Oncofusion

start / graph / ana-literature-recall

Analysisunrated confidence 0 of 5: unrated

We have read a tenth of the recent field, and the gap found a real interactor

also called literature recall · absent laboratories · outward-facing coverage

Every quality check this project had built looked inside the graph. This one looks outward and asks how much of what the field has published since 2020 we have actually read. The first run surfaced a laboratory we had never cited whose work is squarely on our own fusion.

The same thing, for a scientist

Five stated PubMed queries, 2020 onward, pooled to **2,129 distinct papers**, of which the graph cites **208 — 9.8%**. Primary-only recall by query: menin in leukaemia **31.7%**, infant ALL 16.0%, KMT2A-rearranged 15.2%, elongation-complex partners 12.8%, **DOT1L/H3K79 3.7%, since refuted as a query artefact**. **65 senior authors have three or more papers in the pooled window and zero cited anywhere in the graph.** Filtering that list by hand produced one immediate, checkable gap: Rao DS (UCLA), six papers, none cited, including IGF2BP3 as a required amplifier of MLL-Af4 leukaemogenesis — a gene the 100-entry interactome beat had no record of.

Go deeper
Why this measurement did not exist

Counter-argument coverage counts how many of our own claims carry opposition and saturated at 100%. The load-bearing audit counts how concentrated our sources are. The lab-level audit counts how many independent groups stand behind them. None of the three can see a paper we never cited, and the PI's standing instruction is to saturate the biology — which is a claim about the outside.

The measure is relative to five queries someone chose, and they are written into the script rather than hidden in a parameter so that a reader who disagrees with the boundary can change it and re-run.

Recall by query
queryin windowretrievedcitedrecallprimary only
menin in leukaemia3233236419.8%31.7%
infant acute lymphoblastic leukaemia2902904013.8%16.0%
kmt2a-rearranged leukaemia1,2601,20016914.1%15.2%
elongation-complex fusion partners5215216412.3%12.8%
DOT1L and H3K79452452153.3%3.7%

Menin at 31.7% is the beat we read best, which is what you would expect of a project organised around the menin-interface hypothesis. DOT1L at 3.7% is the outlier and it is only partly a query artefact. DOT1L is studied across many diseases, so that query catches work that has nothing to do with leukaemia; but 13 of 354 primary papers is low for a node this graph treats as central, and it fits a pattern the project already knows — DOT1L's clinical answer came in negative, and the beat stopped reading.

The output that is actually actionable

65 senior authors have three or more papers in the pooled window and nothing cited anywhere in the graph, in support or in opposition. Most are artefacts of queries broad enough to catch renal cell carcinoma, non-small-cell lung cancer and a herbal formula, which is the caveat demonstrating itself. Hand filtering leaves about twenty that look on-topic, and four are worth naming:

senior authorpapersa title
Rao DS6The RNA-binding protein IGF2BP3 is critical for MLL-AF4-mediated leukemogenesis
Gao S7WT1 facilitates the self-renewal of leukemia-initiating cells
Vogel T4Nucleophosmin 1 cooperates with DOT1L to preserve peri-nucleolar heterochromatin
Hu Z4Modulators of histone demethylase JMJD1C selectively target leukemic stem cells

Rao was acted on the same cycle and is now gene-igf2bp3. The interactome beat holds over 100 entries and had no IGF2BP-family record, which is the instrument working and also a comment on how that beat was built: enumeration from complex membership finds proteins that bind the fusion, and misses a post-transcriptional regulator that never touches it.

A second list is worth as much: laboratories with six or more papers in the window and at most one cited. Kotecha RS (8 papers, 1 cited) is infant ALL and Interfant; Schneidawind C (6, 1) builds CRISPR KMT2A models. Those are groups the graph has met once and not read.

The DOT1L headline was wrong and this analysis predicted it

This entry reported DOT1L/H3K79 as the worst-covered query at 3.7% of primary papers and recorded, in its own counter-evidence and open questions, that the query was the most contaminated of the five and should be narrowed before the number was acted on.

It was narrowed the next cycle, over the same window with the same definition of READ. The reimplementation reproduces 3.7% exactly on the original query, which is the check that matters. Restricted to leukaemia terms it is 16.4%; restricted to KMT2A/MLL specifically, 19.6%.

tierprimarycitedrecall
original, unrestricted354133.7%
alias-expanded, unrestricted404133.2%
+ leukaemia words671116.4%
+ KMT2A/MLL561119.6%

At 19.6% DOT1L sits above the KMT2A-rearranged query (15.2%) and the elongation-partner query (12.8%). It is mid-field, not an outlier, and the story this entry told around it — a beat that stopped reading after the clinical answer came in — was built on a five-fold artefact.

Two residuals survive and are worth more than the headline was. The alias direction is a real gap: expanding to DOT1, KMT4, KIAA1814, "H3 K79" and "lysine 79" adds 54 papers to the unrestricted query and not one of them is cited. And 11 papers is a thin base for a node this graph treats as central.

Second correction: the pooled figure too

The DOT1L correction fixed one number and implied a worse problem with another. Narrowing the remaining four queries the same way shows none of them was contaminated — adding a leukaemia restriction is a no-op for all four, because each already carries a leukaemia term, and DOT1L was the only query written without one.

So the pooled 9.8% published here was being dragged down by a single broken component. Over the four sound queries it is 15.0%, and 15.8% on the KMT2A-restricted slice. Infant ALL is the only one where narrowing changes much (16.6% → 27.0%), and that is real structure rather than contamination: infant ALL includes non-KMT2A disease this graph correctly does not read closely.

This entry has now been corrected twice. The instrument was sound both times and what failed both times was the same thing — publishing an aggregate before checking whether one of its components was broken.

Still open
  1. Recompute the 65-laboratory absent list on the narrowed query set. It was derived from the contaminated pooled set, so some entries are DOT1L-query artefacts and the remainder is a better reading queue.
  2. Do a human pass over the full 65-laboratory absent list and record the rejections, so that 'not cited' and 'considered and rejected' stop being indistinguishable.
  3. Read Kotecha RS and Schneidawind C, the two groups with many papers in window and one citation each; both work directly on infant KMT2A-rearranged disease.
Every source behind this entry (4)
  1. E1established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.analysis

    Against five stated PubMed queries restricted to 2020 onward and pooled to 2,129 distinct papers, the graph cites 208 — 9.8%. Primary-only recall by query: menin in leukaemia 31.7% (46/145), infant ALL 16.0% (28/175), KMT2A-rearranged 15.2% (127/833), elongation-complex partners 12.8% (56/436), DOT1L/H3K79 3.7% (13/354).

    What this evidence cannot show

    Run by the coordinator on 2026-09-20. Five stated PubMed queries, publication date 2020-01-01 to 2026-12-31, up to 1,200 records each sorted by publication date, pooled to 2,129 distinct papers. A paper counts as READ if any entry cites it in support or in counter-evidence. 'Primary' excludes PubMed types Review, Systematic Review, Editorial, Comment, Letter, Case Reports, Erratum and Meta-Analysis.

    Evidence that argues against this (4)
    • Recall against a keyword query is not recall against a field, and a low number is not by itself a defect. Most hits on any of these queries are reviews, case reports, conference abstracts or papers about a different disease that mention a term once, and a knowledge graph should not cite them. The primary-only column removes some of that and not all. Oncofusion knowledge graph (internal analysis) 2026
    • The 2020 window understates coverage by construction. This graph correctly rests on foundational work from the 1990s — the recombinome series, the original menin and DOT1L papers, Rowley's translocation work — none of which is in scope here. The number measures currency, not completeness. Oncofusion knowledge graph (internal analysis) 2026
    • One of the five per-query numbers in this claim has since been refuted by narrowing: DOT1L/H3K79 moves from 3.7% to 19.6%. The pooled 9.8% therefore understates recall on a properly bounded definition of the field, and the other four queries have not been re-measured the same way. Oncofusion knowledge graph (internal analysis) 2026
    • The pooled 9.8% in this claim is superseded. Over the four queries that were not contaminated it is 15.0%, and 15.8% restricted to KMT2A/MLL. The figure stated here is the aggregate of four sound measurements and one broken one. Oncofusion knowledge graph (internal analysis) 2026
  2. E2strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.analysis

    65 senior authors have at least three papers in the pooled 2020-2026 window and zero papers cited anywhere in the graph, in support or in counter-evidence. A second group have six or more papers in the window and at most one cited, including Kotecha RS (8 papers, 1 cited) and Schneidawind C (6, 1).

    What this evidence cannot show

    Run by the coordinator on 2026-09-20. Five stated PubMed queries, publication date 2020-01-01 to 2026-12-31, up to 1,200 records each sorted by publication date, pooled to 2,129 distinct papers. A paper counts as READ if any entry cites it in support or in counter-evidence. 'Primary' excludes PubMed types Review, Systematic Review, Editorial, Comment, Letter, Case Reports, Erratum and Meta-Analysis.

    Evidence that argues against this (2)
    • The precision of this list is unmeasured and visibly poor: the top entries include senior authors whose titles are about renal cell carcinoma, non-small-cell lung cancer and a herbal formula, caught because 'MLL' and 'ALL' are terrible search terms. It is a reading queue requiring a human pass, not a defect list. Oncofusion knowledge graph (internal analysis) 2026
    • Never citing a laboratory is not the same as not having read it. An agent that searched a topic, read the abstracts and judged them off-beat leaves no trace in the graph, and this project's entries record bounded absences but not bounded rejections. Oncofusion knowledge graph (internal analysis) 2026
  3. E3established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.paper

    The first hand-filtered pass of the absent-laboratory list produced a real gap: Rao DS at UCLA, six papers in the window and none cited, including a Leukemia 2022 report that IGF2BP3 strongly amplifies MLL-Af4-mediated leukaemogenesis and that MLL-Af4 leukaemia-initiating cells require it. The Q5 interactome beat held over 100 entries and no IGF2BP-family record.

    PMID 34321607DOI
    What this evidence cannot show

    Run by the coordinator on 2026-09-20. Five stated PubMed queries, publication date 2020-01-01 to 2026-12-31, up to 1,200 records each sorted by publication date, pooled to 2,129 distinct papers. A paper counts as READ if any entry cites it in support or in counter-evidence. 'Primary' excludes PubMed types Review, Systematic Review, Editorial, Comment, Letter, Case Reports, Erratum and Meta-Analysis.

    Evidence that argues against this (2)
    • One recovered gene does not validate an instrument that also returned 60-odd off-topic laboratories. The honest claim is that the list contains real gaps and unmeasured noise, and that its value depends on a human pass that does not scale. Oncofusion knowledge graph (internal analysis) 2026
    • IGF2BP3's own CRISPR data undercut the urgency: DepMap Chronos mean −0.137 with 1.6% of lines depleted, and −0.295 in the four B-lymphoid index lines. Finding a missing paper is not the same as finding a missing dependency. Oncofusion knowledge graph (internal analysis) 2026
  4. E4established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.analysis

    DOT1L/H3K79 was the worst-covered query at 3.7% of primary papers cited (13 of 354). **This is refuted as a query artefact.** Narrowed to leukaemia the recall is 16.4% (11 of 67) and to KMT2A/MLL specifically 19.6% (11 of 56), which places DOT1L above the KMT2A-rearranged query at 15.2% and the elongation-partner query at 12.8%.

    What this evidence cannot show

    Run by the coordinator on 2026-09-20. Five stated PubMed queries, publication date 2020-01-01 to 2026-12-31, up to 1,200 records each sorted by publication date, pooled to 2,129 distinct papers. A paper counts as READ if any entry cites it in support or in counter-evidence. 'Primary' excludes PubMed types Review, Systematic Review, Editorial, Comment, Letter, Case Reports, Erratum and Meta-Analysis.

    Evidence that argues against this (3)
    • The correction is roughly five-fold and it was foreseeable: this item's own counter-evidence called the query the most contaminated of the five and asked for narrowing before the number was acted on. The failure was quoting a headline without the caveat travelling with it, which is the same failure this project logged when a superlative from a 2,018-gene sample was published and fell at 8,248. Oncofusion knowledge graph (internal analysis) 2026
    • This query is the most contaminated of the five. DOT1L and H3K79 are studied in reprogramming, in solid tumours and in basic chromatin biology, so a large fraction of its 354 primary hits is correctly not cited by a leukaemia graph. The 3.7% is a lower bound on the relevant coverage and the query should be narrowed before the number is acted on. Oncofusion knowledge graph (internal analysis) 2026
    • There is also a defensible reason to read DOT1L less: pinometostat reduced H3K79 methylation as intended in 51 adults and produced 2 complete remissions, so the field's clinical answer is in. Under-reading a target whose pharmacological test has been run is a priority judgement, not an oversight — though this graph separately records non-canonical H3K79me2-dependent survival pathways, which is exactly the literature a 3.7% recall would miss. Blood 2018
Connected to