The chain's biggest risk is generalising a myeloid mouse result to infant B-ALL, and the graph cannot measure how much of it does
also called system census · model attribution gap · which fusion was it measured in
This project's most-documented mistake is taking a result from one leukaemia and one mouse model and quietly applying it to the infant leukaemia it actually studies. Trying to count how often the central mechanism story does that: it cannot be counted, because nothing in the graph records which system each piece of evidence came from.
The same thing, for a scientist
Measured over the seventeen entries that declare membership of `mechanism-kmt2a-fusion-causal-chain`: **110 evidence items citing 85 distinct PMIDs, 19 of which are shared by two or more members** [E1]. That first number is reassuring - the chain is broadly sourced rather than a compression of a handful of papers. The second attempt failed and the failure is the finding. **Assigning each citation to a fusion from its title leaves 59 of 85 unassigned**, because most papers in this field name neither the fusion nor the species in the title [E1]. **The graph carries no field that records the experimental system of an evidence item.** Model system appears, when it appears at all, inside free-text `caveats`. So the one audit that would most directly address this project's commonest documented synthesis failure cannot be run over its own data.
What could be measured, and what could not
| quantity | value | status |
|---|---|---|
| chain-member entries | 17 | measured |
| evidence items across them | 110 | measured |
| distinct PMIDs | 85 | measured |
| PMIDs shared by two or more members | 19 | measured |
| citations whose TITLE names MLL-AF9 | 10 | measured, but the instrument is weak |
| whose title names MLL-ENL | 10 | same |
| whose title names MLL-AF4 or AF4/AFF1 | 11 | same |
| whose title names no fusion at all | 59 of 85 | the problem |
| citations whose species can be determined from graph fields | 0 | not measurable |
| citations whose lineage (myeloid/lymphoid) can be determined from graph fields | 0 | not measurable |
Title-based attribution is not a system census and this entry does not present it as one: a paper titled MLL-rearranged leukemia is dependent on aberrant H3K79 methylation by DOT1L is a murine MLL-AF9 study and its title says neither. Reporting the 10/10/11 split as if it were a model audit would be precisely the kind of accurate-and-false number this project keeps having to retract.
Why this matters more here than anywhere else in the graph
The standing brief names three documented synthesis failures for this project. Two of them are system failures: generalising a myeloid MLL-AF9 result to infant B-lineage KMT2A::AFF1, and treating a single unreplicated paper as established. The second is now measurable - verify-claims.py and the load-bearing-citation analyses address it, and the 85-over-110 spread above is evidence against it for this chain specifically.
The first is not measurable at all. And the causal chain is where it would do the most damage, because a chain figure is the artefact that invites the generalisation: once six steps are drawn in one picture, the reader has no cue that step 1 is a 2004 murine MLL-ENL transformation assay and step 5's human evidence is an shRNA panel. mechanism-kmt2a-fusion-causal-chain therefore carries the system in every row of its own table, by hand, because no field holds it.
This graph already carries the qualitative rule in concept-what-transfers-from-an-mll-af9-result-and-what-does-not. What it does not have is any way to ask how much of a given argument rests on results that do not transfer.
The cheapest fix, and the reason it is not free
An optional system object on each evidence item - {species, lineage, fusion, in_vivo, primary_material} - would make the audit a one-line query and would cost one line per evidence item at writing time. Against that:
- Retrofitting 7,000 evidence items is not feasible by hand, and inferring the field from abstracts with a model would produce exactly the sort of confident, unverified annotation this project's identifier rules exist to prevent.
- Many papers use several systems, so the field is a list, and a list that is usually incomplete is worse than an empty one - it looks measured.
- The honest interim is what the chain entry does: put the system in the claim sentence, where a reader sees it, rather than in a field nobody fills.
A defensible partial version: require the field only for evidence items cited by a mechanism entry that carries a data.chain or data.step key. That is 110 items, one beat's work, and it covers the argument where the generalisation risk is concentrated.
| n member entries | 17 |
|---|---|
| n evidence items | 110 |
| n distinct pmids | 85 |
| n pmids shared by two or more members | 19 |
| title attribution is a valid census | false |
| species determinable from fields | 0 |
| lineage determinable from fields | 0 |
| chain | mechanism-kmt2a-fusion-causal-chain |
title attribution — the numbers behind the chart
{
"MLL-AF9": 10,
"MLL-ENL": 10,
"MLL-AF4": 5,
"AF4_or_AFF1": 6,
"t(4;11)": 2,
"no_fusion_in_title": 59
}fields recording experimental system
[]
proposed field
{
"name": "system",
"shape": {
"species": "human|mouse",
"lineage": "myeloid|lymphoid|other",
"fusion": "fusion-id or null",
"in_vivo": "bool",
"primary_material": "bool"
},
"scope_if_partial": "evidence items cited by mechanism entries carrying data.chain or data.step - 110 items"
}- A chain figure that cannot be system-audited will be redrawn in a grant application without the per-row system column, and the caveat does not survive the redraw - which is the mechanism by which a murine result becomes a statement about infants.
- Adding a system field and filling it incompletely is worse than not having it, because a partially populated field reads as a measurement and licenses exactly the count this entry declines to make.
- Would a text-mining pass over the caveats field recover species and lineage for most evidence items, and could its output be held to the same verification standard as an identifier?
- How many of this graph's mechanism claims would change confidence tier if their evidence were split by species and lineage?
- Is 85 distinct PMIDs across 17 entries typical for this graph, or unusually broad? Without a baseline the number cannot be read as reassurance.
- E1strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.analysis
The seventeen entries declaring membership of the KMT2A-fusion causal chain carry 110 evidence items citing 85 distinct PMIDs, 19 shared by two or more members; attempting to attribute each citation to a fusion from its title leaves 59 of the 85 unattributed, and no graph field records species, lineage or model system for any evidence item.
Evidence census over the seventeen chain-member entries: 110 evidence items, 85 distinct PMIDs, fusion attribution attempted from citation titles · Oncofusion knowledge graph (internal measurement) 2026What this evidence cannot show
Measured 2026-09-20 over the live node files. The counts of evidence items, distinct PMIDs and shared PMIDs are exact. The fusion attribution is a keyword match over citation titles and is reported ONLY to demonstrate that titles are an inadequate instrument - 69% unattributed - not as a model census. The claim that no field records system is a claim about the schema in data/graph/SCHEMA.md and about the type-specific payloads, not about whether the information appears anywhere in free text; it frequently appears in caveats, where it cannot be counted.
Evidence that argues against this (1)
- The information is not absent from the graph, only unqueryable: many entries state the system carefully in caveats, and a text-mining pass over the caveats field would recover a large fraction of it. An instrument gap is not the same as an information gap, and calling it one would overstate the problem. Olsen SN et al. 2022
- E2established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.paper
Abnormal chromatin at MLL-AF9 target loci in leukaemia stem cells was found only for H3K79me2, and Dot1l inactivation downregulated direct MLL-AF9 targets while global gene expression remained largely unaffected.
MLL-rearranged leukemia is dependent on aberrant H3K79 methylation by DOT1L. · Bernt KM, Zhu N, Sinha AU, Vempati S, et al. · Cancer Cell 2011What this evidence cannot show
Abstract read 2026-09-20. Cited here as the worked example of the instrument's failure: this is a murine MLL-AF9 study whose title names neither the species nor the fusion, so it falls into the 59 unattributed citations while being one of the most system-specific results in the chain.
Evidence that argues against this (1)
- The same paper's conclusion was subsequently supported in a human MLL-AF4 context by H3K79 profiling that distinguishes KMT2A-rearranged ALLs from other ALLs, so an unattributed murine citation is not automatically a transfer error - which is exactly why counting them requires the field rather than a heuristic. Krivtsov AV et al. 2008
- E3established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.paper
A conditional Mll-AF4 mouse develops B-precursor ALL or AML whose expression profile overlaps human KMT2A-rearranged ALL, and human KMT2A-rearranged ALLs can be distinguished from other ALLs by their H3K79 profiles.
H3K79 methylation profiles define murine and human MLL-AF4 leukemias. · Krivtsov AV, Feng Z, Lemieux ME, Faber J, et al. · Cancer Cell 2008What this evidence cannot show
Abstract read 2026-09-20. Cited as the chain's only in-vivo AFF1-backbone evidence and as the counter-example to a naive title census: its title names MLL-AF4 and would be attributed correctly, which is why the 11 correctly attributed AFF1 citations look reassuring and are not representative.
Evidence that argues against this (1)
- This study is itself murine for its in-vivo arm, so even a correct fusion attribution does not settle the species question, and a two-axis field (fusion and species) is the minimum that would be useful. Bernt KM et al. 2011
why the chain carries the system in every row of its table by hand
the qualitative transfer rule exists; the quantity it would apply to cannot be measured
- supports →Necessary is not explanatory: four different questions that a knockout answers as if they were one
both propose the same schema repair - record the design and the system on the evidence item
MEASURES WHAT THAT ENTRY REPORTED: the spine beat found no field records species or lineage; this finds the prose does not carry it either in two-thirds of load-bearing items, so the proposed field cannot be backfilled by inference.