Oncofusion

start / graph / ana-unstated-systems-enumeration

Analysisestablished confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.Q18b-unstated-systems

The unstated-system worklist: 26 sources carry a caveat of this class, 18 now resolve, and 13 of 16 new resolutions are imports

also called Q18b worklist · unstated system enumeration · unstated-system resolution rate

We searched the whole graph for entries that give up on a question by saying the source does not state what system it used, found twenty-six such sources, and went and asked the question. Two-thirds answered - and four out of five of the answers turned out to be a different fusion, species, lineage or tissue from the one this project studies.

The same thing, for a scientist

A regular-expression sweep of all `evidence[].caveats` strings across every node in the graph returned **289 caveat sentences in 231 entries** matching a negation pattern. Filtering to sentences that negate a **system** descriptor - cell line, cell type, fusion, translocation, species, organism, model, lineage - rather than a cohort size, a breakpoint or a reagent left **61 caveats in 57 entries across 46 PMIDs** [E1]. Hand-classification split these into **Class A, 26 sources** where the unstated thing is the experimental system, and **Class B, 20 sources** where it is a cohort count, a breakpoint intron, a locus, a stoichiometry or a reagent detail - a different question that this beat did not pursue and records below for the next one [E1]. Two Class A sources were resolved in cycle 49 (PMID 37679565, the CRISPR-ChIP H3K79 partition, and PMID 23990460). Of the remaining **24 attempted this cycle, 16 resolved and 8 did not: a rate of 66.7%, or 18 of 26 (69.2%) including the prior cycle** [E2]. **Thirteen of the sixteen resolutions are imports** - the system is not human KMT2A::AFF1 B-lymphoid disease. Only three land on the index fusion (PMIDs 28076791, 28690313 and 42418681, all through the SEM line), and in two of those the caveat had been *understating* evidence the graph already had [E2]. **All sixteen changed what a dependent entry can claim**, and the corrections owed to twenty-one entry-and-evidence pairs outside this folder are listed in `data.corrections_owed` [E3].

Go deeper
The worklist and its outcomes

Full machine-readable worklist in data.worklist_resolved and data.worklist_unresolved: entry id, evidence id, PMID, the accession that settled it, the system it resolved to, and whether it is an import. The headline table:

nof Class A attempted
Class A sources (system unstated)26
- resolved in cycle 492
- attempted this cycle24
- resolved this cycle1666.7%
- bounded failures833.3%
Resolutions that are imports1381% of resolutions
Resolutions landing on the index fusion319% of resolutions
Dependent entry-and-evidence pairs needing correction21
Class B sources (different unstated thing)20not pursued

The three index-fusion resolutions all go through SEM. No other human KMT2A::AFF1 system appears anywhere in the resolved set - not RS4;11, not a patient sample, not a PDX. That is a finding about the graph's evidence base as much as about its provenance.

Bounded rejection: what was enumerated and set aside

Following the standing instruction to record what was read and not used:

Class B, 20 sources, deliberately not pursued. These carry a caveat of the same grammatical shape - the abstract does not state X - where X is a cohort size, a breakpoint intron, a miRNA locus, a stoichiometry or a reagent identity. They are listed in data.class_b_deferred. They are a legitimate and separate beat: the repository route is wrong for most of them, because a patient count lives in a table and not in a sample annotation. A future beat should note that three of them (breakpoint introns in case reports) are probably resolvable from the papers' own figures and would repay a reading pass rather than a database pass.

Two Class A sources recorded here rather than in their own entries. PMID 30551418 (noncoding-hoxa-as2-as3#E4) and PMID 37517063 (disease-kmt2a-maml2-thymoma#E5) are bounded failures with full routes recorded in data.worklist_unresolved. Both support a single non-load-bearing evidence item, and 37517063 is a pathology case series with no sequencing to deposit, so no repository route could ever resolve it. Every other bounded failure has its own entry.

228 caveats matched the broad pattern and were excluded on reading. The pattern not stated|not specified|does not name matches a very large amount of ordinary hedging - dose not specified, follow-up not stated, mechanism not specified. The filter that separated signal from noise required a system noun in the same sentence as the negation. The excluded 228 are not recorded individually; the regular expressions used are in data.patterns so the sweep is reproducible.

The spot-check that saved the beat, and what it caught

The standing brief requires spot-checking any computed count against a case whose answer is already known. The known case was PMID 37679565, whose four GEO series were established in cycle 49.

The route the concept entry prescribes - esearch db=gds&term=<PMID>[PMID] - returned zero records for that PMID, and zero for all 26. A clean, round, consistent negative across an entire worklist. Without the spot-check this beat would have reported no deposition exists for any unstated-system source and closed with a 0% resolution rate, and the report would have read entirely plausibly.

elink dbfrom=pubmed&db=gds returns all four series for 37679565 and returned series for eight more PMIDs. The failure is documented in ana-geo-pmid-route-returns-zero-for-everything.

Facts and figures
scriptscratchpad Q18b: scan_caveats.py, filter_system.py, run_elink.py, geo_soft_run.py, pmc_run.py
statusrun
result summary26 Class A sources. 18 of 26 resolved (69.2%); 16 of 24 attempted this cycle (66.7%). 13 of 16 resolutions are imports (81%). 3 land on the index fusion, all via SEM. 21 entry-and-evidence corrections owed outside this folder. 20 Class B sources deferred.
import definitiona resolution is an import if the deposited or recovered system differs from human KMT2A::AFF1 B-lymphoid leukaemia on any of four axes: fusion, species, lineage, tissue
inputs
[
  "data/graph/nodes/*/*.json",
  "NCBI esummary/elink/efetch",
  "GEO acc.cgi text endpoint",
  "Europe PMC REST",
  "EMBL-EBI BioStudies",
  "PRIDE Archive"
]
counts
{
  "caveat_sentences_matched_broad": 289,
  "entries_matched_broad": 231,
  "caveats_system_class": 61,
  "entries_system_class": 57,
  "pmids_system_class": 46,
  "class_a_sources": 26,
  "class_b_sources": 20,
  "resolved_prior_cycle": 2,
  "attempted_this_cycle": 24,
  "resolved_this_cycle": 16,
  "bounded_failures": 8,
  "resolution_rate_this_cycle_pct": 66.7,
  "resolution_rate_overall_pct": 69.2,
  "resolutions_that_are_imports": 13,
  "import_fraction_pct": 81.3,
  "resolutions_on_index_fusion": 3,
  "index_fusion_systems_distinct": 1,
  "corrections_owed": 21,
  "corrections_owed_note": "21 entry-and-evidence pairs where a system claim is wrong or materially incomplete, spread across 17 lines of data.corrections_owed because some lines cover several entries citing the same PMID; two further lines are a confidence change (drug-dot1l808) and a route correction (concept-resolving-an-unstated-system-from-geo), making 19 lines and 23 items in total"
}
patterns
{
  "broad": "does not state|not named in the abstract|system is not stated|does not name the|not specified|unclear which cell|abstract does not|not stated|does not say|never states|is not named|unstated",
  "system_filter": "the sentence must also contain one of: cell line|cell type|fusion|translocation|rearrang|species|organism|mouse|murine|human|model system|experimental system|lineage|genotype|blast|primary cells"
}
worklist resolved
{
  "20541477": [
    "ana-partner-provenance-corrections#E3",
    "GSE21299",
    "mouse, Hoxa9-ER BM (deposition is a different experiment)",
    "import"
  ],
  "21726816": [
    "mark-h2bk34ub#E1",
    "none (full text via NCBI efetch db=pmc)",
    "HeLa, human cervical epithelium, zero leukaemia",
    "import"
  ],
  "23412334": [
    "gene-rnf20#E1; hyp-h2bub-direction-is-unknown-in-kmt2a#E1; mark-h2bk120ub#E3",
    "GSE43725",
    "mouse RN2, MLL-AF9/NrasG12D AML",
    "import"
  ],
  "26766589": [
    "gene-kdm4c#E1; hyp-paralogue-buffering-hides-dependencies#E2",
    "E-MTAB-3322",
    "mouse BM HSPC: MOZ-TIF2 8, MLL-GAS7 6, MLL-AF9 2",
    "import"
  ],
  "27068464": [
    "ana-med23-verdict-kmt2a-b-lineage#E3",
    "GSE74651 +4",
    "mouse RN2, MLL-AF9/NrasG12D, 48/48, no comparator",
    "import"
  ],
  "28076791": [
    "ana-index-fusion-h3k79-evidence-that-exists#E2; mechanism-kmt2a-aff1-dot1l-recruitment-route#E8",
    "GSE83671",
    "human, 30/59 SEM (KMT2A::AFF1 pro-B) + 6 other lines",
    "INDEX"
  ],
  "28690313": [
    "mechanism-panobinostat-depletes-h2bub-mll-r-all#E1",
    "GSE78234",
    "human SEM t(4;11), KOPN8 t(11;19), REH t(12;21); 5 nM 24 h",
    "INDEX"
  ],
  "33268767": [
    "noncoding-hoxa-as2-as3#E3",
    "none (full text via Europe PMC)",
    "NB4 (PML::RARA) 32 mentions, THP-1 28",
    "import"
  ],
  "34373735": [
    "import-class-ahd-dependent-mechanism#E4; risk-partner-import-distorts-target-selection#E4",
    "GSE151660",
    "human MOLM-13, 5 uM; in vivo arm NOT deposited",
    "import"
  ],
  "35618797": [
    "cell-type-osteoblast#E5",
    "GSE189928",
    "Prx1-Cre (limb-bud mesenchymal, NOT osteoblast); mouse MLL::AF9 LSC",
    "import"
  ],
  "36264143": [
    "resistance-mll34-utx-tumour-suppressor-failure#E1",
    "GSE186711",
    "mouse WT-MA9 MLL-AF9 42, MOLM13 6, OCI-AML3 NPM1c 6; 2 AUGMENT-101 patients",
    "import"
  ],
  "38374003": [
    "mechanism-lamp5-as1-lncrna-activates-dot1l#E2",
    "GSE150483",
    "human MOLM-13 only, 8 samples, n=1 per condition",
    "import"
  ],
  "40972012": [
    "mechanism-dot1l-normal-b-lymphopoiesis#E1",
    "GSE306634",
    "Mb1-Cre (pro-B stage); mouse Dot1l f/f; 4 B-cell populations",
    "import"
  ],
  "41634341": [
    "mechanism-dot1l-prc1-antagonism#E1",
    "GSE181829/260456/260742",
    "mouse MLLAF9 + human MOLM13/MV4;11/K562/HUDEP-2/mESC; SGC0946 3-5 uM, VTP-50469 300 nM; NO B-lymphoid",
    "import"
  ],
  "41730833": [
    "cell-type-bone-marrow-adipocyte#E2",
    "none retrievable (PXD066693 404)",
    "mouse MLL-AF9; arms still unseparated",
    "import/partial"
  ],
  "42418681": [
    "gene-usp22#E1",
    "GSE337426",
    "human SEM (ALL) + MV4;11 (AML), genotype field reads KMT2A::AFF1",
    "INDEX"
  ]
}
worklist unresolved
{
  "16239140": [
    "ana-menin-dependence-untested-in-septin-and-cytoskeletal-kmt2a-fusions#E1; exp-menin-inhibitor-sensitivity-of-kmt2a-septin-fusions#E2; hyp-elongation-hijack-is-a-special-case-of-forced-oligomerisation#E5",
    "no PMCID; 0 GEO by elink and author search; 0 SRA/BioSample/BioProject",
    "subscription access to Cell 2005, Methods construct list"
  ],
  "17671196": [
    "ana-menin-dependence-untested-in-septin-and-cytoskeletal-kmt2a-fusions#E2",
    "PMC7566887 exists; EPMC fullTextXML HTTP 500 x2; efetch db=pmc 7,476 chars no <body>; 0 GEO",
    "subscription access to Cancer Res 2007, or a later PMC full-text deposit"
  ],
  "30381403": [
    "gene-ccnd3#E2; target-cdk4-6-inhibition-kmt2a-r#E6",
    "PMC6234363 tagged access: free; EPMC 500; efetch db=pmc 18,700 chars no <body>; 0 GEO; 8 keyword hits all unrelated",
    "fetch from the Blood Advances site directly - it is open access"
  ],
  "30551418": [
    "noncoding-hoxa-as2-as3#E4",
    "no PMCID; 0 GEO/SRA/BioSample/BioProject by elink; esearch db=gds 'HOXA-AS2' returns 24 series, all glioblastoma stem cell or microglia studies",
    "subscription access to Biomed Pharmacother 2019. Peripheral: one evidence item, supporting rather than load-bearing, which is why it is recorded here rather than in its own entry"
  ],
  "34551297": [
    "mechanism-dot1l-oglcnacylation-stability#E1",
    "no PMCID; 0 GEO by elink; keyword hit GSE200059 excluded (Series_pubmed_id 35690146, HEK293T, 17 non-matching contributors)",
    "subscription access to Cell Rep 2021"
  ],
  "37517063": [
    "disease-kmt2a-maml2-thymoma#E5",
    "no PMCID; 0 GEO/SRA/BioSample/BioProject; esearch db=gds 'mucoepidermoid' returns 276 series, none this study",
    "subscription access to Fetal Pediatr Pathol 2023. A pathology case series with no sequencing to deposit, so no repository route could ever resolve it - recorded here rather than in its own entry for that reason"
  ],
  "41685902": [
    "drug-dot1l808#E1",
    "no PMCID; 0 GEO/SRA/BioSample/BioProject; esearch db=gds 'DOT1L AND degrader' returns 4 series (ENL degrader study, mutant NPM1 study), none this paper",
    "subscription access to J Med Chem 2026 - the resistance-model Methods and the supporting information"
  ],
  "42537480": [
    "hyp-evicted-fusion-is-degraded-not-displaced#E_igf2bp3_abundance_confound",
    "no PMCID; 0 GEO/SRA/BioSample/BioProject; esearch db=gds 'IGF2BP3 AND leukemia' returns 43 series, top 8 title-checked, none matching",
    "a library copy of Leuk Res 2026: whether the t(4;11) fusion is endogenous or transfected"
  ]
}
class b deferred
{
  "cohort composition": [
    "38306602 breakpoint-kmt2a-partner-frequency-by-age-lineage#E10",
    "31118806 ana-cell-identity-rejected-candidates#E4",
    "30209392 + 31792411 mechanism-kmt2a-mpal-mixed-phenotype#E3/#E5",
    "40858576 mechanism-lineage-switch-kmt2a#E9",
    "30679323 comut-kras-kmt2a-r#E2",
    "35734412 fusion-kmt2a-usp2#E12"
  ],
  "breakpoint or locus identity": [
    "39717192",
    "33718232",
    "33106467",
    "34354240",
    "23798388",
    "24270738",
    "33405950"
  ],
  "assay or reagent detail": [
    "9989505 interface-menin-jund-pocket#E4",
    "39121437 risk-revumenib-intact-pth-elevation#E2",
    "37019990 gene-wdr5#E3",
    "22547686 gene-aff3#E1",
    "42389523",
    "30089904 risk-fa-signal-cell-line-artefact#E1"
  ]
}
corrections owed
[
  "gene-rnf20#E1 / hyp-h2bub-direction-is-unknown-in-kmt2a#E1 / mark-h2bk120ub#E3 - 'diverse models of MLL-rearranged leukaemia' / 'fusions not specified' -> murine RN2 (MLL-AF9/NrasG12D), 3 samples, GSE43725",
  "ana-med23-verdict-kmt2a-b-lineage#E3 - 'KMT2A rearrangement status not stated' -> 48/48 murine MLL-AF9/NrasG12D, AND no non-KMT2A comparator exists, GSE74651+4",
  "ana-index-fusion-h3k79-evidence-that-exists#E2 - drop 'breadth not bounded by this citation' -> 30 of 59 samples are SEM, GSE83671",
  "mechanism-kmt2a-aff1-dot1l-recruitment-route#E8 - same hedge, same resolution, GSE83671",
  "mechanism-panobinostat-depletes-h2bub-mll-r-all#E1 - 'fusions and lines not named' -> SEM t(4;11), KOPN8 t(11;19), REH t(12;21); 5 nM LBH589 24 h, GSE78234",
  "import-class-ahd-dependent-mechanism#E4 / risk-partner-import-distorts-target-selection#E4 - in vitro arm is MOLM-13 (GSE151660); the in vivo arm is NOT deposited and stays unresolved",
  "cell-type-osteoblast#E5 - LARGEST: 'the osteolineage cre the abstract does not name' -> Prx1-Cre, a limb-bud mesenchymal driver. The negative cannot be read as an osteoblast negative, GSE189928",
  "ana-partner-provenance-corrections#E3 - replace the species inference with the fact: Mus musculus, female C57BL/6; note the deposited series is Hoxa9-ER, not an MLL fusion, GSE21299",
  "mechanism-dot1l-normal-b-lymphopoiesis#E1 - 'Cre driver and stage not stated' -> Mb1-Cre (Cd79a-Cre), acting from the pro-B stage, GSE306634",
  "mark-h2bk34ub#E1 - 'non-leukaemic context' -> HeLa; the full text contains zero occurrences of the stem 'leukem', PMC4119175",
  "gene-kdm4c#E1 / hyp-paralogue-buffering-hides-dependencies#E2 - the deposition is MOZ-TIF2 8, MLL-GAS7 6, MLL-AF9 2, all mouse; the human 'MLL3' xenograft deposited nothing and its partner is unnamed in the full text, E-MTAB-3322",
  "resistance-mll34-utx-tumour-suppressor-failure#E1 - 'KMT2A-r, NPM1-mutant or both' -> BOTH in separate arms: WT-MA9 mouse MLL-AF9 42, MOLM13 6, OCI-AML3 (DNMT3A;NPM1) 6, plus 2 AUGMENT-101 patients one of each, GSE186711",
  "mechanism-lamp5-as1-lncrna-activates-dot1l#E2 - 'MLL fusions not named' -> the H3K79 arm is 8 MOLM-13 samples, one replicate per condition, GSE150483",
  "mechanism-dot1l-prc1-antagonism#E1 - all five clauses close: mouse MLLAF9 + human MOLM13/MV4;11/K562/HUDEP-2/mESC/H9; SGC0946 3-5 uM; VTP-50469 300 nM; and 'conserved across cell types' spans myeloid, erythroid and pluripotent with ZERO B-lymphoid, GSE181829/GSE260456/GSE260742",
  "gene-usp22#E1 - 'a KMT2A-rearranged B-ALL line without specifying it' -> SEM, with an explicit genotype field reading KMT2A::AFF1; paired line MV4;11. The inhibitor clause of the caveat stands, GSE337426",
  "noncoding-hoxa-as2-as3#E3 - 'AML cell lines not stated' -> NB4 (PML::RARA, 32 mentions) and THP-1 (28); only THP-1 is KMT2A-rearranged, PMC7710717",
  "cell-type-bone-marrow-adipocyte#E2 - model is murine MLL-AF9; the arms remain unseparable because the listed GEO accessions are re-analysed public cohorts and PXD066693 returns 404",
  "drug-dot1l808#E1 - lower confidence on the menin-resistance claim to weak: the resistance route is undefined and unrecoverable",
  "concept-resolving-an-unstated-system-from-geo - step (i) of data.route is broken: 'esearch db=gds on the paper's distinctive method name' via term=<PMID>[PMID] returns 0 for all 26 PMIDs including 37679565. Replace with elink dbfrom=pubmed&db=gds, and add the data-availability statement as a co-equal route"
]
spot check
{
  "known_case": "PMID 37679565, four GEO series established in cycle 49",
  "prescribed_route_result": "0 records",
  "working_route_result": "GSE192562, GSE192561, GSE192555, GSE192544",
  "consequence_if_unchecked": "the beat would have reported a 0% resolution rate and no deposition for any source"
}
How this could bite us back
  • An audit that resolves two-thirds of a class and reports a rate invites the reading that one third remains uncertain, when in fact the unmeasured population - unstated systems nobody flagged - is larger than both and entirely unquantified.
  • Every resolution here raises the graph's apparent rigour while lowering the evidential support for its central hypothesis, so a reader who skims the rate and not the imports will draw the opposite conclusion from the correct one.
  • Applying twenty-one corrections will make many entries read as weaker, which creates an institutional incentive to apply the favourable two and defer the rest.
Still open
  1. How many unstated systems exist in entries that never wrote a caveat about it? The 26 counted here are the ones somebody noticed, and the sweep is blind to the rest by construction.
  2. Class B - 20 sources where a cohort count, breakpoint or reagent is unstated - is an untouched beat. How many of those resolve from the papers' own tables rather than from a repository?
  3. Three of the eight bounded failures are 2026 subscription papers whose data may be deposited under embargo. What does this worklist look like re-run in twelve months?
Every source behind this entry (3)
  1. E1established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.database

    A regular-expression sweep over evidence caveats in every node file returned 289 matching sentences in 231 entries; filtering to sentences negating a system descriptor left 61 caveats in 57 entries across 46 PMIDs, which hand-classification split into 26 Class A sources where the experimental system is unstated and 20 Class B sources where a cohort count, breakpoint, locus, stoichiometry or reagent is unstated.

    Q18b caveat sweep over data/graph/nodes/*/*.json in the oncofusion knowledge graph · Oncofusion Knowledge Graph (internal sweep, verified against GEO) 2026
    What this evidence cannot show

    Run on 2026-09-19 over the working tree as it stood, which includes another agent's in-flight files, so the denominator will shift slightly as that work lands. The sweep reads only evidence[].caveats and would miss an unstated system flagged in section prose, a summary or an open question instead; it also cannot see an entry that never noticed the gap, which is the larger and unmeasured population. The Class A / Class B split is a judgement call made by one reader and the boundary cases are listed so it can be re-drawn.

    Evidence that argues against this (1)
    • The sweep measures caveats that were written, not systems that are unstated - an entry whose author never noticed the gap is invisible to it, so 26 is a lower bound on the class and probably a substantial underestimate. Gilan O et al. 2023
  2. E2established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.dataset

    Of 24 Class A sources attempted this cycle, 16 resolved to a named system and 8 did not, a rate of 66.7%; 13 of the 16 resolutions are imports away from human KMT2A::AFF1 B-lymphoid disease and only 3 land on the index fusion, all three through the SEM cell line.

    What this evidence cannot show

    Each resolution is recorded in its own entry with the accession and the samples read; the counts here are a tally of those sixteen entries. 'Resolved' means a named organism and cell system for the deposited data or the recovered full text, not a complete account of the study - three resolutions are explicitly partial (PMIDs 34373735, 20541477 and 41730833). The import classification is this beat's judgement against a four-axis definition (fusion, species, lineage, tissue) stated in data.import_definition.

    Evidence that argues against this (1)
    • A 67% resolution rate on caveats that were WRITTEN says nothing about the rate on unstated systems that nobody flagged, and the flagged ones are plausibly the easier half because an author who noticed the gap was often an author who had already looked. Rogic S et al. 2026
  3. E3strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.dataset

    The sixteen resolutions imply corrections to twenty-one entry-and-evidence pairs outside this folder, ranging from removing an unnecessary hedge on index-fusion evidence to withdrawing an osteolineage attribution that the resolved Cre driver does not support.

    What this evidence cannot show

    Listed in data.corrections_owed and reported rather than applied, per the standing rule that an agent does not edit files it does not own. Each correction names the entry, the evidence id, the sentence at issue and the accession that settles it. The single largest is cell-type-osteoblast#E5, where the driver resolves to Prx1-Cre and the negative result therefore cannot be read as an osteoblast negative at all.

    Evidence that argues against this (1)
    • Two of the twenty-one corrections make a dependent entry stronger rather than weaker, so this is not a defect list; an audit framed purely as error-finding would have missed that the graph was also underselling its own index-fusion evidence. Kerry J et al. 2017
Connected to