and — where marked ‡ — the running header and page number removed where a sentence crosses the page 3/4 break. The space in OECD TG 442 C is the article's own. These have not been put through dismech's reference validator, because AOP-Wiki aside, this article's full text is not in the reference cache: references_cache/PMID_38385441.md and references_cache/DOI_10.1080_10408444.2024.2308816.md both hold abstract-only records (content_type: unavailable on the DOI record), and the publisher refuses automated fetches. Treat them as transcribed, not machine-verified.

Key Events against the assays aligned to them

Key Event titles and identifiers are AOP-Wiki's, read from the 2026-10-03 bulk XML export; the article itself uses only KE1–KE4 and does not cite AOP-Wiki identifiers, so the first two columns are a join, not a quotation.

Every validated guideline assay is aligned to exactly one Key Event, and two of the five Events have no validated non-animal assay at all. The second clause is the article's own emphasis, stated twice:

Currently, there are no NAMs addressing KE4 (the activation and proliferation of a T cell response), which are sufficiently progressed for implementation in OECD TG or for use in a Next Generation Risk Assessment (NGRA)

It remains difficult to envisage robust, sensitive, and reproducible assays addressing this KE in vitro in the near future.

Methods that sit beside the battery

Named in the article, aligned to a Key Event's biology but not carrying a guideline.

One assay is aligned to a Key Event its biology does not match

Worth carrying out of the table, because it is a documented misalignment rather than a clean row, and it is the kind of thing a KE cross-reference would have to be able to record:

Interestingly, the IL 8 assay, which is the only OECD validated cytokine assay (Watanabe et al. 2007 ; Sand et al. 2018 ) was developed as a marker for APC activation (KE3) rather than keratinocyte inflammation (KE2), possibly due to the cell type used in the assay

The article generalizes the same caution across all of TG 442E — "While read outs from all these assays have been aligned to KE3, some of these assays are also representative of the events leading up to KE3." So KE3 alignment for this battery is an assignment, not a measurement boundary.

Models and approaches spanning more than one Key Event

The article treats single-assay-to-single-KE alignment as the floor, not the usable unit:

With the exception of the GARD TM skin assay in certain circumstances (JRC EC 2021 ), these NAMs should not be used alone for hazard and potency categorization, but the output should be combined with information from other NAMs within a DA.

Two different kinds of thing span several Key Events, and they are worth keeping apart.

An assay that itself reads out several Key Events. This is the only place the article uses the phrase:

SENS-IS and GARDskin TM dose response assays (Cottrez et al. 2015 , 2016 ; Gradin et al. 2021 ), which cover multiple KEs, show some promise in providing more quantitative skin sensitizing potency assessment by predicting a skin sensitization class (strong/moderate/weak/non-sensitizing). However, it remains to be seen how these can be utilized in risk assessment.

A defined approach (DA) that integrates outputs from assays aligned to different Key Events. These are multi-KE by construction rather than by readout.

Three things this table deliberately does not say.

• "SARA-ICE" is not the article's name for it. The article says "SARA model" throughout and describes the NICEATM collaboration as an adaptation addressing UN GHS classification, citing a manuscript in preparation (Reynolds et al. 2023). It never writes the string SARA-ICE. If that name is wanted here it needs its own source.
• No DA's Key Event inputs are enumerated. The article says DAs combine outputs from assays that are each aligned to a Key Event, which makes them multi-KE, but it does not state which assays feed SARA, the Bayesian ITS, the ANN, or the regression models. Writing a per-model KE list from this source alone would be invention.
• The three OECD 497 DAs are unnamed in the article. Their identities (the 2o3 and ITS-style approaches) are not quotable from here.

Where the article's own KE4 mapping lives

The LLNA-to-KE4 mapping in the table above is stated only in the Figure 2 caption, with a correction attached ("rather than established sensitization"). The body text says just that the LLNA is the sole validated assay measuring T cell proliferation, without naming a Key Event. Anyone reusing that row should know the alignment rests on a caption.

No assay is aligned to the adverse outcome anywhere in the article. Patch testing appears only as diagnosis of ACD in patients, which is a different claim from an assay instrumenting a Key Event.

What does dismech tell us about IL-8?

The misalignment above is stated in the article as a suspicion — the assay "was developed as a marker for APC activation (KE3) rather than keratinocyte inflammation (KE2), possibly due to the cell type used in the assay." AOP 40 nonetheless has the whole of TG 442E, IL-8 Luc included, aligned to KE3. dismech holds an independent body of curated IL-8 mechanism, assembled from disease literature by curators who were not looking at AOP 40, and it can be asked what cell it thinks produces IL-8 and what it thinks follows.

The census has two tiers, and the distinction matters for how much weight the answer carries. 54 pathophysiology nodes across 46 files in kb/ mention IL-8 somewhere. Of those, 14 nodes across 13 files carry it structurally — bound in a descriptor, named in the node's own name, or named in a causal edge's description — and those are the ones below. The other 40 nodes across 37 files mention it only in free-text prose inside the node description, where no query can reach it and nothing ties it to a particular cell.

Among the 14 structural nodes, the bound cell_types are:

No structurally-annotated IL-8 node in dismech places production in a dendritic cell, and in the description-only tier exactly one node does. That node is Pouchitis → Mucosal Immune Dysregulation, which binds neutrophil, dendritic cell and T cell on a node whose description names four cytokines at once ("TNF-alpha, IL-1beta, IL-6, IL-8") alongside neutrophil infiltration and Th1/Th17 activity, and which has no outgoing edges. So the dendritic cell there is part of the node's cellular cast, not an assertion that a dendritic cell produces IL-8 — but it is the one place in the corpus where the two appear on the same node, and the claim has to be stated with that exception rather than without it.

Keratinocyte remains the most frequent single answer. That is not a refutation of the KE3 alignment — a cell line's readout is a property of the assay, not of the disease literature — but it is independent corroboration of the article's own doubt, from a corpus with no stake in the question.

Almost every IL-8 causal edge in dismech ends in neutrophil recruitment. Ten entries, each curated separately — eight of them terminating in a recruitment or infiltration node:

Source node and downstream target are written exactly as they appear in the file, so each row can be checked by grepping the entry.

Two rows are the exceptions, and neither points at a dendritic cell. Atopic Dermatitis routes IL-8 into Th2 polarization and IgE alongside IL-6 and IL-33 rather than to neutrophils. Familial Pityriasis Rubra Pilaris names CXCL8 on the upstream edge — NF-kappaB driving its transcription — so that edge ends in cytokine production itself, and the chain then continues into IL-23/IL-17 amplification rather than into recruitment. Nothing in the corpus routes IL-8 to dendritic-cell activation, which is the Key Event the assay reading it out is aligned to.

Where IL-8 appears as a measurement rather than as a mechanism, it carries no direction. Two entries hold it in biochemical — Shigellosis, as one of a group ("Inflammatory cytokines (IL-1β, IL-18, IL-8)") with frequency: VERY_FREQUENT and the elevation stated in notes, and Amniotic Fluid Embolism, whose presence field reads "Markedly increased in an individual fatal case". Neither sets a direction. So the corpus's clearest statements that IL-8 goes up in a disease are the ones a query for an increased readout would miss entirely, which is the same encoding problem as the three pathograph spellings below, reached from the marker side.

Shigellosis holds it in both places, which is worth noting because the second one is invisible to the census above: its Inflammasome sensing and pyroptotic inflammation node names IL-8 release among the pyroptotic cytokines and binds macrophage and enterocyte, but only in free text and with no outgoing edges at all. That is the shape of the description-only tier in general — a real mechanistic claim, with real bound cells, that every structural query misses.

What this says to AOP-Wiki, and what it does not

The useful claim is narrow and worth keeping narrow. dismech cannot say which Key Event an assay should be aligned to; that is a judgement about what a cell line reports under a test guideline. What it can say is what the disease literature holds about the biology the assay is named for — and here that answer points at keratinocyte and epithelial production feeding neutrophil recruitment, which is the KE2 compartment and its immediate consequence, not dendritic-cell activation.

Two things follow that are worth putting to the AOP 40 authors:

• KE 826 (Activation, Keratinocytes) has a measurement-methodology field, and the IL-8 Luc assay is not in it. If the assay's readout is in part a KE2 marker, as the review states, the methodology text on the Key Event is where that belongs — independently of which guideline the assay sits in.
• The misalignment is a property of the edge, not of either node. Recording it needs a place to say "this assay is aligned to KE3 and its readout is a KE2 marker." Neither AOP-Wiki's Key Event records nor dismech's node annotations can currently hold that sentence, which is the same gap question 1 is about, reached from the opposite direction.

Three caveats on the dismech side

• The queryable signal is far thinner than the corpus mentions. IL-8 appears in 267 strings across 77 files in kb/, but only six of those are ontology-grounded descriptor bindings; 53 are evidence snippets and 50 are node descriptions. The finding above comes from reading prose, not from a query that could be re-run as a gate — and the first version of this section proved the point by missing a node. A census over node names, descriptor bindings and causal-edge descriptions does not see a mechanism stated only in a node's own description on a node with no outgoing edges, which is how Shigellosis's inflammasome node and 39 others were absent from the table until a reviewer found one of them.
• Increase is encoded three incompatible ways. SAPHO Syndrome and Spotted Fever Rickettsiosis bind GO:0032637 interleukin-8 production with modifier: INCREASED; Chlamydia Pneumoniae Pneumonia binds GO:0032757 positive regulation of interleukin-8 production with no modifier; Autoimmune Pancreatitis states it in the node name alone, binding no process descriptor at all. All three assert the same thing. A filter on modifier: INCREASED finds two of them, a filter on the GO term finds one, and nothing short of reading node names finds the third.
• The strongest-looking row is the most carefully hedged. Autoimmune Pancreatitis is the only node in kb/ named for increased IL-8 expression, and its curator recorded causal_link_type: INDIRECT_UNKNOWN_INTERMEDIATES, directness: INDIRECT, and an explanation stating that the evidence is differential qPCR between AIP subtypes and "was not a chemotaxis or IL-8 blockade experiment." Do not cite it as demonstrated causation. Subacute Inflammatory Demyelinating Polyneuropathy is hedged the same way: CSF IL-8 as a biomarker of blood-nerve barrier dysfunction, with the node description saying outright that the evidence is from GBS and CIDP and is "not SIDP-specific."

What this case adds to the alignment project

• It is the first case where the MIE is both instrumented and curated. The lead and liver pilots both found the molecular initiating event unmapped and unmeasured. Here KE1 is the most thoroughly assayed Event of the five — three guideline methods plus PPRA and four in silico tools — and Contact Dermatitis now carries a direct counterpart node, Covalent hapten-protein adduct formation (biological_scale: MOLECULAR), wired from Hapten exposure and backed by a peptide-adduct snippet (PMID:20631061). So for the first time across the three pilots, both ends of an MIE correspondence exist.

That node is also the clearest instance yet of the binding problem this project keeps running into: its notes record that GO has no term for non-enzymatic covalent modification of a protein by a xenobiotic electrophile, and that CHEBI files hapten as a role and so outside the bindable enum. An AOP MIE that nothing in GO or CHEBI can name is a different kind of gap from an unmapped one. - Most of the chain is curated, and the gaps are not where the earlier pilots' were. Against AOP 40's five Events, the allergic arm of Contact Dermatitis holds counterparts for KE1, KE3 (Skin innate immune activation during sensitization, binding Langerhans cell and dendritic cell) and KE4 (Allergen-specific T-cell priming, binding GO:0042110 T cell activation with CD8-positive T cell and dendritic cell). Two observations fall out of that:

KE2 is skipped on the arm that needs it. The entry does carry Keratinocyte activation and proinflammatory mediator release with GO:0032980 keratinocyte activation, which is a clean KE 826 counterpart — but it sits on the irritant chain (irritant exposure → barrier disruption → keratinocyte activation → inflammatory cell recruitment), not the hapten chain, which runs adduct formation → innate activation → T-cell priming with no keratinocyte step. AOP 40 is a sensitization pathway, so its KE2 belongs on the allergic arm. This is a concrete curation lead, and it is the same compartment the IL-8 section argues the IL-8 Luc assay is really reading.

The AO is not missing, it is fused with KE4. AO 827 sensitisation, skin is the established sensitized state, and that is what Allergen-specific T-cell priming describes — "establishing the adaptive immune memory required for later flares" — in the same node as the T-cell activation that is KE4. Past that point the entry continues beyond the AOP's scope into elicitation: re-exposure-driven recruitment, resident memory T-cell flare, and the Contact dermatitis phenotype (HP:0032282). So the earlier pilots' endpoint gap does not reproduce here. What replaces it is a harder modelling question: one dismech node standing for two AOP-Wiki Events. - It supplies two tests for any future Key Event cross-reference slot, and neither can be met by a bare identifier. A slot recording "this node corresponds to KE3" must also be able to record the IL-8 case, where the assay's own biology sits at KE2 while the alignment is at KE3 — and the KE4/AO case, where one node answers to two Events at once. The second needs an explicit aggregation relation, not just a list of identifiers. - It is the first case running in the other direction. The lead and liver pilots ask what an AOP can contribute to dismech. The IL-8 section asks what dismech's corpus says about a Key Event assignment the AOP's own review article flags as uncertain, and gets a usable answer. That is a second, separate reason to keep curating against AOPs, and it is the one that gives the collaboration something to send back upstream.

Next steps

• [ ] Put the IL-8 KE2/KE3 question to the AOP 40 authors, together with the measurement-methodology observation on KE 826.
• [ ] Decide whether the three encodings of increased IL-8 in kb/ should be normalised, and if so to which one. This is a corpus-wide question about modifier versus positive regulation of GO terms, not specific to IL-8.
• [ ] Check whether the same KE2/KE3 reading holds for the other TG 442E assays, which the article says are "also representative of the events leading up to KE3."
• [ ] Curate a keratinocyte-activation step onto the allergic arm of Contact_Dermatitis. The entry's existing GO:0032980 node sits on the irritant chain, so AOP 40's KE2 has no counterpart on the chain that corresponds to it. This is the one Event of the five with no node on the right arm, and it is the compartment the IL-8 section argues the IL-8 Luc assay is actually reading.
• [ ] Decide how a node standing for two AOP-Wiki Events should be recorded, using Allergen-specific T-cell priming (KE4 + AO 827) as the test case. This is the aggregation-relation question, and it is now a worked instance rather than a hypothetical.