Oncofusion

start / graph / ana-how-many-laboratories-is-this-graph-standing-on

Analysisunrated confidence 0 of 5: unrated

3,288 papers are 2,025 research groups, the per-entry overcount is 11.7% and barely moved while the graph doubled - and TWO entries now rest on three papers from a single laboratory, where the first run of this analysis found none

also called source independence by lab · paper versus group counting

Counting papers overstates how many separate teams stand behind a claim, and here it overstates it by about a eighth. That fraction has stayed the same as the graph doubled in size.

The same thing, for a scientist

Ledger priority 3. **RE-RUN CYCLE 58 ON A GRAPH NEARLY TWICE THE SIZE, AND THIS ENTRY'S ORIGINAL HEADLINE IS NOW FALSE.** Every independence figure this project reports counts **papers**, and two papers from one laboratory are one group's opinion. Clustering the cited PMIDs by shared last author plus heavy author overlap now gives **2,025 groups from 3,288 papers**, and per entry **10,723 paper-citations collapse to 9,465 group-citations - an 11.7% overcount**. **The durable result is the ratio, not either snapshot.** On 2026-09-20 this was 1,731 papers, 1,041 groups, 5,996 to 5,230 citations, 12.8%. The graph has since almost doubled - 1,111 to 1,939 entries - and the overcount moved only from 12.8% to 11.7%. Source-count inflation in this corpus is **structural at roughly an eighth**, not something that grows with the corpus, and that is more useful than either figure alone. **What is no longer true:** the original title said *'not one entry rests on three papers from a single laboratory'* and the body said 0 of 1,111. **Two of 1,939 now do** - `mechanism-is-tifab-a-direct-relb-target-two-papers-one-system` (3 papers, all Zhao C) and `ana-the-chain-cannot-be-system-audited-from-the-graphs-own-fields` (3 papers, all Armstrong SA). The first names its own problem in its id, which is the honest form. The claim-level failure the ledger predicted is therefore real but vanishingly rare: **0.1% of entries**. **Concentration remains a graph-level property.** Largest groups: Stam RW **45** papers, Marschalek R **40**, Vakoc CR 31, Armstrong SA 31, Bushweller JH 30, Cleary ML 27, Shilatifard A 18. The instrument flags its own proxy failures: Vakoc, Bushweller and Cleary are marked `probable_name_collision` and carry cohesion scores of 0.148-0.195, meaning those labels are almost certainly more than one laboratory each, so those counts are **upper bounds** on any single group's influence.

Go deeper
What the recount changes, and what it does not
measurepapersgroups
distinct sources cited1,7311,041
citation instances across entries5,9965,230

The per-entry overcount is 12.8%. That is the amount by which every independence number in this graph - "83.6% of counter-notes cite a different source", "N distinct PMIDs" - has been flattering itself. It is a real correction and it is smaller than the ledger's wording implied.

And the specific worry does not materialise. Zero entries rest on 3+ papers from one group. The ledger's note said "this graph has already found a beat citing one paper per lab and calling it a programme" - the reverse concern, many papers one lab, is absent at the claim level.

Where the concentration actually sits is the whole graph. The fifteen largest groups:

grouppapersgrouppapers
Marschalek R33Greaves M13
Armstrong SA24Pieters R13
Cleary ML22Menendez P13
Bushweller JH21Shilatifard A12
Vakoc CR18Meshinchi S12
Stam RW14Hess JL, Milne TA, Slany RK11, 11, 10

Marschalek's 33 papers are the recombinome series, which this graph cites for breakpoint distributions, partner frequencies and spliced-fusion rates throughout - and which supplied the single most consequential correction of this project, the 1-in-1,262 AFF1 spliced-fusion figure. One archive, one group, load-bearing across many beats. That is not a defect, but it is a fact the graph should state rather than obscure behind a PMID count.

The collision problem, and a check that failed before one that worked

Surname-plus-initial is not an identifier. "li h" came back as an 11-paper group containing seven chimeric-RNA papers from one laboratory and at least one histone-crotonylation paper from a different Li H entirely.

First attempt - author overlap ("cohesion"), and it does not work. The reasoning was that a real lab shares trainees across papers. Measured as the fraction of paper-pairs sharing two or more authors, it flags Cleary 0.17, Greaves 0.18, Vakoc 0.21 - all single long-career PIs whose trainees turned over completely across decades - and it misses the collision that motivated it, scoring "li h" at 0.31, above the threshold. Cohesion measures career length, not identity.

Second attempt - MeSH topic overlap, which catches it. Mean pairwise Jaccard over MeSH descriptors:

groupcohesionMeSH similarity
Pieters R0.850.216tight and on-topic
Stam RW1.000.202tight and on-topic
Marschalek R0.280.187long career, on-topic
Cleary R / Greaves M / Vakoc CR0.17-0.210.133-0.154long career, on-topic
Shilatifard A0.590.113flagged
li h0.310.114flagged - confirmed collision
Bushweller JH0.250.104flagged

It is better and it is still not clean. It catches "li h", but it also flags Bushweller and Shilatifard, who are single laboratories with genuinely broad output - structural biology across several fusion families, and chromatin biology across several complexes. A low MeSH score is a reason to read the titles, not a verdict, and reading the titles is what actually settled all three.

Facts and figures
remeasured cycle 58
{
  "date": "2026-09-22",
  "n_entries": 1939,
  "n_pmids_resolved": 3288,
  "n_groups": 2025,
  "paper_citations": 10723,
  "group_citations": 9465,
  "overcount_fraction": 0.1173,
  "prior_run_2026_09_20": {
    "n_entries": 1111,
    "n_pmids_resolved": 1731,
    "n_groups": 1041,
    "paper_citations": 5996,
    "group_citations": 5230,
    "overcount_fraction": 0.1278
  },
  "the_ratio_is_the_finding": "the graph grew from 1,111 to 1,939 entries and the overcount moved 12.8% -> 11.7%, so the inflation is structural at about an eighth rather than a growing problem",
  "entries_on_a_single_group": [
    {
      "entry": "mechanism-is-tifab-a-direct-relb-target-two-papers-one-system",
      "n_pmids": 3,
      "groups": [
        "zhao c"
      ],
      "note": "names its own single-system problem in its id, which is the honest form"
    },
    {
      "entry": "ana-the-chain-cannot-be-system-audited-from-the-graphs-own-fields",
      "n_pmids": 3,
      "groups": [
        "armstrong sa"
      ]
    }
  ],
  "largest_groups": {
    "stam rw": 45,
    "marschalek r": 40,
    "vakoc cr": 31,
    "armstrong sa": 31,
    "bushweller jh": 30,
    "cleary ml": 27,
    "shilatifard a": 18,
    "menendez p": 15,
    "meshinchi s": 15
  },
  "the_instrument_flags_its_own_failures": "vakoc cr (cohesion 0.148), bushweller jh (0.195) and cleary ml (0.148) are marked probable_name_collision, and greaves m at 13 papers has cohesion 0.18 with an explicit LOW COHESION warning - so these labels are more than one laboratory each and the per-group counts are UPPER BOUNDS",
  "why_the_stale_figure_survived": "this directory holds BOTH result.json and results.json. run.py writes the SINGULAR; the plural was an orphan from a 2026-09-19 version using a different method (senior-author keying: 688 labs, 1.54 papers per lab, 3,510 evidence items) that nothing updated. Every other analysis here writes the plural, so a reader reaches for it by convention and gets a superseded method's numbers. Fixed cycle 58: run.py now writes both."
}
Still open
  1. Re-cluster using ORCID or PubMed's own author disambiguation rather than surname-plus-initial. That replaces a proxy with an identifier and would settle the three flagged groups.
  2. Recompute the counter-argument independence figure - '83.6% of counter-notes cite a different source' - at group level rather than paper level. That number is quoted in the ledger and has never had this correction applied.
  3. How much of the graph would survive removing the largest group's papers? Marschalek supplies 33, load-bearing for breakpoint and partner-frequency claims across many beats, and nothing measures what depends on them.
Every source behind this entry (4)
  1. E1moderate confidence 3 of 5: One strong study, or several smaller consistent ones.analysis

    The 1,731 distinct PMIDs cited in this graph cluster into 1,041 research groups by shared last author plus heavy author overlap. Across 1,111 entries, 5,996 paper-citations collapse to 5,230 group-citations, a 12.8% overcount in every paper-based independence figure the project reports.

    What this evidence cannot show

    Run 2026-09-20; author lists resolved for 1,731 of 1,733 PMIDs. **Last author is a proxy for laboratory and a crude one.** It splits a two-PI collaboration whose senior slot rotates, merges two people with the same surname and initial, and cannot see a trainee who started an independent lab - the last of which makes this a **lower bound on concentration**, since real clustering by institution and funding would merge more. The 12.8% figure therefore understates the overcount.

    Evidence that argues against this (1)
    • A lower bound cuts against the reassuring reading of this result as much as the worrying one. If trainee-founded labs were merged into their parent groups, both the group count and the 'zero entries on a single group' result would move, and in this field - where several of the largest groups trained each other - that is not a remote possibility. Oncofusion knowledge graph (internal analysis) 2026
  2. E2moderate confidence 3 of 5: One strong study, or several smaller consistent ones.analysis

    No entry in the graph rests on three or more papers from a single research group; the concentration is at the graph level rather than the claim level, with the largest single group (Marschalek) supplying 33 of 1,731 cited papers.

    What this evidence cannot show

    'Rests on' counts PMIDs in evidence claims and counter-notes; an entry can still lean on one group through a single paper cited for several claims, which this measure does not see. The threshold of three is arbitrary, and entries with two papers from one group were not counted.

    Evidence that argues against this (1)
    • The per-entry result is reassuring and the graph-level one is not. Marschalek's recombinome series is cited across many beats for breakpoint distributions and partner frequencies, and it is a **single diagnostic archive with its own referral pattern**. No entry over-leans on it, but the graph does, and a per-entry measure is constructed not to see that. Oncofusion knowledge graph (internal analysis) 2026
  3. E3strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.analysis

    Author-overlap cohesion fails as a test for surname collisions: it flags single long-career laboratories (Cleary 0.17, Greaves 0.18, Vakoc 0.21) and misses the confirmed collision it was built for (li h, 0.31). MeSH topic overlap detects that collision at 0.114 but also flags two single laboratories with broad output.

    What this evidence cannot show

    Both thresholds were set after seeing the distribution, on fifteen groups, which is too few to calibrate anything. The 'confirmed collision' was confirmed by reading eight titles, not by an identifier - **ORCID or PubMed author-disambiguation identifiers would settle it properly and were not used.**

    Evidence that argues against this (1)
    • Neither metric is good enough to run unsupervised, which means the group counts above carry an unknown error from undetected collisions. Since a collision **inflates** group size and therefore **understates** independence, the direction is conservative - but it is unmeasured. Oncofusion knowledge graph (internal analysis) 2026
  4. E_k58_remeasuredestablished confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.database

    Re-run on 2026-09-22 over 1,939 entries, the graph's 3,288 resolved PMIDs cluster into 2,025 research groups, and 10,723 per-entry paper-citations collapse to 9,465 group-citations - an 11.7% overcount. Against the 2026-09-20 run (1,111 entries, 1,731 PMIDs, 1,041 groups, 12.8% overcount) the corpus nearly doubled while the overcount moved barely one point, so source-count inflation here is structural at roughly an eighth. TWO entries now rest on three papers from a single group, where the first run found none: mechanism-is-tifab-a-direct-relb-target-two-papers-one-system (Zhao C) and ana-the-chain-cannot-be-system-audited-from-the-graphs-own-fields (Armstrong SA).

    What this evidence cannot show

    A count over one run of one instrument, which is why it is established rather than strong - it is not a claim about how many laboratories exist in this field. Senior authorship is a PROXY for a laboratory and the script says where it fails: four of the largest labels (vakoc cr, bushweller jh, cleary ml, greaves m) carry cohesion below 0.2 and three are flagged probable_name_collision, so they merge distinct groups sharing a common surname and the per-group paper counts are upper bounds. Merging on shared surnames UNDERSTATES the group count and therefore OVERSTATES the overcount; splitting a genuine collaboration into two labels does the reverse. Only PMIDs are covered - a DOI-only or URL-only citation is outside the denominator entirely.

    Evidence that argues against this (2)
    • An 11.7% overcount is small enough to argue the original ledger priority was not worth the instrument. Counting papers rather than groups inflates source counts by about an eighth, and no claim in this graph turns on a one-eighth difference in how many sources back it. The concentration that does matter - Stam 45 papers, Marschalek 40 - was already visible by counting authors and did not need union-find to find.
    • The stability of the ratio across a doubling may be an artefact of what got added rather than a structural property. Cycles 55-58 added heavily from database records, GEO depositions and this project's own analyses, which carry few or no PMIDs and so enter neither numerator nor denominator. A corpus that grew by reading more papers from the same labs would move this number, and this one did not grow that way.
Connected to