Oncofusion

start / graph / ana-load-bearing-citations

Analysisunrated confidence 0 of 5: unrated

One paper holds up 70 entries, and one dataset holds up 167

also called load-bearing citations · citation concentration audit · single point of failure

We counted what this knowledge graph is actually resting on: one paper supports seventy entries and one dataset supports a hundred and sixty-seven. If either turned out to be misread, a great deal would move at once.

The same thing, for a scientist

3,353 evidence items across **1,399 distinct sources** (1,034 papers), mean 2.4 items per source, and only **9 multi-item entries resting on a single source** — the aggregate is diverse. The exposure is in the tail. **PMID 37019990, the 2023 KMT2A recombinome, supports 75 items in 69 entries across 14 beats**, nearly three times the next paper. **The DepMap Chronos dataset, accessed through one Open Targets endpoint, reaches further still: 219 items in 167 entries (23% of the graph) across 11 beats, and it is the SOLE support for 9 entries.** And **39 of the 60 widest-reach sources are carried at more than one confidence level**; the recombinome is graded `established` 49 times, `strong` 22, `moderate` 3 and `weak` once. With item-level counter-argument coverage now at **3,369/3,369 (100%)**, the coverage number no longer discriminates, so this audit adds its successor: **83.5% of 3,660 counter-notes cite a different source from the claim they oppose; 16.0% cite the same one**, which is the paper's own caveat rather than independent evidence.

Go deeper
The aggregate is healthy and the tail is not

An earlier audit, ana-evidence-redundancy, reports how many distinct papers the graph cites. That is the numerator of the wrong fraction: a graph can cite a thousand papers and still have every load-bearing claim resting on twenty of them. This measures the denominator.

quantityvalue
entries735
evidence items3,353
distinct sources supporting them1,399
of which papers1,034
multi-item entries resting on ONE source9
distinct sources used in counter-evidence1,093
of those, used ONLY as opposition455

That last row is the reassuring one. Four cycles of counter-argument sweeps brought in 455 sources that appear nowhere as support, so the opposition was written by going and finding literature rather than by recycling the citations already in the entry.

Reach: what happens if one paper is wrong

Ranked by number of ENTRIES supported, not by number of citations — a paper cited fifteen times inside one entry is a thorough entry, while a paper cited once each in fifteen entries is a load-bearing beam.

sourceentriesitemsbeats
DepMap Chronos gene effect, via one Open Targets endpoint16721911
The KMT2A recombinome of acute leukemias in 2023707614
NCBI esearch counts for KMT2A/MLL disease terms52549
Outcome of Infants Younger Than 1 Year With ALL242410
MEN1 mutations mediate clinical resistance242410
AUGMENT-101 (revumenib, KMT2A-r)23238
Blinatumomab added to chemotherapy in infant ALL17178
Upfront menin-inhibitor resistance17174

The two widest-reach sources are both instruments rather than experiments. The recombinome is a breakpoint and partner census from 3,401 patients; the DepMap scan is this project's own computation. Neither is a result someone else replicated, and both are used across the whole graph as if they were settled background.

The dependency dataset is the exposure, and this cycle made it worse

167 entries — 23% of the graph — cite DepMap Chronos gene effect through a single Open Targets endpoint, and it is the sole support for 9 of them. That would be fine if what the numbers MEAN were stable. It is not:

  • ana-lineage-confound-myeloid-tf-dependencies showed the index arm is myeloid-enriched, so CEBPA and SPI1 ranked high as lineage artefacts.
  • ana-effect-size-vs-z-reranking showed the selectivity statistic divides by comparator SD, which ranks narrowness rather than effect size at Spearman −0.92 in the depleted tail.
  • Floor compression: a gene depleted in 90–100% of all lines has no dynamic range left, so HCFC1's −0.073 and KAT8's +0.004 are uninterpretable rather than small.

The underlying measurements are not wrong. What changed three times is what a number drawn from them supports, and most of those 167 entries were written before the second and third corrections.

The same paper is graded four different ways

39 of the 60 widest-reach sources carry more than one confidence level. The three worst span three levels.

sourceitemsconfidence levels assigned
The KMT2A recombinome of acute leukemias in 202375established 49, strong 22, moderate 3, weak 1
The landscape of somatic mutations in infant MLL-rearranged leukaemia15strong 11, established 2, moderate 1, weak 1
Human MLL/KMT2A gene exhibits a second breakpoint cluster region15strong 10, moderate 4, speculative 1
DepMap KMT2A-selectivity scan207moderate 108, strong 81, weak 18
REVUFORJ US prescribing information21established 20, moderate 1

Some spread is correct: a breakpoint distribution from 3,401 patients is established, and an inference about what that distribution implies is not. But the sampled items do not sort that way — two entries state the same 98% major-BCR figure, one as established and one as strong. The grading is tracking who wrote the entry.

Concentration by beat

Share of a beat's sourced items coming from its top three sources. Small beats concentrate by arithmetic, so read this against the distinct-source count.

beatsourced itemsdistinct sourcestop-3 share
Q15b Menin in normal tissue642548.4%
Q13 Runnable analyses / coordinator915137.4%
Q16 Scan hits402035.0%
Q14 Phase-2 readiness895032.6%
Q11d KMT2A-side subtraction1023631.4%
…
Q4 Cell type / state2331339.4%
Q11 Comparative fusion biology2401607.9%

Q11 comparative is the best-sourced beat in the graph and Q15b the most concentrated. Q13 being third is expected and still worth saying out loud: the coordinator's own analyses cite the coordinator's own analyses.

Coverage saturated, so here is the measure that replaces it

Item-level counter-argument coverage reached 100% of 3,369 evidence items across all 739 entries this cycle. A metric at its ceiling stops telling you anything, and the obvious successor is whether the opposition is INDEPENDENT of what it opposes.

counter-notesvalue
total3,660
citing a DIFFERENT source from the claim3,056 (83.5%)
citing the SAME source as the claim586 (16.0%)
no resolvable citation18

A note citing the same source is the paper's own caveat. That is legitimate opposition — often the most honest kind, since it quotes the authors against their own headline — and it is not independent evidence. The 586 break down by the claim's citation type as papers 327, databases 108, datasets 68 and this project's own analyses 41; the database and dataset rows are mostly scope arguments against a fact the same database reports, which is the correct form and has no alternative source to cite.

Separately, 464 sources appear in counter-evidence and nowhere as support, which is the strongest single indication that the sweeps went and found literature rather than recycling what the entries already held.

A correction to this analysis, made after it was published

The first version of this audit keyed every citation by PMID, else DOI, else URL. That collapses every internal analysis citing the same API endpoint into one source, and sixteen distinct DepMap analyses in this graph all cite the Open Targets GraphQL URL. The result was a headline reading "the graph's widest-reach source is its own scan, sole support for 16 entries" when what the data said was "sixteen separate analyses each cite an Open Targets query" — a differently shaped and much weaker claim.

Internal analyses are now keyed by TITLE, which is distinct per analysis. Databases keep URL keying, because gnomAD really is one source however many times it is queried. After the fix the widest-reach non-paper source is the DepMap Chronos dataset itself at 167 entries and 219 items, sole support for 9 rather than 16, and the recombinome's reach corrects upward to 70 entries and 76 items.

The finding survives the correction and its description did not. That is worth recording because this audit exists to catch exactly this class of error in other entries.

Still open
  1. Cluster sources by research group before counting distinct sources, so a laboratory's five papers count as one observation rather than five.
  2. Decide a rule for when a census paper is cited as `established` versus `strong`, then apply it to the recombinome's 75 items in one pass.
  3. Re-read the 9 entries for which the DepMap dataset is sole support, against the two corrections the scan has since accumulated.
  4. Read the 327 same-source counter-notes on PAPER claims and separate the ones quoting an author's own caveat from the ones that could have cited independent literature and did not.
Every source behind this entry (6)
  1. E1established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.analysis

    Across 739 entries and 3,369 evidence items the graph draws on 1,431 distinct sources of which 1,046 are papers, and only 8 multi-item entries rest on a single distinct source. Counter-evidence draws on 1,133 distinct sources, 465 of which appear nowhere as support.

    What this evidence cannot show

    Run by the coordinator on 2026-09-20 over all 735 entries. Sources keyed by PMID, else DOI, else NCT, else normalised URL, else normalised title, so a paper reachable under two different URLs counts twice — the revumenib label appears as two keys, an SPL version and an NDA record. Counter-evidence citations are deliberately excluded from every support count, because a paper cited to oppose a claim is not holding it up.

    Evidence that argues against this (3)
    • Distinct-source counts measure breadth and not independence. Papers from one laboratory, or a paper and the review that restates it, count as two sources while carrying one observation, and this project has already found the landscape beat citing one paper per laboratory and calling it a programme. The honest version of this number would cluster sources by research group before counting. Oncofusion knowledge graph (internal analysis) 2026
    • Nine single-source entries is a floor, not a count of fragile entries: the measure only catches entries whose every item cites the SAME source, and an entry citing one paper four times plus one database once escapes it while being just as dependent. Oncofusion knowledge graph (internal analysis) 2026
    • Measured since, and the number is lower than this item states. Clustering every cited PMID by senior author gives 1,049 papers standing on 682 laboratories, and counter-evidence independence falls from 83.7% at SOURCE level to **79.8% at LABORATORY level**: 155 counter-notes oppose a claim using a different paper from the same group. This item's own caveat predicted exactly that, and the prediction was right. Oncofusion knowledge graph (internal analysis) 2026
  2. E2established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.analysis

    The 2023 KMT2A recombinome (PMID 37019990) supports 77 evidence items across 71 entries in 14 of the graph's 28 beats, nearly three times the reach of the next paper, and all 76 citations carry the same PMID and DOI with no identifier drift.

    What this evidence cannot show

    Run by the coordinator on 2026-09-20 over all 735 entries. Sources keyed by PMID, else DOI, else NCT, else normalised URL, else normalised title, so a paper reachable under two different URLs counts twice — the revumenib label appears as two keys, an SPL version and an NDA record. Counter-evidence citations are deliberately excluded from every support count, because a paper cited to oppose a claim is not holding it up. **Citation corrected and figures refreshed 2026-09-19.** This item carried the recombinome paper as its source for a claim about how many of THIS GRAPH's items that paper supports — a fact the paper cannot contain. It now cites the audit that counted it. Counts regenerated after the shallow-reading classifier fix: 77 items across 71 entries, previously 76 across 70. Fourth instance of the wrong-in-kind class found in cycle 50.

    Evidence that argues against this (3)
    • High reach for this paper is largely appropriate and not a defect. It is a census of 3,401 patients, it is the only source for several facts the graph needs repeatedly — partner frequencies, breakpoint distribution, spliced fusions — and there is no competing census to cite alongside it. The risk is concentration, not misuse. Leukemia 2023
    • A census is also the source type whose numbers are most detection-limited: the partners it counts are the partners the contributing laboratories could characterise, so rare partners are undercounted and the concentration of breaks in the major cluster region is an upper bound on itself. Entries quoting its percentages as settled inherit that. Oncofusion knowledge graph (internal analysis) 2026
    • The heuristic that found this does not generalise into a check. Scanning for claims containing self-referential vocabulary ('this graph', 'entries', 'this beat') that cite a paper returns 14 items, of which **4 are the defect and 10 are legitimate** — a claim can cite a paper for a substantive fact and mention the graph in passing, as in 'the KMT2A::USP2 literature contains a structural claim that this graph does not carry'. At 29% precision this belongs in a reading queue and not in a validator, which is the same conclusion the opposed-claims detector reached about its own class. Oncofusion knowledge graph (internal analysis) 2026
  3. E3established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.analysis

    DepMap Chronos gene effect, accessed through a single Open Targets endpoint, is the graph's widest-reach non-paper source: 219 evidence items across 167 entries (23% of the graph) in 11 beats, and the sole support for 9 entries.

    What this evidence cannot show

    Run by the coordinator on 2026-09-20 over all 735 entries. Sources keyed by PMID, else DOI, else NCT, else normalised URL, else normalised title, so a paper reachable under two different URLs counts twice — the revumenib label appears as two keys, an SPL version and an NDA record. Counter-evidence citations are deliberately excluded from every support count, because a paper cited to oppose a claim is not holding it up. Corrected after first publication: the original keying collapsed sixteen distinct internal analyses that share the Open Targets endpoint URL into one source, overstating sole support as 16.

    Evidence that argues against this (2)
    • The instrument's meaning has changed twice since most of those 167 entries were written. The index arm is myeloid-enriched, so myeloid-identity genes rank high as artefacts; and the statistic divides by comparator SD, which ranks narrowness rather than effect size at Spearman −0.92 in the depleted tail. Entries quoting a rank from the scan inherit both corrections without carrying them. Oncofusion knowledge graph (internal analysis) 2026
    • Against that: a self-run scan is the source type most easily re-run, and it has been — four times, at 647, 2,018, 8,248 and 17,728 genes, each time with the candidate ordering stable. A paper misread once stays misread; an instrument in the repository can be pointed at the question again. Oncofusion knowledge graph (internal analysis) 2026
  4. E4strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.analysis

    39 of the 60 widest-reach sources are cited at more than one confidence level. The recombinome spans four levels (established 49, strong 22, moderate 3, weak 1), and two entries state the same 98% major-BCR figure from it at different confidence.

    What this evidence cannot show

    Run by the coordinator on 2026-09-20 over all 735 entries. Sources keyed by PMID, else DOI, else NCT, else normalised URL, else normalised title, so a paper reachable under two different URLs counts twice — the revumenib label appears as two keys, an SPL version and an NDA record. Counter-evidence citations are deliberately excluded from every support count, because a paper cited to oppose a claim is not holding it up.

    Evidence that argues against this (2)
    • Spread is not automatically error, and this measure cannot tell the two apart. A paper reporting both a descriptive census and an interpretive claim SHOULD be cited at different confidences for the two, so the correct number of inconsistently graded sources is somewhere below 39 and this analysis does not say where. Oncofusion knowledge graph (internal analysis) 2026
    • The existing `find-contradictions.py` already flags confidence splits pairwise and has reported nine of them for several cycles without anyone acting, which suggests the reporting was never the bottleneck. Ranking by reach may not change that. Oncofusion knowledge graph (internal analysis) 2026
  5. E5established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.analysis

    Of 3,660 counter-evidence notes in the graph, 3,056 (83.5%) cite a different source from the claim they oppose, 586 (16.0%) cite the same source, and 18 carry no resolvable citation. The same-source notes break down by the claim's citation type as papers 327, databases 108, datasets 68 and internal analyses 41. Separately, 464 distinct sources appear in counter-evidence and nowhere as support.

    What this evidence cannot show

    Measured on 2026-09-20, the cycle in which item-level counter-argument coverage reached 3,369/3,369. Independence here means only that a different identifier was cited; it does not test whether the two sources are independent in substance.

    Evidence that argues against this (3)
    • Citing a different identifier is a weak definition of independence. Two papers from one laboratory, or a paper and the review restating it, count as independent under this measure while carrying one observation — and this project has already found a beat citing one paper per laboratory and calling it a programme. The 83.5% is therefore an upper bound on genuine independence. Oncofusion knowledge graph (internal analysis) 2026
    • Same-source opposition is not a defect and treating 16.0% as a shortfall would push the graph in the wrong direction. A scope argument against a database fact has no alternative source to cite — gnomAD is the only place to say what gnomAD measures — and quoting a paper's own caveat against its headline is the most direct form of opposition available. The right target for this number is not 100%. Oncofusion knowledge graph (internal analysis) 2026
    • Measured since, and the number is lower than this item states. Clustering every cited PMID by senior author gives 1,049 papers standing on 682 laboratories, and counter-evidence independence falls from 83.7% at SOURCE level to **79.8% at LABORATORY level**: 155 counter-notes oppose a claim using a different paper from the same group. This item's own caveat predicted exactly that, and the prediction was right. Oncofusion knowledge graph (internal analysis) 2026
  6. E6established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.analysis

    The first version of this audit keyed sources by PMID, else DOI, else URL, which collapsed sixteen distinct internal DepMap analyses sharing one Open Targets endpoint URL into a single source and overstated its sole support as 16 entries. Keying internal analyses by title corrects sole support to 9 and the recombinome's reach upward to 70 entries and 76 items.

    What this evidence cannot show

    Found by reading the sixteen named entries rather than by any checker, which is the uncomfortable part: no validator in this repository would have caught it.

    Evidence that argues against this (2)
    • Keying by title has its own failure mode in the opposite direction: an analysis renamed between cycles splits into two sources, and a paper cited once under its preprint title and once under its published title already does. Neither keying is correct in general and the audit should report which rule it used, as it now does. Oncofusion knowledge graph (internal analysis) 2026
    • The correction changes the description and not the exposure. 167 entries still rest on one dependency dataset whose interpretation this project has corrected three times, which was the point of the section. Oncofusion knowledge graph (internal analysis) 2026
Connected to