Oncofusion

start / graph / ana-arid1bs-flat-dependency-was-paralogue-buffering-and-my-own-caveat-was-right

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

ARID1B is not dispensable - it is paralogue-buffered. In ARID1A-mutant lines its CRISPR dependency goes from -0.0724 to -0.3709 (p=0.0002) while ARID1A's own falls away (+0.1925, p=0.044) and ARID2/SMARCA4 stay flat. Cycle 80's 'dispensable everywhere' framing was wrong and the caveat that entry carried was right

also called the ARID1B paralogue-buffering test · why a flat CRISPR score is not dispensability · cycle 80's caveat, measured

Last cycle I reported that switching off the gene ARID1B barely affects any leukaemia cell line, and I flagged one alternative explanation I could not rule out: ARID1B has a near-identical partner gene, ARID1A, that may simply cover for it, so a screen that switches off only one would see nothing either way. This is that test, and the alternative explanation wins: in cell lines whose ARID1A is already broken by mutation, switching off ARID1B becomes about five times more damaging, while switching off ARID1A itself stops mattering in those same lines - which is the internal check that the mutation records mean what they say. Two related genes show no such shift, so the effect is specific to the pair that substitute for each other; the honest correction is that ARID1B is not unnecessary, it is covered for.

The same thing, for a scientist

**Statement.** `ana-depmap-arid1b-is-not-a-dependency-anywhere-and-men1-outscores-kmt2a-itself` (cycle 80) reported ARID1B at **−0.0833** in KMT2A-rearranged and **−0.0979** in other leukaemia screens and framed it as *"dispensable in all 57 leukaemia screens"*. It also named, in its own counter-evidence, the one reading it could not exclude: ARID1A and ARID1B are **mutually exclusive paralogous cBAF subunits**, so a single-gene screen cannot see a dependency the paralogue buffers. **Tested within the same dataset. Buffering wins.** | gene | ARID1A-mutant | ARID1A-WT | difference | perm p | |---|---:|---:|---:|---:| | **ARID1B** | **−0.3709** | **−0.0724** | **−0.2985** | **0.00020** | | **ARID1A** *(instrument check)* | −0.0429 | −0.2354 | **+0.1925** | **0.04390** | | ARID2 *(specificity)* | −0.3574 | −0.4027 | +0.0452 | 0.60147 | | SMARCA4 *(specificity)* | −0.3326 | −0.3639 | +0.0313 | 0.76856 | n = 11 ARID1A-mutant against 1,197 wild-type DepMap screens, 20,000-permutation two-sided p, pan-cancer. **The instrument check is what makes this readable and it was declared in advance.** If Cellosaurus's ARID1A mutation calls did not capture functional loss, the split would be noise. They do: **ARID1A's own dependency all but vanishes in the lines called ARID1A-mutant** (−0.2354 → −0.0429, +0.1925, p = 0.044), which is exactly what a gene already inactivated should look like. The script encodes failure of this check as an UNINTERPRETABLE verdict rather than leaving it to judgement. **And the specificity controls are flat.** ARID2 and SMARCA4 are both substantial dependencies that do **not** move with ARID1A status (+0.045 and +0.031, p ≥ 0.60). So this is not "ARID1A-mutant lines are generally sicker and everything looks more essential" — the shift is specific to the paralogue that substitutes for ARID1A. **ARID1B at −0.3709 is a real dependency**, comparable in magnitude to ARID1A's own −0.2354 in wild-type lines. cBAF matters; ARID1B's flat score was redundancy, not irrelevance. **What this does NOT overturn.** Cycle 80's *fusion-selectivity* result stands untouched: ARID1B's KMT2A-rearranged-minus-other difference was **+0.0145, p = 0.799**, and nothing here bears on it. What is corrected is the **framing** — "dispensable everywhere" becomes "dispensable only while ARID1A is intact". A target-directed proposal for ARID1B in KMT2A-rearranged leukaemia still has no support, because that would require fusion-selectivity, which remains absent.

Go deeper
The pre-registered PRIMARY stratum was underpowered, and I did not swap it for the one that worked

This needs stating plainly because the temptation was real and immediate.

The pre-registration named the LOF stratum as primary — only truncating ARID1A calls (nonsense Ter/* or frameshift fs) — on the stated reasoning that missense of unknown consequence dilutes a buffering signal toward zero. That is still the right reasoning.

It holds 4 usable lines and cannot be scored. Of 17 Cellosaurus ARID1A-Mutation lines with a DepMap id, 7 carry a truncating call and 4 of those have gene-effect data — below the minimum group size of 5 declared in advance.

So the primary analysis returned UNDERPOWERED, and that is the primary outcome. The ANY stratum (11 lines, every Mutation call) is secondary, it does give a verdict, and its own instrument check passes. results.json emits both verdicts under separate keys and a how_to_read_this field saying the primary could not be run. The first version of the scoring logic silently reported only the primary's UNSCOREABLE and hid a p = 0.0002 result; the second reports both.

What the underpowered primary costs: the ANY stratum includes missense variants whose functional consequence is unknown, so some of its 11 "mutant" lines may have intact ARID1A. That biases the measured difference toward zero, which means −0.2985 is a floor, not a point estimate — the effect in genuinely ARID1A-null lines is likely larger. It cuts the right way for the conclusion, which is precisely why it must be said out loud rather than left as a reassurance.

Zero leukaemia lines are in this test, and that is the real limit

The LOF-leukaemia-only stratum returned 0 mutant lines against 54 wild-type. ARID1A mutations are rare in leukaemia, and there is not one truncating-ARID1A leukaemia line with DepMap data.

So this analysis is pan-cancer and says nothing directly about KMT2A-rearranged leukaemia. The mechanism it establishes — ARID1A/ARID1B paralogue buffering in cBAF — is a general property of the complex, demonstrated across 1,208 screens. Whether it operates in KMT2A-rearranged cells is an inference from complex biology, not a measurement here.

That inference is worth making anyway, because the murine genetics this whole line of work rests on (PMID 35022500, Arid1b deletion promoting MLL-AF9 leukaemogenesis) is a single-gene perturbation in an ARID1A-intact background. Buffering predicts that the same deletion in an ARID1A-compromised background would look very different, and that is an experiment nobody has run.

The trade was declared in the script's own docstring before the data was pulled — pan-cancer for power, leukaemia-only computed and reported with its n however small — rather than discovered afterwards and presented as a choice.

Facts and figures
depmap releaseNOT STATED BY THE OPEN TARGETS API
pulled2026-09-23
n cellosaurus arid1a mutation lines with depmap ID17
n truncating calls7
n truncating with gene effect data4
min group declared in advance5
indifference band declared in advance0.1
n permutations20000
seed20260923
primary stratum LOF verdictUNDERPOWERED - cannot be scored (n_mutant=4)
secondary stratum ANY verdictP1 FIRES - paralogue buffering explains ARID1B's flatness
instrument check passed ANYtrue
leukaemia only stratum0 ARID1A-truncating lines against 54 wild-type - there is not one truncating-ARID1A leukaemia line with DepMap data, so this analysis is pan-cancer and measures nothing in KMT2A-rearranged cells
what this correctscycle 80's framing of ARID1B as 'dispensable in all 57 leukaemia screens'. It is dispensable ONLY WHILE ARID1A IS INTACT. -0.3709 is a real dependency, comparable to ARID1A's own -0.2354 in wild-type lines.
what this does NOT overturncycle 80's FUSION-SELECTIVITY result. ARID1B's KMT2A-rearranged-minus-other difference was +0.0145, p = 0.799, and nothing here bears on it. A target-directed proposal for ARID1B in this disease would need fusion-selectivity, which remains absent.
why minus 0 2985 is a floorthe ANY stratum includes missense of unknown consequence, so some 'mutant' lines may have intact ARID1A, biasing the measured difference toward zero. The effect in genuinely ARID1A-null lines is likely larger. This cuts in favour of the conclusion, which is why it is stated rather than relied on quietly.
a scoring error i caught in myselfthe first version of the scoring logic evaluated ONLY the pre-declared primary stratum, reported UNSCOREABLE, and silently hid a p = 0.0002 secondary result. The second version emits both verdicts under separate keys plus a how_to_read_this field stating that the primary could not be run. Reporting only the primary would have been defensible; reporting only the secondary would have been goalpost-moving; reporting both is the only honest option and it took a deliberate rewrite to get there.
absence classnever-measured
absence class noteThe decisive experiment - Arid1b deletion in an Arid1a-compromised background in a KMT2A-rearranged model - has never been done by anyone. The murine result this whole line rests on (PMID 35022500) is a single-gene perturbation in an ARID1A-intact background, and buffering predicts it would look different otherwise. Not measured-not-analysed: no dataset anywhere contains it, and DepMap cannot supply it because zero truncating-ARID1A leukaemia lines are screened.
open targets dataVersion26.06, fetched live in cycle 86 via {meta{dataVersion{year month iteration}}}. This is the OPEN TARGETS PLATFORM release and NOT a DepMap release name - the DepMap release itself is genuinely not stated, which is what the 'release not stated' caveats in these entries correctly mean. Recording it because it is the only version anchor available and these entries were omitting it.
results ANY pancancer
ARID1B-0.371ARID1B: -0.3709ARID2-0.357ARID2: -0.3574SMARCA4-0.333SMARCA4: -0.3326ARID1A-0.043ARID1A: -0.0429
results ANY pancancer — the numbers behind the chart
{
  "ARID1B": {
    "mutant": -0.3709,
    "wt": -0.0724,
    "diff": -0.2985,
    "p": 0.0002
  },
  "ARID1A": {
    "mutant": -0.0429,
    "wt": -0.2354,
    "diff": 0.1925,
    "p": 0.0439
  },
  "ARID2": {
    "mutant": -0.3574,
    "wt": -0.4027,
    "diff": 0.0452,
    "p": 0.60147
  },
  "SMARCA4": {
    "mutant": -0.3326,
    "wt": -0.3639,
    "diff": 0.0313,
    "p": 0.76856
  }
}
Still open
  1. Lineage-conditioned version of this contrast. ARID1A-mutant lines are enriched for endometrial, ovarian, gastric and cholangiocarcinoma, and lineage rather than ARID1A status could drive ARID1B dependency. analyses/depmap-lineage-conditioned/ is the template and this was NOT run - it is the single biggest hole in this result.
  2. Does the effect survive dropping each of the 11 mutant lines in turn? At n=11 a leave-one-out is cheap and would say whether -0.2985 rests on one or two lines. Not run.
  3. DepMap publishes its own mutation calls with functional annotation, which would give a cleaner LOF stratum than Cellosaurus free-text parsing for Ter/*/fs. That route was not tried and would likely lift the primary stratum above the minimum group size.
Every source behind this entry (1)
  1. E1moderate confidence 3 of 5: One strong study, or several smaller consistent ones.analysis

    In DepMap CRISPR (Chronos) data via the Open Targets GraphQL API, ARID1B's mean gene effect is -0.3709 across 11 screens of lines carrying a Cellosaurus-annotated ARID1A mutation and -0.0724 across 1,197 other screens (difference -0.2985, 20,000-permutation two-sided p = 0.00020). In the same contrast ARID1A's own dependency falls from -0.2354 to -0.0429 (+0.1925, p = 0.04390), the pre-declared instrument check that the mutation annotation captures functional loss. ARID2 (+0.0452, p = 0.60) and SMARCA4 (+0.0313, p = 0.77) do not move, the pre-declared specificity controls.

    What this evidence cannot show

    CONFIDENCE IS moderate, NOT strong, and four things hold it there. FIRST, THE PRE-REGISTERED PRIMARY STRATUM (truncating ARID1A calls only) IS UNDERPOWERED at 4 usable lines and could not be scored; the result quoted is from the SECONDARY stratum, and both verdicts are emitted separately rather than the secondary being substituted for the primary. SECOND, n = 11 against 1,197 is severely unbalanced and 11 is small - a single outlier line moves the mutant mean by roughly 0.09. THIRD, ZERO LEUKAEMIA LINES carry a truncating ARID1A call with DepMap data (the leukaemia-only stratum is 0 vs 54), so this is a pan-cancer result about cBAF and says nothing measured about KMT2A-rearranged leukaemia. FOURTH, Cellosaurus 'Mutation' is not a curated functional call and the ANY stratum includes missense of unknown consequence, so some 'mutant' lines may have intact ARID1A. The DepMap release version is not stated by the API; figures carry the pull date 2026-09-23.

    Evidence that argues against this (3)
    • The most serious objection is selection on the wrong axis: ARID1A-mutant cancer lines are enriched for particular lineages (endometrial, ovarian, gastric, cholangiocarcinoma), and lineage - not ARID1A status - could drive ARID1B dependency. The flat ARID2 and SMARCA4 controls argue against a general 'these lines are sicker' confound but do NOT exclude a lineage-specific ARID1B requirement that happens to co-occur with ARID1A mutation. A lineage-conditioned version of this contrast is the test, it is the same shape as analyses/depmap-lineage-conditioned/, and it was NOT run.
    • The missense dilution argument cuts both ways and I have stated only the favourable direction in the body. If missense ARID1A variants are functionally neutral they bias the difference toward zero, making -0.2985 a floor. But if the 11-line ANY group is driven by a handful of genuinely null lines and the rest are noise, the effect is concentrated in an even smaller effective n than 11, and the permutation p - which treats all 11 as exchangeable - overstates the evidence. Both readings are consistent with the data and the second is less flattering.
    • This analysis corrects a framing I published one cycle earlier, and a re-measurement run by the person who published the original is a weak instrument for retiring it. The specific risk here is the opposite of the usual one: having written the buffering caveat myself, I had an incentive to find it vindicated. What limits that is that the buffering direction, the indifference band, the instrument check and the specificity controls were all declared in the script's docstring before the API was called, and the primary stratum's failure is reported rather than suppressed.
Connected to