SWI/SNF split by sub-complex and by lineage: every KMT2A-selective subunit is in the ncBAF module and none is in canonical BAF
also called Q47 SWI/SNF DepMap lineage split
We scored 28 SWI/SNF genes for how much KMT2A-rearranged leukaemia cell lines need them, splitting the cell lines into myeloid and B-lineage so that a myeloid gene could not masquerade as a KMT2A gene. Only four genes passed, and three of them belong to the same minority flavour of the machine.
The same thing, for a scientist
`analyses/swi-snf-lineage-split/run.py` scores 28 SWI/SNF genes on DepMap Chronos gene effect via the Open Targets GraphQL API, restricted to leukaemia lines, with the KMT2A-rearranged index arm split into **8 myeloid** and **5 B-lineage** lines and each half scored in standard deviations of a **43-line other-leukaemia** arm. Ensembl ids come from the cached HGNC complete set and the API's `approvedSymbol` is asserted against the requested symbol before a row is kept. Thresholds and code are inherited unchanged from `analyses/depmap-interactome5-lineage-split` so the numbers are comparable across beats [E1]. **Result.** Four panel genes are depleted in both halves (z <= -0.75): **SMARCD1, BRD9, BICRA, SS18** - plus MEN1 and KMT2A as the positive controls, which both pass. Three of the four are ncBAF-defining or ncBAF-shared. **Not one cBAF-defining subunit (ARID1A, ARID1B, DPF1/2/3), cBAF/PBAF core subunit (SMARCA4, SMARCA2, SMARCB1, SMARCC1/2, SMARCE1, ACTL6A) or PBAF-defining subunit (ARID2, PBRM1, PHF10, BRD7) passes** [E2]. Three genes are classified myeloid-identity artefacts: **SMARCD2** (myeloid -1.61, B-lineage -0.04), ACTB and SS18L1 [E2]. The genome-wide scan agrees: BRD9 39, SMARCD1 59, SS18 224, BICRA 517, against SMARCA4 16,333 and ACTL6A 17,410 [E3].
The full table
z = SDs below the 43-line other-leukaemia arm. pan = mean gene effect across all lines returned.
| gene | module | pan | myeloid z | B-lin z | gap | verdict |
|---|---|---|---|---|---|---|
| MEN1 | control | -0.36 | -2.29 | -2.28 | -0.01 | lineage-independent |
| KMT2A | control | -0.21 | -1.48 | -1.22 | -0.26 | lineage-independent |
| DOT1L | control | -0.15 | -0.53 | -0.88 | 0.35 | weak or ambiguous |
| BRD9 | ncBAF | 0.04 | -1.11 | -1.62 | 0.51 | lineage-independent |
| BICRA | ncBAF | 0.09 | -1.14 | -0.94 | -0.20 | lineage-independent |
| SS18 | ncBAF + cBAF | 0.04 | -0.92 | -1.27 | 0.35 | lineage-independent |
| SMARCD1 | shared BAF60 | -0.21 | -0.86 | -1.66 | 0.80 | lineage-independent |
| SMARCD2 | shared BAF60 | -0.16 | -1.61 | -0.04 | -1.57 | MYELOID ARTEFACT |
| SS18L1 | neural BAF | 0.01 | -0.83 | -0.35 | -0.48 | MYELOID ARTEFACT |
| ACTB | core | -0.81 | -1.14 | 0.12 | -1.26 | MYELOID ARTEFACT |
| SMARCA4 | ATPase | -0.36 | 0.57 | -0.23 | 0.80 | none |
| SMARCA2 | ATPase | 0.03 | -0.18 | 0.38 | -0.56 | none |
| SMARCB1 | cBAF/PBAF core | -0.84 | -0.19 | 0.27 | -0.46 | none |
| SMARCC1 | core | -0.23 | -0.19 | 0.06 | -0.25 | none |
| SMARCC2 | core | -0.10 | -0.55 | -0.76 | 0.21 | none |
| SMARCE1 | cBAF/PBAF | -0.67 | -0.42 | 0.45 | -0.87 | none |
| ACTL6A | core | -1.57 | 0.44 | 0.10 | 0.34 | none |
| ARID1A | cBAF | -0.23 | 0.04 | 0.19 | -0.15 | none |
| ARID1B | cBAF | -0.07 | 0.37 | -0.30 | 0.67 | none |
| DPF1 | cBAF | -0.08 | 0.34 | -1.08 | 1.42 | none |
| DPF2 | cBAF | -0.40 | -0.36 | -0.43 | 0.07 | none |
| DPF3 | cBAF | 0.05 | -0.30 | 0.99 | -1.29 | none |
| SMARCD3 | BAF60 | 0.06 | -0.49 | -1.09 | 0.60 | none |
| ARID2 | PBAF | -0.40 | -0.22 | 1.09 | -1.31 | none |
| PBRM1 | PBAF | -0.14 | -0.06 | 0.08 | -0.14 | none |
| PHF10 | PBAF | -0.06 | 0.12 | -0.15 | 0.27 | none |
| BRD7 | PBAF | -0.32 | -0.21 | -0.12 | -0.08 | none |
| BICRAL | ncBAF | -0.01 | 0.27 | -0.41 | 0.67 | none |
The positive controls are what make the panel readable: MEN1 at -2.29/-2.28 and KMT2A at -1.48/-1.22 pass cleanly, so the instrument can detect a real lineage-independent KMT2A dependency 2.
What this cannot show
Five limits, all of which apply to every number above.
- Paralogue buffering. Single-gene CRISPR knockout is blind to a dependency covered by a paralogue, and this complex is full of them: SMARCA4/SMARCA2, ARID1A/ARID1B, SMARCD1/2/3, DPF1/2/3, BICRA/BICRAL, SS18/SS18L1. The canonical-BAF nulls are therefore systematically under-powered, and this is the single most likely reason the cBAF column is empty.
- No infant lines. DepMap contains no infant-derived KMT2A-rearranged B-ALL line. The five B-lineage index lines are adolescent/adult.
- z is not an effect size. BRD9, BICRA and SS18 all have pan-essentiality at or above zero, which is exactly the narrowness that inflates a z. Their raw differences are -0.39, -0.14 and -0.21.
- Three ncBAF genes are not three observations. If the lines need an intact ncBAF, BRD9, BICRA and SS18 measure one thing.
- Lineage by label. The split uses the line's lineage annotation, not an expression signature, and U-937 (PICALM::MLLT10) is treated as a non-KMT2A myeloid reference although it shares elongation biology.
| script | analyses/swi-snf-lineage-split/run.py |
|---|---|
| status | run |
| generated | 2026-09-20 |
| n genes | 28 |
| n kmt2a myeloid lines | 8 |
| n kmt2a b lineage lines | 5 |
| n comparator lines | 43 |
| result summary | Every SWI/SNF gene that survives the lineage split is ncBAF-defining, ncBAF-shared or the shared BAF60 subunit SMARCD1. No cBAF-defining, cBAF/PBAF-core or PBAF-defining subunit survives. SMARCD2, the subunit of the anchor paper, is a myeloid-identity artefact (gap -1.57). |
inputs
[ "DepMap Chronos gene effect via Open Targets GraphQL", "analyses/depmap-null-distribution/hgnc_complete_set.txt" ]
survivors
[ "SMARCD1", "BRD9", "BICRA", "SS18", "MEN1", "KMT2A" ]
lineage artefacts
[ "ACTB", "SMARCD2", "SS18L1" ]
known confounds
[ "paralogue buffering of single-gene knockouts across every slot in this complex", "no infant-derived KMT2A-rearranged B-ALL line exists in DepMap", "z is in comparator-arm SDs and is not an effect size", "three ncBAF genes are one observation if the requirement is complex integrity", "lineage assigned by label, not expression signature" ]
- Acting on this result means prioritising BRD9 and ncBAF, and if the empty cBAF column is a paralogue artefact that prioritisation is wrong in the most expensive possible way - it would redirect a programme away from the module the drugs actually hit.
- None of these numbers is from an infant model, and the disease this graph exists for has no cell line in the dataset.
- The pan-essentiality figures that make BRD9 and BICRA look like a clean therapeutic window are measured across cancer lines, not normal tissue, and say nothing about paediatric marrow or development.
- What does a SMARCA4+SMARCA2 double knockout do in these lines? The single-gene nulls cannot answer the question the panel is really asking.
- Does an expression control change the SMARCD1 result? If SMARCD1 is simply the dominant BAF60 paralogue in B-lineage lines, its rank is an expression readout.
- Would a combinatorial CRISPR screen over the paralogue pairs in this complex recover a canonical-BAF dependency that every single-gene scan misses?
- E1strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.dataset
The analysis scores 28 SWI/SNF genes on DepMap Chronos gene effect via Open Targets GraphQL, leukaemia lines only, with the KMT2A-rearranged arm split into 8 myeloid and 5 B-lineage lines and each half scored in SDs of a 43-line other-leukaemia arm, asserting the API's approvedSymbol against the requested symbol before keeping a row.
SWI/SNF panel lineage-split DepMap dependency scan, 28 genes · Oncofusion knowledge graph, analyses/swi-snf-lineage-split/run.py · DepMap Chronos gene effect via Open Targets GraphQL 2026What this evidence cannot show
Code and thresholds inherited unchanged from analyses/depmap-interactome5-lineage-split; DepMap release version is not stated by the API, so this is not reproducible to a specific release. Ensembl ids come from a cached HGNC file rather than a live fetch.
Evidence that argues against this (1)
- The approvedSymbol assertion protects against a wrong-gene Ensembl resolution but not against a gene being scored in too few lines; the script drops genes with fewer than 3 lines per half, which silently removes any gene whose data are sparse. Oncofusion knowledge graph et al. 2026
- E2moderate confidence 3 of 5: One strong study, or several smaller consistent ones.dataset
Four SWI/SNF panel genes are depleted in both lineage halves - SMARCD1, BRD9, BICRA, SS18 - and none of them is a cBAF-defining, cBAF/PBAF-core or PBAF-defining subunit; SMARCD2, SS18L1 and ACTB are classified myeloid-identity artefacts; MEN1 and KMT2A pass as positive controls.
SWI/SNF panel lineage-split DepMap dependency scan, 28 genes · Oncofusion knowledge graph, analyses/swi-snf-lineage-split/run.py · DepMap Chronos gene effect via Open Targets GraphQL 2026What this evidence cannot show
Eight myeloid and five B-lineage index lines, so a single line moves a mean; thresholds (-0.75 for both halves, -0.35 for the artefact call) are inherited conventions, not derived from a null. Changing the threshold changes the membership of the survivor list.
Evidence that argues against this (1)
- An empty cBAF column is exactly what paralogue buffering would produce, and every cBAF slot in this panel has a paralogue while BRD9 and BICRA's partner BICRAL do not cover each other here. The result is therefore confounded with redundancy structure. Mashtalir N et al. 2018
- E3moderate confidence 3 of 5: One strong study, or several smaller consistent ones.dataset
The genome-wide scan over the same lines agrees: BRD9 ranks 39, SMARCD1 59, SS18 224, BICRA 517, against SMARCA4 16,333 (+0.0791), ARID1A 12,593 (+0.0276), SMARCB1 14,024 (+0.0415) and ACTL6A 17,410 (+0.1454).
Genome-wide KMT2A-rearranged selectivity scan, 17,728 genes (DepMap Chronos via Open Targets) · Oncofusion knowledge graph, analyses/depmap-kmt2a-selective-scan-full/run.py · DepMap Chronos gene effect via Open Targets GraphQL 2026What this evidence cannot show
11 index against 43 comparator lines; Cellosaurus fusion annotation may be incomplete, biasing toward the null. The two scans share the underlying data, so agreement between them is not independent replication.
Evidence that argues against this (1)
- Pharmacological dual ATPase inhibition kills MLL1-rearranged AML cells in vivo, which a rank of 16,333 for SMARCA4 does not predict; a genetic scan and a drug are measuring different things. Fiskus W et al. 2024
- E4moderate confidence 3 of 5: One strong study, or several smaller consistent ones.paper
No published study has separated cBAF from ncBAF in KMT2A-rearranged leukaemia: every MLL-rearranged SWI/SNF functional paper predates the 2018 definition of ncBAF and perturbed shared or cBAF subunits.
Glioma tumor suppressor candidate region gene 1 (GLTSCR1) and its paralog GLTSCR1-like form SWI/SNF chromatin remodeling subcomplexes. · Alpsoy A, Dykhuizen EC · J Biol Chem 2018What this evidence cannot show
Bounded by publication date rather than by a search: the ncBAF-defining papers are 2018 (PMIDs 29374058, 30397315, 30343899) and the MLL-rearranged SWI/SNF functional literature cited in this beat is 2013-2015. A post-2018 complex-resolved leukaemia study could exist and be framed at complex level; `ncBAF AND leukemia` returns 15 records which were not individually title-checked.
Evidence that argues against this (1)
- BRD9's AML dependency was established in 2016, before ncBAF was named, and was already effectively an ncBAF result - so the field had complex-specific data before it had complex-specific language. Hohmann AF et al. 2016
the module-level result the complex distinction is needed to state
BRD9 is the best-ranked survivor, lineage-independent in both halves
SMARCD2's KMT2A dependency does not survive the lineage split (gap -1.57)
no SMARCA4 selectivity in either lineage half
SMARCD1 survives the split with the largest B-lineage z in the panel outside the controls
BICRA survives the split
SS18 survives the split
the dependency half of the same beat's evidence base
the data pattern the hypothesis is read off