How much of this graph's mechanism is somebody else's fusion? A counted answer: 54% of fusion-bearing evidence, and 0% from a dedicated human KMT2A::AFF1 experiment
also called partner provenance audit · imported mechanism census · whose fusion is this
We took a random sample of the entries that carry mechanism in this graph and asked, for each one, which leukaemia fusion the evidence actually came from. Of the entries whose key experiment used a fusion at all, more than half used a fusion that is not the infant one this project is about, and not a single one rested on a dedicated experiment in human infant-type cells.
The same thing, for a scientist
**Denominator and sampling.** The mechanism-bearing folders are `mechanism-core` (26), `histone-marks` (45), `interactome` through `interactome6` (21/22/21/21/16/25), `subtraction` (14), `subtraction-kmt2a` (21), `localization` (24) and `cells` (41) - **297 entries**. We drew a stratified random sample of **4 entries per folder, 48 total (16.2%)**, with `random.seed(20260919)` over the sorted file list in each folder, and classified the **first evidence item** of each, which is this graph's convention for the load-bearing citation [E1]. **Result.** 24 of 48 sampled entries have load-bearing evidence obtained in a system that contains a leukaemogenic fusion at all; the other 24 rest on structural biology, normal-cell biology, yeast, *Drosophila*, database records or literature counts, where the question does not arise. Within those 24: | provenance of the load-bearing experiment | n | % of 24 | |---|---:|---:| | a KMT2A fusion other than the index one (AF9, AF6, EEN, ELL, EP300, ENL-line) | 10 | 42% | | a non-KMT2A fusion or a different disease (PICALM::MLLT10, NUP98-r, adult AML) | 3 | 12% | | **subtotal: somebody else's fusion** | **13** | **54%** | | mixed panels that include index-fusion lines (DepMap, multi-chimera co-IP) | 5 | 21% | | fusion present but not named by the source | 4 | 17% | | the index fusion KMT2A::AFF1 (murine *Mll-Af4*) | 1 | 4% | | the index disease, partner mixed (primary infant KMT2A-r) | 1 | 4% | **Zero of 24** rest on a dedicated experiment in a human KMT2A::AFF1 system. Human index-fusion material enters the sample only inside mixed DepMap panels, where it contributes four cell lines. Twelve of the 24 (50%) are murine and twelve are myeloid-only. **The flagging news is good, and it is not the problem.** Of the 13 'somebody else's fusion' entries, **12 state the mismatch in the evidence item's own caveats** - naming the fusion, the species, or explicitly disclaiming transfer to KMT2A::AFF1. One does not: `gene-leo1` (see `ana-partner-provenance-corrections`). An unflagged-import rate of **1 in 13 (8%)** is far better than the brief's premise expected [E2]. **So the failure is not per-entry honesty; it is that the flag does not aggregate.** Every entry discloses its own limitation and no entry discloses the pattern, so a reader who follows any single thread is correctly warned and a reader who reads the graph as a mechanism is not. This entry is the missing aggregate.
How we sampled, so someone can repeat it
folders = ["mechanism-core","histone-marks","interactome","interactome2","interactome3",
"interactome4","interactome5","interactome6","subtraction","subtraction-kmt2a",
"localization","cells"] # 297 entries
random.seed(20260919)
for f in folders:
pick = random.sample(sorted(glob.glob(f"data/graph/nodes/{f}/*.json")), 4)
Stratifying by folder rather than sampling uniformly is deliberate: cells holds 41 entries and subtraction holds 14, and a uniform draw would have reported mostly the cell-biology beat. The cost is that the sample is not proportional to the graph, so the percentages describe the beats, not the entries. Both framings are defensible; this one answers 'is this beat importing?' for twelve beats rather than 'what fraction of all entries import?'.
We classified the first evidence item only. Entries here typically lead with their strongest citation and add DepMap, UniProt and bounded-absence items afterwards, so E1 is the load-bearing one - but this is a convention, not a rule, and an entry whose real support is E3 would be misclassified. That is the single biggest limitation of this count 1.
A whole-graph string census, as a cross-check with a different bias
String presence is a crude instrument, but it runs over all 297 entries rather than 48, and it fails in the opposite direction to the hand classification (it over-counts, because a mention in a caveat counts the same as a result). Across the 297:
| token class | entries | % |
|---|---|---|
| mentions AFF1/AF4 anywhere | 248 | 84% |
| mentions mouse/murine | 211 | 71% |
| mentions AF9/MLLT3 | 174 | 59% |
| mentions ENL/MLLT1 | 119 | 40% |
| mentions a t(4;11) line (SEM, RS4;11, MV4-11, KOPN-8, HB11;19) | 71 | 24% |
| mentions an AML line (MOLM-13, THP-1, NOMO-1, OCI-AML, ML-2) | 39 | 13% |
| mentions AF10/MLLT10 | 39 | 13% |
| mentions AF9/ENL/AF10 and never AFF1/AF4 | 28 | 9% |
The two instruments disagree in an informative way. By strings the graph looks AFF1-saturated (84%); by load-bearing evidence it is 54% somebody else's fusion. The gap is the caveat text: AFF1 appears in most entries as the thing the evidence is not about. 1
Graph-wide (all 848 entries), the t(4;11) cell lines are named in 116 entries and the AML lines in 80, so the cell-line balance is genuinely better than the fusion balance - human index-fusion lines are used widely even where the mechanism came from mouse MLL-AF9.
The import classes, counted
Counting by class rather than by instance, over all 848 entries, by the defining primary citations of each class:
| class | entries citing it | index fusion has the interface? |
|---|---|---|
| murine MLL-AF9 keystone papers (PMIDs 21741597, 20159607, 24813891, 20541477, 19956800) | 46 | n/a - species and lineage import |
| AHD-dependent mechanism (23260655, 25921540, 32954361, 34373735, 34174329, 38747396, 21896721) | 20 | no - AFF1 carries no AHD |
| DOT1L-AF10 OM-LZ (25464900, 29563185, 31527241, 33993518) | 14 | no - AFF1 carries no leucine zipper |
| index-fusion H3K79 evidence (18977325, 28076791, 31243293, 26711339) | 13 | yes - this is the index fusion |
| CRISPR-ChIP (37679565) | 7 | no - mouse MLL-AF9, see ana-crispr-chip-fusion-identity-resolved |
Each class has its own entry: import-class-murine-mll-af9-myeloid, import-class-ahd-dependent-mechanism, import-class-dot1l-af10-om-lz, import-class-human-aml-lines-for-infant-b-all, import-class-non-kmt2a-fusion-systems 1.
Scoring the damage in both directions
The brief asked for an honest split, so here it is.
Load-bearing and unflagged: 1 of 13 (8%). gene-leo1, PMID 20541477, whose E1 caveat discusses complex-level attribution and never says the paper's fusion is MLL-AF9 or that the assay is murine - while gene-rtf1, citing the same paper, says both 2.
Load-bearing and already flagged: 12 of 13 (92%). The counter-argument sweeps of cycles 41-48 did their job at the item level.
Imports with independent index-fusion evidence nobody cited: at least one large one. H3K79 methylation in KMT2A::AFF1 is not an inference from MLL-AF9. Krivtsov et al. 2008 profiled H3K79 in a murine Mll-AF4 B-ALL model and in human MLL-rearranged ALL, and showed DOT1L suppression silences MLL-AF4 target genes 3; Kerry et al. 2017 mapped MLL-AF4 spreading with H3K79me2/3 and DOT1L-inhibitor sensitivity in human cells 4. That paper is cited 9 times in this graph and Krivtsov 2008 is cited once. So the phenotype is index-supported even though the mechanism is AF9-supported, and ana-index-fusion-h3k79-evidence-that-exists records the difference.
An import is not automatically wrong. Partner-shared biology is real: all common KMT2A chimeras co-purify the super elongation complex 5, and endogenous AF10 is required for transformation by MLL-AF9, MLL-AF6 and NUP98-NSD1 - fusions that do not contain AF10 6. The correct posture is not to discount imported evidence but to know which claims survive the partner boundary and which do not. The ones that do not are the ones that name an interface.
What would have caught the CRISPR-ChIP case earlier
Nothing in the existing process, and that is the actionable finding. The cycle-48 entry did everything right: it read the abstract, noticed the fusion was unnamed, recorded that in caveats, and listed it as the hypothesis's top open question.
What it did not do is treat an unnamed system as a lookup rather than a caveat. The answer was one esearch db=gds away. A caveat that says 'the source does not state its system' should trigger a GEO/ArrayExpress check before it is accepted, because a deposited dataset states organism and cell type per sample and is free even when the paper is not.
| script | ad hoc python over data/graph/nodes/*/*.json, seed 20260919 |
|---|---|
| status | run |
| result summary | 13 of 24 fusion-bearing load-bearing evidence items (54%) come from a fusion other than KMT2A::AFF1; 0 of 24 from a dedicated human KMT2A::AFF1 experiment; 12 of 13 non-index imports are already flagged in their own caveats; 1 is not. |
inputs
[ "297 mechanism-bearing entries", "48-entry stratified random sample", "848-entry whole-graph string census" ]
sample
{
"population": 297,
"sampled": 48,
"fraction_pct": 16.2,
"seed": 20260919,
"strategy": "4 per folder across 12 mechanism-bearing folders"
}counts — the numbers behind the chart
{
"fusion_bearing": 24,
"non_fusion_bearing": 24,
"other_kmt2a_partner": 10,
"non_kmt2a_fusion_or_other_disease": 3,
"mixed_panel_including_index": 5,
"fusion_not_named_by_source": 4,
"index_fusion_murine": 1,
"index_disease_partner_mixed": 1,
"dedicated_human_kmt2a_aff1_experiment": 0,
"murine_of_fusion_bearing": 12,
"myeloid_only_of_fusion_bearing": 12,
"non_index_flagged": 12,
"non_index_unflagged": 1
}- A target nominated from a mechanism established only in MLL-AF9 inherits that fusion's interface. If the interface is absent from KMT2A::AFF1 - as the AHD is - a chemical series optimised against it can be perfectly good chemistry against a mechanism the index disease does not run.
- The 92% flagging rate is itself a hazard: an audit that finds most entries already honest invites the conclusion that nothing needs doing, when the actual defect is that per-entry honesty does not aggregate into a visible pattern.
- Repeating every imported mechanism in a t(4;11) line is not free. Infant B-ALL models are fewer, slower and harder to transduce than murine MLL-AF9 myeloid cultures, which is a large part of why the literature looks like this; a demand for index-fusion replication of everything would stall the field rather than correct it.
- Does the 54% hold in the beats this sample did not reach - therapeutics, resistance, genome, atlas? Those were excluded as not mechanism-bearing, but a drug-mechanism entry imports in exactly the same way.
- If the classification were re-run on every evidence item rather than the first, would the import fraction rise or fall? Later evidence items are disproportionately DepMap and UniProt, which are index-inclusive, so the whole-item fraction is probably lower - and would be a different and also useful number.
- Is there any mechanism in this graph that has been shown in KMT2A::AFF1 and NOT in MLL-AF9? A one-directional import pattern and a genuinely shared biology look the same from the AF9 side; they differ in whether anyone ever looked from the AFF1 side first.
- E1moderate confidence 3 of 5: One strong study, or several smaller consistent ones.dataset
In a seed-fixed stratified random sample of 48 of the 297 mechanism-bearing entries in this graph, 24 have load-bearing evidence from a fusion-containing system; of those, 13 (54%) come from a fusion other than KMT2A::AFF1, 1 from murine Mll-Af4, 1 from primary infant KMT2A-rearranged disease, 5 from mixed panels that include index-fusion lines, and 4 from sources that do not name the fusion; 0 rest on a dedicated human KMT2A::AFF1 experiment.
Provenance scan of 297 mechanism-bearing entries in this graph (folders mechanism-core, histone-marks, interactome1-6, subtraction, subtraction-kmt2a, localization, cells), cycle 49 · Oncofusion knowledge graph, partner-provenance beat · this project 2026What this evidence cannot show
Our own count, not peer reviewed, computed on 2026-09-19. Three limitations that cut in different directions. (i) Only the FIRST evidence item was classified, on the convention that this graph leads with its load-bearing citation; an entry whose real support is E3 is misclassified. (ii) The sample is stratified by folder, not proportional to entry count, so the percentages describe the twelve beats and not the graph. (iii) Classification of 'which fusion' was read from the entry's own claim and caveat text plus the source abstracts already fetched by this beat, not from the source full texts; where an entry's caveat was silent we read the abstract, and four cases remain 'not named by the source' because the abstract does not say. n = 48 is small: the 54% figure has a binomial 95% interval of roughly 33-74%.
Evidence that argues against this (1)
- A whole-graph string census over the same 297 entries gives the opposite impression - 84% mention AFF1/AF4 and only 9% mention AF9/ENL/AF10 without ever mentioning AFF1 - so a reader using a different, coarser instrument would conclude this graph is AFF1-saturated. Both numbers are correct; they measure different things, and the string census counts a caveat the same as a result. Oncofusion knowledge graph et al. 2026
- E2moderate confidence 3 of 5: One strong study, or several smaller consistent ones.analysis
Of the 13 sampled entries whose load-bearing evidence comes from a non-index fusion, 12 state the mismatch in the evidence item's own caveats, and one - gene-leo1, citing PMID 20541477 - does not name either the fusion or the species, although a sibling entry citing the same paper does.
Partner-provenance audit: how much of this graph's mechanism comes from a fusion that is not KMT2A::AFF1 · Oncofusion knowledge graph (internal analysis) 2026What this evidence cannot show
Abstract fetched from PubMed on 2026-09-19 and read; full text not read. The abstract states the assays are MLL- and MLL-AF9-mediated transcriptional activation of Hoxa9 and MLL-AF9-mediated immortalisation, and does not mention AFF1 or AF4. Whether a caveat 'flags' an import is a judgement call: we counted an entry as flagged if it names the actual fusion, names the species, or explicitly disclaims transfer to KMT2A::AFF1. A stricter rule requiring all three would score far more entries as partially flagged. **Citation corrected 2026-09-19.** This item carried PMID 20541477 as its source, which is wrong in kind: the claim is about how THIS GRAPH annotates its imports, and that paper is a 2010 study of the PAF complex which says nothing about this project. It now cites the audit that produced the count, with the PAF paper kept as counter-evidence for the one specific entry it concerns. This is the third instance of the same class found in one cycle, alongside `ana-reading-depth` E3 and `fusion-tcf3-hlf` E7, and `scripts/verify-claims.py` ranked this one lowest in the entire graph at 0.013 — the checker is detecting the class reliably even though nothing distinguishes it from 'the abstract cannot contain this detail' without reading.
Evidence that argues against this (2)
- An 8% unflagged rate measured on 13 entries is a very noisy estimate - the 95% interval spans roughly 0.2% to 36% - and it was measured in the folders that have already been through counter-argument sweeps. Folders outside this sample, and entries whose flag sits in the body rather than in caveats, are not covered by this number. Oncofusion knowledge graph et al. 2026
- The single unflagged import this claim counts has since been corrected: `gene-leo1` E1 now states that the assays are MLL-AF9 and murine, and that the KMT2A-N-terminal PAFc contact transfers to KMT2A::AFF1 while the synergy and immortalisation results do not. So the 12-of-13 figure describes the graph as it stood on 2026-09-19 and is already out of date in the graph's favour, which is the usual fate of a self-audit. Cancer Cell 2010
- E3strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.paper
A conditional murine Mll-AF4 model produces B-precursor ALL whose expression profile overlaps human MLL-rearranged ALL; ChIP-chip H3K79 methylation profiles correlate with Mll-AF4-associated expression in murine ALL and in human MLL-rearranged leukaemias, human MLL-rearranged ALLs are distinguishable from other ALLs by their H3K79 profiles, and DOT1L suppression inhibits expression of critical MLL-AF4 target genes.
H3K79 methylation profiles define murine and human MLL-AF4 leukemias. · Krivtsov AV, Feng Z, Lemieux ME, Faber J, Vempati S, Sinha AU, Xia X, Jesneck J, Bracken AP, Silverman LB, Kutok JL, Kung AL, Armstrong SA · Cancer Cell 2008What this evidence cannot show
Abstract fetched from PubMed and read on 2026-09-19; full text not read, every identifier copied from that fetched record. ChIP-chip, not ChIP-seq, and not calibrated - so this establishes profile correlation, not quantitative occupancy. The mouse model is a knock-in chimeric Mll-AF4, not the human t(4;11) product. 'Suppression of DOT1L' in 2008 means RNAi, with the off-target and incomplete-knockdown caveats of that era. This paper is cited exactly once elsewhere in this graph.
Evidence that argues against this (1)
- Correlation of H3K79 profiles with expression is not a recruitment mechanism, and this paper does not identify how MLL-AF4 reaches DOT1L. It is the strongest index-fusion evidence for the phenotype and it is silent on the interface, which is precisely the gap recorded in mechanism-kmt2a-aff1-dot1l-recruitment-route. Biswas D et al. 2011
- E4strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.paper
At a subset of gene targets, MLL-AF4 binding spreads into the gene body with spreading of menin binding, increased transcription, increased H3K79me2/3, disrupted H3K36me3 and unmethylated CpG regions; compared with other H3K79me2/3-marked genes, MLL-AF4 spreading-gene expression is preferentially downregulated by DOT1L inhibitors.
MLL-AF4 Spreading Identifies Binding Sites that Are Distinct from Super-Enhancers and that Govern Sensitivity to DOT1L Inhibition in Leukemia. · Kerry J, Godfrey L, Repapi E, Tapia M, Blackledge NP, Ma H, Ballabio E, O'Byrne S, Ponthan F, Heidenreich O, Roy A, Roberts I, Konopleva M, Klose RJ, Geng H, Milne TA · Cell Rep 2017What this evidence cannot show
Abstract fetched from PubMed and read on 2026-09-19; full text not read, every identifier copied from that fetched record. Human MLL-AF4 cell-line ChIP-seq; the abstract does not name the lines, so the number and identity of the t(4;11) models is not established by this citation alone. 'Spreading' is defined by the authors' own peak-calling convention.
Evidence that argues against this (1)
- This is a correlative occupancy map plus a drug-sensitivity contrast. It shows H3K79me2/3 accumulates where MLL-AF4 spreads; it does not show that MLL-AF4 recruits DOT1L, and an equally good reading is that DOT1L follows elongating polymerase wherever the fusion raises transcription. Godfrey L et al. 2019
- E5established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.paper
AF4 and ENL family proteins form a higher-order complex with P-TEFb (AEP) that is normally recruited to MLL-target chromatin, and MLL oncoproteins fused with AEP components constitutively form MLL/AEP hybrid complexes; MLL-AF6, whose partner is cytoplasmic, forms no such hybrid yet still constitutively recruits AEP by unknown means.
A higher-order complex containing AF4 and ENL family proteins with P-TEFb facilitates oncogenic and physiologic MLL-dependent transcription. · Yokoyama A, Lin M, Naresh A, Kitabayashi I, Cleary ML · Cancer Cell 2010What this evidence cannot show
Abstract fetched from PubMed and read on 2026-09-19; full text not read, every identifier copied from that fetched record. Biochemical purification plus murine myeloid transformation assays; the claim that AEP recruitment is the strategy 'most frequently used' is a statement about the partners tested, not a survey of all KMT2A partners. This is cited here as the strongest case FOR partner-shared biology, so it is worth noting it is also the paper that shows one partner reaches the same endpoint by an unknown different route.
Evidence that argues against this (1)
- Shared recruitment of AEP does not imply shared recruitment of DOT1L: reconstitution shows DOT1 forms a stable complex with AF9 and does not interact with AF9-AF4-P-TEFb complexes, so the AEP-containing and DOT1L-containing assemblies are distinct even among partners that share AEP. Biswas D et al. 2011
- E6strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.paper
Endogenous AF10 regulates progressive H3K79 methylation and HOX gene expression, and AF10 inactivation impairs transformation by MLL-AF9, MLL-AF6 and NUP98-NSD1 - fusions that do not themselves contain AF10.
AF10 regulates progressive H3K79 methylation and HOX gene expression in diverse AML subtypes. · Deshpande AJ, Deshpande A, Sinha AU, Chen L, Chang J, Cihan A, Fazio M, Chen CW, Zhu N, Koche R, Dzhekieva L, Ibáñez G, Dias S, Banka D, Krivtsov A, Luo M, Roeder RG, Bradner JE, Bernt KM, Armstrong SA · Cancer Cell 2014What this evidence cannot show
Abstract fetched from PubMed and read on 2026-09-19; full text not read, every identifier copied from that fetched record. Mouse models throughout, and all three fusions tested are myeloid; none is KMT2A::AFF1. Cited here as the best available argument that a mechanism can cross a partner boundary, which is a general argument and not evidence about the index fusion.
Evidence that argues against this (1)
- This result also destabilises the CRISPR-ChIP partition at its most important boundary: if AF10 is required for transformation by fusions that do not contain it, the 'endogenous' MLLT10-directed unit is not a housekeeping activity separable from the oncogenic one, and the two-unit model is cleaner in a reporter assay than in a leukaemia. Gilan O et al. 2023
answers the audit's question and reports that partner prognosis was the wrong proxy for it
the audit that dispatched this beat
- supports →KMT2A::AFF1 (MLL-AF4), t(4;11)(q21;q23)
quantifies how little of the asserted mechanism was measured in the index fusion
MLL-AF9 is the single largest source of imported mechanism in this graph
the sharpest instance the audit was dispatched to resolve
the DOT1L literature is where the import fraction is highest
the audit that defines and counts the concept
the audit's sharpest mechanistic consequence
the third bucket the audit was asked to score
the audit's correction list
the count this hypothesis rests on
- ← part ofWhat this beat read and did not use: bounded absences, bounded rejections and one bounded unreadable
the beat's negative record
the database-side counterpart of 0 of 24 sampled load-bearing experiments resting on a dedicated human KMT2A::AFF1 experiment