Oncofusion

start / graph / ana-depmap-arid1b-is-not-a-dependency-anywhere-and-men1-outscores-kmt2a-itself

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

DepMap agrees with the mouse genetics: ARID1B is not a FUSION-SELECTIVE dependency (+0.0145, p=0.80) and reads -0.08 in both groups - but CYCLE 81 SHOWED THAT FLATNESS IS PARALOGUE BUFFERING, not dispensability. ARID1A and ARID2 ARE dependencies and are equally so in both groups - so no BAF subunit tested is fusion-selective. And MEN1 (-1.26) is a harder and more selective dependency than KMT2A itself (-0.81)

also called the ARID1B DepMap test · is BAF a KMT2A-selective dependency · MEN1 outscores KMT2A in its own rearranged lines

A 2022 mouse study found that deleting the gene ARID1B makes MLL-AF9 leukaemia worse rather than better, which is the opposite of a drug target, so this analysis asked whether large-scale human gene-knockout screens agree - and they do: switching ARID1B off barely affects any leukaemia cell line, whether or not it carries a KMT2A rearrangement. Two related genes, ARID1A and ARID2, are genuinely needed for these cells to grow, but they are needed just as much in leukaemias without a KMT2A rearrangement, so none of them is a target specific to this disease; the unplanned finding is that MEN1, the protein the current drugs actually block, scores as a harder and more selective requirement than KMT2A itself does.

The same thing, for a scientist

**Statement.** Q119 (cycle 79) found that `gene-arid1b`'s only functional result in the KMT2A family points away from a dependency: per **PMID 35022500** (*"Differential roles of BAF and PBAF subunits, Arid1b and Arid2, in MLL-AF9 leukemogenesis"*, Leukemia 2022, Bluemn T et al. — title and journal fetched from esummary before this analysis was written), deleting *Arid1b* at initiation or maintenance **promotes** MLL-AF9 leukemogenesis. This tests that against human CRISPR data, pre-registered. **DepMap CRISPR (Chronos) via Open Targets, 11 KMT2A-rearranged against 46 other leukaemia screens, 20,000-permutation two-sided p:** | gene | mean rearranged | mean other leukaemia | difference | perm p | |---|---:|---:|---:|---:| | **ARID1B** | **−0.0833** | **−0.0979** | **+0.0145** | 0.799 | | ARID1A | −0.4118 | −0.4277 | +0.0159 | 0.881 | | ARID2 | −0.5631 | −0.5628 | −0.0003 | 0.998 | | SMARCA4 | −0.4469 | −0.5260 | +0.0791 | 0.477 | | **MEN1** | **−1.2629** | −0.3745 | **−0.8885** | **0.00005** | | **KMT2A** | **−0.8113** | −0.2542 | **−0.5571** | **0.00010** | **The controls are decisive, which is what licenses reading the rest.** MEN1 and KMT2A both show strong rearranged-selective dependency at p ≤ 1e-4, so the Cellosaurus line split is working. Pre-registered verdict: **P1 HELD, inside the declared ±0.10 indifference band.** **Three results, in order of how much they matter.** **1. ARID1B is not a dependency anywhere.** −0.08 is close to zero on a scale where −1 is the median common essential. This is not merely "non-selective" — the gene is dispensable in all 57 leukaemia screens. Human CRISPR and the murine genetics agree, and a target-directed proposal for ARID1B in this disease has no support from either. **2. No BAF or PBAF subunit tested is fusion-selective.** ARID1A (−0.41) and ARID2 (−0.56) are real dependencies — the instrument can plainly see BAF dependency, so ARID1B's flatness is not an instrument failure — but both are **equally** required with and without a KMT2A rearrangement (|difference| ≤ 0.016, p ≥ 0.88). SMARCA4 likewise (−0.45 vs −0.53, p = 0.48). **That is a constraint on `hyp-ncbaf-not-canonical-baf-is-the-kmt2a-dependency`:** whatever distinguishes ncBAF from cBAF here, it is not visible as selective essentiality of ARID1A, ARID1B, ARID2 or SMARCA4, and a hypothesis resting on subunit-level selectivity would need a different subunit or a different readout. **3. Unplanned, and the most interesting: MEN1 is a harder and more selective dependency than KMT2A itself.** −1.2629 against KMT2A's −0.8113, and a −0.8885 group difference against KMT2A's −0.5571. **The drug target scores better than the fusion's own 5′ gene in the fusion's own lines.** That is a direct, quantitative argument for the menin-interface hypothesis from a dataset nobody in this corpus had pointed at it, and it was a positive control rather than the question.

Go deeper
Why "no dependency" does not mean "no function" here

The result that ARID1B is dispensable has a specific alternative reading that this analysis cannot exclude, and it is the one the same literature points at.

ARID1A and ARID1B are mutually exclusive paralogous subunits of cBAF. The dual-loss literature Q119 surfaced reports that losing both causes acute bone-marrow failure, and its authors explicitly advise against paralogue-directed therapy. A single-gene CRISPR screen cannot see a dependency that a paralogue buffers: if ARID1A substitutes for ARID1B, knocking out ARID1B alone reads as −0.08 whether or not the complex matters.

So the honest claim is narrow: ARID1B is not a single-gene dependency in any leukaemia line screened, and it is not fusion-selective. Whether cBAF containing ARID1B has a function in KMT2A-rearranged leukaemia is untouched by this. The measurement that would settle it is a synthetic-lethal or dual-perturbation screen, which DepMap's single-gene Chronos data cannot provide.

The murine result is the stronger evidence on direction, because it was a genetic deletion in vivo with a leukaemogenesis read-out, not a proliferation screen in an established line. The value of this analysis is that it fails to contradict it, in the human system, with a working positive control — not that it independently proves the direction.

What the pre-registration bought

Written before the API was called, and it mattered in two ways.

The refutation condition was declared and did not fire. "ARID1B mean ≤ −0.3 in rearranged AND more negative than other leukaemia AND perm p < 0.05" would have put published mouse genetics in direct conflict with human CRISPR — a finding either way, and the reason the analysis was worth running rather than assuming.

The ±0.10 indifference band stopped a false positive in its tracks. ARID1B's difference is +0.0145 and SMARCA4's is +0.0791. Without a band declared in advance, SMARCA4's 0.08 in the "rearranged lines need it less" direction is exactly the kind of number that gets written up as a lead. It is inside the band and p = 0.48, and it is reported as uninformative. This project adopted indifference bands after a cycle-68 verdict flipped across a fixed cutoff on a deviation introduced for an unrelated reason; this is the third beat to use one and the first time one has suppressed a result rather than qualified one.

ARID2 was included for a reason that has nothing to do with statistics. The source paper's title names both Arid1b and Arid2. Testing only ARID1B would have been quoting the half of a paper that suits the prediction — the same shape as three prior errors in this corpus's record.

Facts and figures
depmap releaseNOT STATED BY THE OPEN TARGETS API
pulled2026-09-23
n cellosaurus kmt2a fusion lines with depmap ID16
n rearranged screens used11
n other leukaemia screens used46
n permutations20000
seed20260923
indifference band declared in advance0.1
refutation threshold declared in advance-0.3
verdictP1 HELD, inside the declared indifference band - ARID1B is not selectively required, and the difference is too small to speak to direction
controls validtrue
why controls matter hereMEN1 and KMT2A both show rearranged-selective dependency at p <= 1e-4, so the Cellosaurus line split works. Had they not, no ARID1B number in this table could be read, and the script encodes that as an UNINTERPRETABLE verdict rather than leaving it to judgement.
the unplanned findingMEN1 (-1.2629, diff -0.8885) is a harder AND more selective dependency than KMT2A itself (-0.8113, diff -0.5571) in KMT2A-rearranged lines - the drug target outscores the fusion's own 5' gene
constraint on an existing hypothesishyp-ncbaf-not-canonical-baf-is-the-kmt2a-dependency cannot rest on subunit-level selective essentiality: ARID1A, ARID1B, ARID2 and SMARCA4 are all non-selective between rearranged and other leukaemia lines
instrument failure in building thismy first draft read `depMapEssentiality` as a dict when the API returns a LIST of tissue blocks each carrying its own `screens` array. It crashed rather than returning silently - but a defensive `or {}` in that position would have yielded ZERO screens for every gene and printed a table of blanks that could have been read as 'no dependency anywhere'. Fixed by reading analyses/depmap-haem-lineage-control/run.py, which eleven committed analyses already rely on, instead of guessing the shape.
absence classmeasured-not-analysed
absence class noteThe DepMap Chronos profiles for ARID1A, ARID1B, ARID2 and SMARCA4 have been publicly available and queryable for the whole life of this project, and Q119 named this as 'the most informative next number on that gene' in its own hand-back. Nobody here ran it until cycle 80. Not never-measured (the screens exist and are public), not literature-absence-bounded (no literature search is involved) - the data was measured, released and left unanalysed by this corpus.
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
MEN1-1.263MEN1: -1.2629KMT2A-0.811KMT2A: -0.8113ARID2-0.563ARID2: -0.5631SMARCA4-0.447SMARCA4: -0.4469ARID1A-0.412ARID1A: -0.4118ARID1B-0.083ARID1B: -0.0833
results — the numbers behind the chart
{
  "ARID1B": {
    "rearranged": -0.0833,
    "other": -0.0979,
    "diff": 0.0145,
    "p": 0.79946
  },
  "ARID1A": {
    "rearranged": -0.4118,
    "other": -0.4277,
    "diff": 0.0159,
    "p": 0.88136
  },
  "ARID2": {
    "rearranged": -0.5631,
    "other": -0.5628,
    "diff": -0.0003,
    "p": 0.99755
  },
  "SMARCA4": {
    "rearranged": -0.4469,
    "other": -0.526,
    "diff": 0.0791,
    "p": 0.47738
  },
  "MEN1": {
    "rearranged": -1.2629,
    "other": -0.3745,
    "diff": -0.8885,
    "p": 0.00005
  },
  "KMT2A": {
    "rearranged": -0.8113,
    "other": -0.2542,
    "diff": -0.5571,
    "p": 0.0001
  }
}
five missing lines resolved same cycle
{
  "the_open_question": "5 of the 16 Cellosaurus KMT2A-fusion lines with a DepMap id contributed no usable screen; if the lexical disease filter dropped them, every mean in this table shifts",
  "answer": "THE FILTER DROPPED NOTHING. All five have NO SCREEN AT ALL in the Open Targets DepMap data - no gene-effect record of any kind, disease label None.",
  "the_five": [
    "ACH-002209 ALL-PO",
    "ACH-001106 KOPN-8",
    "ACH-002256 Karpas-45",
    "ACH-002273 ML-2",
    "ACH-000006 Mono-Mac-6"
  ],
  "and_the_eleven_that_remain": "every one carries a disease label containing 'Leukemia' - HB11;19, MOLM-13, MOLM-14, MV4-11, NOMO-1, OCI-AML-2, RS4;11, SEM, SEMK2, SHI-1, THP-1 - so the lexical filter leukemia/leukaemia/lymphoblastic/myeloid kept 11 of 11 and excluded none",
  "consequence": "no figure in this entry moves. The rearranged group is 11 because 11 is all the data there is, not because a filter chose 11.",
  "worth_noting_separately": "ML-2 and KOPN-8 are lines this corpus reasons about - ML-2 is the KMT2A::AFDN line Q118 and Q120 discuss, KOPN-8 a KMT2A::MLLT1 infant line - and NEITHER has a DepMap screen. Any future dependency claim about those two has no CRISPR data behind it.",
  "checked": "2026-09-23, same cycle, by re-querying Cellosaurus and the MEN1 screen list and printing every line with its disease label"
}
CORRECTED CYCLE 81 it was paralogue buffering and this entrys own caveat was right
{
  "what_this_entry_framed": "ARID1B as 'dispensable in all 57 leukaemia screens'",
  "what_was_measured_in_cycle_81": "in ARID1A-mutant lines ARID1B's dependency goes -0.0724 -> -0.3709 (difference -0.2985, 20,000-perm p = 0.00020), while ARID1A's own dependency falls away in those same lines (-0.2354 -> -0.0429, +0.1925, p = 0.0439 - the pre-declared instrument check) and ARID2 (+0.0452, p = 0.60) and SMARCA4 (+0.0313, p = 0.77) do not move at all",
  "so_the_correct_framing": "ARID1B is NOT dispensable. It is PARALOGUE-BUFFERED - dispensable only while ARID1A is intact. -0.3709 is a real dependency, comparable to ARID1A's own -0.2354 in wild-type lines. cBAF matters here; the flat score measured redundancy.",
  "the_caveat_in_this_ENTRY_was_right": "this entry's own counter_evidence said 'on that reading -0.0833 measures redundancy, not dispensability, and says nothing about whether ARID1B-containing cBAF is functionally important'. That is now measured and it is the correct reading. The summary_expert's stronger claim outran the entry's own counter-evidence, which is a failure mode this corpus has recorded before: a warning in a field does not protect the prose.",
  "what_is_NOT_overturned": "the FUSION-SELECTIVITY result, which is what this entry was dispatched to test. ARID1B's KMT2A-rearranged-minus-other difference was +0.0145, p = 0.799, and cycle 81 bears on it not at all. A target-directed proposal for ARID1B in KMT2A-rearranged leukaemia still has no support, because that needs fusion-selectivity and there is none.",
  "and_the_limit_of_the_correction": "cycle 81 is PAN-CANCER. Zero truncating-ARID1A leukaemia lines have DepMap data (0 vs 54), so buffering is established for cBAF across 1,208 screens and is an INFERENCE, not a measurement, in KMT2A-rearranged cells.",
  "see": "ana-arid1bs-flat-dependency-was-paralogue-buffering-and-my-own-caveat-was-right"
}
Still open
  1. CLOSED 2026-09-23, same cycle: the five lines have NO SCREEN AT ALL, not a filter exclusion. All 11 lines with data carry a 'Leukemia' disease label, so the lexical filter kept 11 of 11. No figure moves. Separately worth knowing: ML-2 and KOPN-8 have no DepMap screen.
  2. Does a paralogue-buffering reading survive? The decisive measurement is ARID1B dependency in an ARID1A-mutant or ARID1A-low background. DepMap has ARID1A-mutant lines and their identities are queryable, so this is a within-dataset test that was not run.
  3. MEN1 outscoring KMT2A deserves its own pre-registered analysis rather than a footnote in a control. The confound to design against is that CRISPR KMT2A knockout hits both the wild-type allele and the fusion's 5' portion, so the comparison is not gene-for-gene.
Every source behind this entry (1)
  1. E1strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.analysis

    In DepMap CRISPR (Chronos) data accessed via the Open Targets GraphQL API, ARID1B's mean gene effect is -0.0833 across 11 KMT2A-rearranged leukaemia screens and -0.0979 across 46 other leukaemia screens (difference +0.0145, 20,000-permutation two-sided p = 0.799). ARID1A (-0.41) and ARID2 (-0.56) are substantial dependencies but equally so in both groups; SMARCA4 is -0.4469 vs -0.5260 (p = 0.48). The positive controls behave: MEN1 -1.2629 vs -0.3745 (difference -0.8885, p = 0.00005) and KMT2A -0.8113 vs -0.2542 (difference -0.5571, p = 0.00010).

    What this evidence cannot show

    THE DEPMAP RELEASE VERSION IS NOT STATED BY THE OPEN TARGETS API, so these numbers carry a pull date (2026-09-23) and not a release, and must not be quoted without it - a limitation every one of the eleven other DepMap analyses in this repository shares. Group sizes are 11 and 46, so the rearranged arm is small and a single outlier line moves the mean by roughly 0.09 - the same order as the indifference band. Cellosaurus fusion annotation returned 16 KMT2A-fusion lines with a DepMap id, of which 11 had leukaemia-labelled screens with a non-null gene effect; the other 5 are silently absent and I did not check why. Disease labels were matched on the substrings leukemia/leukaemia/lymphoblastic/myeloid, which is a lexical filter and may miss or mislabel screens. Most importantly: a SINGLE-GENE screen cannot detect a dependency buffered by a paralogue, and ARID1A/ARID1B are mutually exclusive cBAF paralogues, so 'not a single-gene dependency' is the claim and 'no function' is not.

    Evidence that argues against this (3)
    • The strongest counter-reading of the ARID1B result is paralogue buffering, and the same literature supports it: ARID1A/ARID1B dual loss causes acute bone-marrow failure, which means the complex matters even though neither single knockout is selectively lethal. On that reading -0.0833 measures redundancy, not dispensability, and says nothing about whether ARID1B-containing cBAF is functionally important in KMT2A-rearranged leukaemia. This analysis cannot distinguish the two and the dual-perturbation screen that could does not exist in DepMap's single-gene Chronos data.
    • The MEN1-outscores-KMT2A observation is a POSITIVE CONTROL read as a result, which is exactly the move this corpus has criticised elsewhere. It was not pre-registered, the comparison between two different genes' Chronos scores assumes the scale is comparable across genes in a way a within-gene group contrast does not require, and CRISPR knockout of KMT2A removes both the wild-type allele and the 5' portion of the fusion, so a weaker KMT2A score may reflect the fusion's independence from wild-type KMT2A rather than a lesser requirement. It is reported because it is large (-1.26 vs -0.81) and because suppressing an unplanned control result would be worse, not because the design supports it.
    • The published murine result this tests is a single paper, and its own title reports DIFFERENTIAL roles for Arid1b and Arid2 - so the 'BAF subunit loss promotes leukaemogenesis' framing is partner-specific within the complex and should not be generalised to BAF. Notably ARID2 is the STRONGEST dependency in this table (-0.56), which is not what a naive reading of 'BAF subunits are tumour suppressors here' would predict, and that tension is unresolved.
Connected to