Oncofusion

start / graph / ana-the-four-leukaemia-lines-point-the-other-way-and-none-carries-a-kmt2a-fusion

Analysismoderate confidence 3 of 5: One strong study, or several smaller consistent ones.

I withheld these four numbers last cycle to keep an underpowered contrast out of a narrative, and they point the OTHER WAY: the four truncating-ARID1A leukaemia lines mean -0.0401 against a leukaemia wild-type of -0.0923, a naive +0.0523 where pan-cancer gives -0.2475. And NONE of the four carries a KMT2A fusion, so the intersection this whole thread needs is EMPTY, not merely small

also called the four leukaemia lines, looked at · the intersection is empty · what cycle 87 chose not to compute

Last cycle I noted that only four blood-cancer cell lines have a clearly broken copy of the partner gene, decided that was too few to test, and deliberately did not look at their numbers so that a weak comparison could not creep into my write-up - while recording that this also meant any contradictory signal in them would be invisible to me. This cycle I looked. They point the opposite way to the main finding, and more decisively, not one of the four carries the gene rearrangement this project exists to study, so the specific combination needed to test the idea in this disease does not exist in the dataset at all. The main pan-cancer result is unaffected, but carrying it across into this disease now has four data points against it and no route to a direct test.

The same thing, for a scientist

**Statement.** Cycle 87 reported *"4 truncating-ARID1A leukaemia lines against 55 wild-type"*, declined to run the contrast because 4 is below the minimum group size of 5 declared at cycle 81, and wrote in its own counter-evidence: *"I did not compute what those four lines' ARID1B effects are, precisely so that the underpowered contrast could not leak into the narrative — but that also means I cannot rule out that they point the opposite way… Someone should look, label it underpowered, and report it."* Looked. | line | DepMap id | ARID1B effect | disease | |---|---|---:|---| | P31FUJ | ACH-000770 | **−0.1498** | Acute Myeloid Leukemia | | JURKAT | ACH-000995 | **−0.1243** | T-Lymphoblastic Leukemia/Lymphoma | | L1236 | ACH-000702 | **+0.0154** | Hodgkin Lymphoma | | OCI-AML4 | ACH-001613 | **+0.0984** | Acute Myeloid Leukemia | **Mean of the four: −0.0401. Leukaemia wild-type: −0.0923 (n = 59). Naive difference: +0.0523** — the *opposite sign* to the pan-cancer truncating difference of **−0.2475**. **They point the other way.** That is not a contrast and must never be quoted as one: n = 4 against a declared minimum of 5, individual values spanning −0.1498 to +0.0984 on four points, and four different diseases among four lines. **And the decisive fact is not about power at all. NONE OF THE FOUR CARRIES A KMT2A FUSION.** Cross-checked against the 16 Cellosaurus KMT2A-fusion lines with a DepMap id: zero overlap. **So the intersection this entire thread needs — truncating ARID1A **and** KMT2A-rearranged — is EMPTY.** Every ARID1B entry since cycle 81 has carried the limit "buffering is pan-cancer and an inference, not a measurement, in KMT2A-rearranged cells", and cycle 87 upgraded that to "one line short of testable". **That was wrong in the direction that flattered the thread.** It is not one line short. It is unreachable in this dataset: there is no line that is both, and adding a fifth leukaemia line would not change that unless it were also KMT2A-rearranged.

Go deeper
What this does and does not damage

Undamaged: the pan-cancer result. −0.2475 at p = 5e-05 over 124 truncating lines, surviving lineage conditioning at −0.2516, with a tight missense null and a passing instrument check. Four leukaemia lines pointing the other way is entirely compatible with a real effect that is absent or reversed in one tissue — and with n = 4 it is compatible with noise.

Damaged: the bridge. Eight cycles of work have carried a caveat saying the leukaemia relevance is an inference. That inference now has four data points against it and no route to a direct test. A reader deciding whether ARID1B buffering matters for KMT2A-rearranged leukaemia should weight this more heavily than the elegance of the pan-cancer result, because it is the only leukaemia evidence there is.

And the framing "one line short of testable" was itself the error. It implied the question was nearly answerable and invited a future cycle to wait for a fifth line. The real barrier is an empty intersection, which no amount of additional ARID1A-mutant leukaemia lines fixes unless one happens to be KMT2A-rearranged. Naming a gap by the threshold it misses, rather than by the thing that is absent, made it sound closer than it is.

On having withheld the numbers in the first place

Cycle 87's reasoning for not computing these was that an underpowered contrast should not leak into a narrative, and that reasoning is sound. But the same entry recognised the cost and wrote it down, which is the only reason this got looked at one cycle later rather than never.

The lesson is not "do not withhold". It is that a deliberate non-measurement is a debt, and it has to be recorded in a place that is read — cycle 87 put it in an open question and in the counter-evidence, and both were read. Had it been left implicit, the thread would have carried "one line short" indefinitely.

The sharper version: the value of looking was not the four numbers, which are uninterpretable. It was the KMT2A cross-check, which I only thought to run because I was already looking at the four lines by name. The empty intersection was discoverable at any point in the last eight cycles by one Cellosaurus query, and nobody ran it because the limit was always phrased as a count rather than as a question about overlap.

Facts and figures
mean of the four-0.0401
leukaemia wild type mean-0.0923
n leukaemia wild type59
naive difference0.0523
pan cancer truncating difference for comparison-0.2475
directionOPPOSITE in sign to the pan-cancer effect
n kmt2a fusion lines with depmap ID16
overlap with the four0
THE DECISIVE FACTthe intersection of (truncating ARID1A) and (KMT2A-rearranged) is EMPTY in this panel. Not one line short of a threshold - unreachable. Adding a fifth ARID1A-mutant leukaemia line does not help unless it is also KMT2A-rearranged.
what cycle 87 said'4 truncating-ARID1A leukaemia lines ... ONE LINE SHORT of testable in the disease this project exists for'
why that framing was the errornaming a gap by the threshold it misses, rather than by the thing that is absent, made it sound closer than it is and invited a future cycle to wait for a fifth line. The barrier is an empty overlap, not a count.
how this was discoverable eight cycles agoone Cellosaurus query. Nobody ran it because the limit was always phrased as a COUNT of leukaemia lines rather than as a question about OVERLAP with the fusion.
what is undamagedthe pan-cancer result: -0.2475 at p = 5e-05 over 124 lines, surviving lineage conditioning at -0.2516, with a tight missense null and a passing instrument check. Four leukaemia lines pointing the other way is compatible with a real effect absent in one tissue, and at n=4 is compatible with noise.
what is damagedthe bridge to KMT2A-rearranged leukaemia. That inference now has four data points against it and no route to a direct test in this dataset.
open targets dataVersion26.06
pulled2026-09-23
absence classnever-measured
absence class noteThere is no DepMap line that is both truncating-ARID1A and KMT2A-rearranged, so the measurement that would settle whether cBAF paralogue buffering operates in this disease has not been made and cannot be made from this panel. Not measured-not-analysed: the screens do not exist. The nearest substitute is a dual-perturbation experiment in an existing KMT2A-rearranged line, which is bench work, and patient material almost certainly contains the combination even though the cell-line panel does not.
the four lines
P31FUJ-0.150P31FUJ: -0.1498JURKAT-0.124JURKAT: -0.1243L12360.015L1236: 0.0154OCIAML40.098OCIAML4: 0.0984
the four lines — the numbers behind the chart
[
  {
    "line": "P31FUJ",
    "depmap_id": "ACH-000770",
    "arid1b_effect": -0.1498,
    "disease": "Acute Myeloid Leukemia",
    "kmt2a_rearranged": false
  },
  {
    "line": "JURKAT",
    "depmap_id": "ACH-000995",
    "arid1b_effect": -0.1243,
    "disease": "T-Lymphoblastic Leukemia/Lymphoma",
    "kmt2a_rearranged": false
  },
  {
    "line": "L1236",
    "depmap_id": "ACH-000702",
    "arid1b_effect": 0.0154,
    "disease": "Hodgkin Lymphoma",
    "kmt2a_rearranged": false
  },
  {
    "line": "OCIAML4",
    "depmap_id": "ACH-001613",
    "arid1b_effect": 0.0984,
    "disease": "Acute Myeloid Leukemia",
    "kmt2a_rearranged": false
  }
]
Still open
  1. Does ARID1A truncating mutation co-occur with KMT2A rearrangement in PATIENT cohorts? The cell-line panel lacking the combination says little about the disease, and cBioPortal serves paediatric and adult AML/ALL cohorts that could answer it in one query. That is the version of this question worth asking next.
  2. Audit cBioPortal's mutationType calls. Four analyses now rest on an instrument that has never been checked, and the empty intersection depends on it as much as the positive result does.
  3. Does Cellosaurus under-annotate KMT2A fusions? The empty-intersection claim rests on its 16 fusion-positive DepMap lines; a line rearranged but uncurated would be invisible, and DepMap's own fusion calls are a second source that was not consulted.
Every source behind this entry (1)
  1. E1moderate confidence 3 of 5: One strong study, or several smaller consistent ones.analysis

    The four DepMap leukaemia lines carrying a truncating ARID1A call (cBioPortal CCLE 2025 mutationType) have ARID1B gene effects of -0.1498 (P31FUJ, ACH-000770, AML), -0.1243 (JURKAT, ACH-000995, T-ALL), +0.0154 (L1236, ACH-000702, Hodgkin lymphoma) and +0.0984 (OCI-AML4, ACH-001613, AML), mean -0.0401, against a leukaemia wild-type mean of -0.0923 over 59 screens - a naive difference of +0.0523, opposite in sign to the pan-cancer truncating difference of -0.2475. Cross-checked against the 16 Cellosaurus KMT2A-fusion lines carrying a DepMap id, none of the four is KMT2A-rearranged.

    What this evidence cannot show

    THE FOUR-LINE COMPARISON IS NOT A CONTRAST AND MUST NEVER BE QUOTED AS ONE. n = 4 against a minimum group size of 5 declared at cycle 81; individual values span -0.1498 to +0.0984; the four lines are four different diseases (two AML, one T-ALL, one Hodgkin lymphoma); no permutation test was run and none would be legitimate. The +0.0523 naive difference is reported to remove an asymmetry - cycle 87 knew the number existed and chose not to look - NOT as evidence of a reversed effect. The EMPTY-INTERSECTION claim is the load-bearing one here and it rests on Cellosaurus fusion annotation, which returned 16 KMT2A-fusion lines with a DepMap id; if Cellosaurus under-annotates KMT2A fusions in lines it has not curated, a fifth line could exist that is both. Everything also still rests on cBioPortal's mutationType calls, an instrument used four times now and never audited.

    Evidence that argues against this (3)
    • Two of the four lines are arguably not leukaemia in the sense this project means. L1236 is Hodgkin lymphoma, a B-cell malignancy with a very unusual dependency profile, and JURKAT is T-lymphoblastic - neither resembles the myeloid or B-lymphoblastic KMT2A-rearranged context. Dropping them leaves two AML lines at -0.1498 and +0.0984, whose mean is -0.0257 and whose spread is larger than the pan-cancer effect being tested. On that reading the 'they point the other way' framing rests on a disease grouping that is too coarse, and the honest statement shrinks further to 'there is no usable leukaemia evidence in either direction'.
    • The empty-intersection result is less final than the entry's language implies. DepMap screens a small and non-random slice of leukaemia lines; the absence of a truncating-ARID1A KMT2A-rearranged line in 16 fusion-positive screens says more about what has been screened than about what exists in patients. ARID1A truncating mutations are documented in leukaemia at low frequency, so patient material almost certainly contains the combination - it is the CELL LINE PANEL that lacks it, and that distinction is elided by saying the intersection 'is empty'.
    • Reporting the four numbers now, having declined last cycle, is itself a choice that can be read as motivated: the pan-cancer result was strong and the leukaemia numbers are weak, so publishing them after the strong result is established is a less costly order than publishing them first. I did not choose that order deliberately, but I cannot demonstrate that, and a reader is entitled to note that the deflating measurement arrived one cycle after the flattering one.
Connected to