Oncofusion

start / graph / ana-at-124-lines-there-was-a-real-lineage-confound-and-the-effect-survives-it-and-the-missense-null-is-tight

Analysisstrong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.

Redone at 11x the power: at n=124 the truncating-ARID1A group IS materially lineage-concentrated (TVD permutation p = 0.00020, against p = 0.42 at n=11), so this time the lineage test had something real to remove - and the effect survives it at -0.2516. The missense null is tight, CI95 (-0.0242, +0.0473), excluding -0.10. And there are now 4 truncating-ARID1A leukaemia lines, one short of testable

also called cycles 84 and 85 redone on the truncating stratum · this time there was a confound to remove · the missense null is real, not underpowered

Two earlier checks on this finding were correct pieces of work done on a badly undersized sample of eleven cell lines, and a better source of mutation records raises that to a hundred and twenty-four, so both were redone. The tissue-of-origin check now matters in a way it did not before - at the larger size the mutated lines really are concentrated in particular tissues, where at eleven they were not - and the finding survives that correction essentially unchanged. A second check settles a worry I had recorded against my own work: lines carrying only a single-amino-acid change really do show no effect, rather than the data simply being too thin to tell. And the number of blood-cancer lines with a clearly broken copy of the gene has gone from zero to four, which is one short of the minimum needed to test the question in the disease this project is actually about.

The same thing, for a scientist

**Statement.** Cycle 86 powered cycle 81's pre-registered PRIMARY stratum by switching from Cellosaurus free text (17 lines, 7 truncating, 4 with data) to cBioPortal CCLE `mutationType` joined through `DEPMAP_ID` (244 lines, 173 truncating, **124 with data**). **Cycles 84 and 85 both ran on the eleven-line any-mutation proxy** and cycle 86 named redoing them. Done. | | value | |---|---| | truncating vs WT (unconditioned) | **−0.2475**, p = 0.00005, CI95 **(−0.3156, −0.1831)** | | **truncating vs WT, lineage-conditioned** | **−0.2516**, stratified p = **0.00005** | | usable strata | **29** (against 9 at n=11), **2** mutant lines dropped | | missense-only vs WT | **+0.0117**, p = 0.66, CI95 **(−0.0242, +0.0473)**, width 0.0715 | | ARID1A instrument, truncating | **+0.1636**, p = 0.00005, CI95 (0.115, 0.2102) | **Q1 FIRES — and this time the test was informative, which is the finding.** Cycle 85 deflated cycle 84 by showing that at n=11 the mutant group's lineage distribution was **not** unusually concentrated (permutation p = 0.42), so the stratified test had little to remove and "lineage explains none of it" confirmed an absence rather than defeating a threat. **At n=124 the concentration permutation gives p = 0.00020.** The truncating group **is** materially lineage-concentrated, the lineage test therefore had a real confound to remove, and **the effect survives conditioning at −0.2516 against an unconditioned −0.2475.** **So cycle 85's deflation was correct for n=11 and does not transfer to n=124.** Both cycles were right about their own sample; the conclusion they could not reach is now available. **Q2 answers a caveat I wrote against my own cycle-86 result.** That entry's counter-evidence said *"'absent in missense' is really 'not detectably different from zero at n=41'"* and that a graded −0.08 would tell a partial-loss-of-function story instead of a threshold one. **The bootstrap settles it: CI95 (−0.0242, +0.0473), width 0.0715, excluding −0.10.** The missense null is tight, not underpowered, and the dose-response is a **threshold**, not a gradient. **And the asymmetry is instructive.** ARID1A's *own* missense estimate is −0.0783 with CI95 **(−0.1851, +0.0202)** — width 0.2053, **containing −0.10**. So ARID1B's missense null is real while ARID1A's is genuinely uncertain. A missense ARID1A allele may well retain partial function; the data cannot say, and it is the same 41 lines. **Q3: four truncating-ARID1A leukaemia lines, against 55 wild-type.** Every ARID1B entry since cycle 81 has carried the limit *"zero leukaemia lines are in the test."* It is no longer zero — but **4 is below the minimum group size of 5 declared at cycle 81**, so it is reported as a count and **not** as a contrast. The standing limit holds, and it is now quantified: **one line short of testable in the disease this project exists for.**

Go deeper
The criterion-informativeness clause paid off on its first use

Cycle 86 ended by naming the one sentence that would have caught all four of this project's failed pre-registered criteria: what result would make this criterion itself uninformative? This analysis is the first to carry that clause, and it fired immediately.

For Q1 the clause read: "if the mutant group's lineage distribution is not materially different from the pool's, a passing Q1 confirms an absence rather than defeating a threat — so the concentration permutation is computed HERE and reported beside it."

Without that clause this entry would have reported "lineage conditioning: FIRES, p = 5e-05" and left the reader to assume a threat had been defeated. With it, the concentration permutation is in the same table and says p = 0.00020 — which is what licenses the stronger reading. At n=11 the identical verdict would have been near-worthless, and cycle 85 had to spend a whole cycle discovering that after the fact.

The general form: a criterion and the test of that criterion's informativeness belong in the same pre-registration and the same output table. Computing the second one a cycle later is how cycle 84's headline came to be overstated for a cycle.

Facts and figures
n truncating lines available173
n truncating with gene effect124
n missense only with gene effect41
n wild type1040
why that comparison is the findingat n=11 the lineage test had little confound to remove, so cycle 85 correctly deflated cycle 84's headline to 'confirming an absence'. At n=124 the concentration permutation gives p = 0.00020 - there IS a real confound - and the effect survives conditioning at -0.2516 against an unconditioned -0.2475. Cycle 85's deflation was right for its sample and does not transfer.
the criterion informativeness clause paid off on first usecycle 86 named the one sentence that would have caught all four of this project's failed pre-registered criteria - 'what result would make this criterion itself uninformative?'. This analysis is the first to carry it, and without it the entry would have reported 'lineage conditioning FIRES' and left a reader to assume a threat was defeated. A criterion and the test of its own informativeness belong in the same pre-registration and the same output table.
what the asymmetry showsARID1B's missense null is TIGHT (width 0.0715, excludes -0.10) while ARID1A's own is genuinely UNCERTAIN (width 0.2053, contains -0.10) on the same 41 lines. A missense ARID1A allele may retain partial function and the data cannot say.
the standing limit now quantified4 truncating-ARID1A leukaemia lines against 55 wild-type. Not zero any more, but below the minimum group size of 5 declared at cycle 81 - ONE LINE SHORT of testable in the disease this project exists for.
open targets dataVersion26.06
pulled2026-09-23
absence classnever-measured
absence class noteThe measurement that would settle whether ARID1A/ARID1B buffering operates in KMT2A-rearranged leukaemia does not exist in any public dataset: it needs a truncating-ARID1A leukaemia line in a CRISPR screen and there are four, one short of this project's own declared minimum. Not measured-not-analysed - the screens have not been done. The nearest available substitute is a dual-perturbation experiment in a KMT2A-rearranged line, which is bench work.
ARID1B
lineage_conditioned-0.252lineage_conditioned: -0.2516truncating_vs_wt-0.247truncating_vs_wt: -0.2475missense_vs_wt0.012missense_vs_wt: 0.0117
lineage concentration at n124
perm_p0.000perm_p: 0.0002observed_tvd0.498observed_tvd: 0.4979n_perm5,000n_perm: 5000
ARID1B — the numbers behind the chart
{
  "truncating_vs_wt": {
    "diff": -0.2475,
    "perm_p": 0.00005,
    "boot_ci95": [
      -0.3156,
      -0.1831
    ]
  },
  "lineage_conditioned": {
    "diff": -0.2516,
    "strat_perm_p": 0.00005,
    "n_strata_usable": 29,
    "n_mut_dropped": 2
  },
  "missense_vs_wt": {
    "diff": 0.0117,
    "perm_p": 0.66417,
    "boot_ci95": [
      -0.0242,
      0.0473
    ],
    "ci_width": 0.0715,
    "excludes_minus_0_10": true
  },
  "leukaemia_only": {
    "n_truncating": 4,
    "n_wt": 55,
    "verdict": "UNDERPOWERED - count, not contrast"
  }
}
ARID1A instrument check
{
  "truncating_vs_wt": {
    "diff": 0.1636,
    "perm_p": 0.00005,
    "boot_ci95": [
      0.115,
      0.2102
    ]
  },
  "missense_vs_wt": {
    "diff": -0.0783,
    "perm_p": 0.12434,
    "boot_ci95": [
      -0.1851,
      0.0202
    ],
    "ci_width": 0.2053,
    "excludes_minus_0_10": false
  }
}
lineage concentration at n124 — the numbers behind the chart
{
  "observed_tvd": 0.4979,
  "perm_p": 0.0002,
  "n_perm": 5000
}
lineage concentration at n11 cycle 85
{
  "perm_p": 0.42083
}
Still open
  1. Audit cBioPortal's mutationType calls. Everything from cycle 86 onward rests on an instrument used twice and never checked, and a depth-correlated truncating-call bias would reproduce this whole pattern.
  2. What are the four truncating-ARID1A leukaemia lines' ARID1B effects? I deliberately did not compute them so an underpowered contrast could not leak into the narrative - but that also means a contrary signal in those four would be invisible to me. Someone should look, label it underpowered, and report it.
  3. Are truncating ARID1A calls genuinely lineage-clustered in a way missense calls are not? If so, the n=11 and n=124 lineage results are two different tests rather than a small and a large version of one, and the story connecting them is mine rather than the data's.
Every source behind this entry (1)
  1. E1strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.analysis

    On the 124-line truncating-ARID1A stratum from cBioPortal CCLE mutationType calls joined to DepMap Chronos, ARID1B's gene-effect difference against 1,040 wild-type screens is -0.2475 unconditioned (20,000-permutation p = 0.00005, bootstrap CI95 -0.3156 to -0.1831) and -0.2516 when truncating status is permuted only within lineage strata (29 usable strata, 2 mutant lines dropped, stratified p = 0.00005). The truncating group's lineage distribution IS materially concentrated relative to the pool: TVD 0.4979 with permutation p = 0.00020, against p = 0.42083 for the eleven-line any-mutation group at cycle 85. The missense-only estimate is +0.0117 with bootstrap CI95 -0.0242 to +0.0473, width 0.0715, excluding -0.10; ARID1A's own missense estimate is -0.0783 with CI95 -0.1851 to +0.0202, width 0.2053, containing -0.10. Four truncating-ARID1A leukaemia lines exist against 55 wild-type, below the minimum group size of 5 declared at cycle 81.

    What this evidence cannot show

    THE LEUKAEMIA LIMIT STILL STANDS AND IS THE ONE THAT MATTERS FOR THIS PROJECT: 4 truncating-ARID1A leukaemia lines is below the minimum of 5 declared at cycle 81, so buffering remains an INFERENCE, not a measurement, in KMT2A-rearranged cells - now quantified as one line short rather than absent. The whole result also still rests on cBioPortal's mutationType calls, an instrument this project first used in cycle 86 and has NOT audited; a bias in truncating-call detection correlating with sequencing depth, and therefore possibly with screen quality, would produce this pattern. The lineage proxy is a disease STRING, not a curated ontology, so it over-splits - which costs power and cannot manufacture signal, but two strings for one tissue discard usable comparisons. Bootstrap intervals assume the lines are exchangeable within stratum, which the lineage concentration result shows they are not in the unconditioned case. Open Targets dataVersion 26.06; the underlying DepMap release name is not stated by that API.

    Evidence that argues against this (3)
    • The tidy reading - 'cycle 85 was right for n=11 and this is right for n=124' - assumes the two strata are measuring the same thing, and they are not. The eleven-line group was ANY ARID1A mutation; this is TRUNCATING only. So the change in lineage concentration from p = 0.42 to p = 0.00020 could reflect the different SELECTION rather than the different size: truncating ARID1A mutations may genuinely cluster in endometrial, ovarian and gastric lineages in a way that missense calls do not. That would mean the two results are not a small and a large version of one test but two different tests, and the narrative tying them together is my own construction.
    • A tight missense confidence interval is not the same as a demonstrated absence of mechanism. CI95 (-0.0242, +0.0473) excludes a -0.10 effect but comfortably includes zero AND small positive values, and 41 lines classified as missense-only by a single annotation source could include genuinely inactivating missense variants - ARID1A missense in a critical domain can be loss-of-function. The clean threshold story requires that missense-only means functional, and nothing here establishes that; the ARID1A instrument check's own missense CI (width 0.2053, containing -0.10) is evidence that it may not.
    • Four leukaemia lines is close enough to the threshold that the decision not to run the contrast is itself consequential, and a reader should know the number was looked at. 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. Reporting the count while withholding the values is the conservative choice and it is not a neutral one.
Connected to