Oncofusion

start / graph / ana-q114-suv39h1-is-not-at-the-kmt2a-locus-in-either-patient-derived-glioma-line

Analysismoderate confidence 3 of 5: One strong study, or several smaller consistent ones.Q114-suv39h1-at-the-locus

Read at last, against a pre-registered class-matched null: in SU-DIPG13 and SU-DIPG36 the SUV39H1 CUT&RUN mean at the KMT2A exon-11 CTCF anchor is 0.0000 and 1.6660 - the first an EXPLICIT zero record, not missing data - and no one of the seven KMT2A intervals in either line exceeds 1.21x the genome-wide random-tile mean. The pre-registered NEGATIVE prior held.

also called SUV39H1 at KMT2A in DIPG · Q114 DIPG arm · the writer is not at the anchor

In two tumour lines taken from children with a brainstem glioma, the enzyme that writes the main silencing tag does not sit on the KMT2A gene anywhere along it - and at the exact 147-letter spot this project cares about, one of the two lines has a measured zero. It is a clean negative, and it says nothing about infant leukaemia, because neither cell is a blood cell.

The same thing, for a scientist

**In patient-derived H3K27M diffuse midline glioma neurospheres (SU-DIPG13, SU-DIPG36), SUV39H1 does not occupy the *KMT2A* locus.** This is the first time a human SUV39H1 occupancy track has been read at this locus by anybody. `absence-no-h3k9-writer-or-reader-localised-at-the-kmt2a-locus` classified the writer as `never-measured`; `ana-q113-...` found the deposits and deliberately left them unread to avoid a collision [E1][E3]. **The numbers.** Mean signal over each interval, scored against (i) **59,821 width-matched windows centred on ENCODE CTCF point-source summits** [E4] and (ii) 200,000 random tiles of the same width, which is context and is never scored [E3]: | interval | width | SU-DIPG13 mean | pct vs CTCF anchors | SU-DIPG36 mean | pct vs CTCF anchors | |---|---:|---:|---:|---:|---:| | **exon11 (the anchor)** | 147 | **0.0000** | **0.00** | 1.6660 | 74.17 | | ctcf_element | 1,048 | 0.6622 | 34.18 | 0.9932 | 55.10 | | ctcf_summit_1kb | 1,001 | 0.6933 | 35.77 | 0.9122 | 37.68 | | dcal_cut_1kb | 1,000 | 0.6940 | 35.80 | 0.6270 | 33.40 | | bcr_core | 7,466 | 0.8889 | 37.21 | 0.7535 | 25.28 | | kmt2a_tss_2kb | 2,001 | 0.2192 | 8.72 | 1.2193 | 54.39 | | kmt2a_gene | 90,377 | 0.9381 | 39.74 | 1.1109 | 49.66 | **Question A - does SUV39H1 occupy this locus at all? NOT OCCUPIED, both lines, every interval.** The pre-registered rule needed mean >= 2x the random-tile centre AND percentile >= 50 in both tracks. The random-tile *median* is exactly 0 in these sparse CUT&RUN tracks (57.7-71.1% of random 147 bp tiles are uncovered), so the fold was scored against the random-tile *mean*, disclosed as deviation D5; it changes nothing in either direction. **The largest fold reached at any *KMT2A* interval in either line is 1.21x** (`kmt2a_gene`, SU-DIPG36). Nothing in 90 kb of this gene beats the average random tile by more than a fifth. **Question B - is the anchor special? UNINTERPRETABLE by this arm's own rule; descriptively, ordinary.** The pre-registration declared this arm **C4-blind before the data were read**, because no CTCF experiment exists in any DIPG or H3K27M line in any archive [E4], and §7 says an arm failing C4 is UNINTERPRETABLE. That verdict is not being upgraded here. Descriptively: **0 of 2** SUV39H1 tracks put the anchor above the 75th percentile, and the median of the two exon-11 percentiles is **37.06** - inside the FAILS band [25, 75] had the arm been interpretable. **The SU-DIPG13 zero is a measured zero, not a gap.** `pybigtools` 0.3.0 changed its uncovered-base default to NaN between versions, so the pre-registration required a per-file probe. Run on a known-empty interval and on the anchor itself, **all three modes returned identical values in all four files** (`default`, `fillna=0`, `fillna=None`), and `records()` returns **one raw interval covering the exon-11 window in SU-DIPG13, of value zero**. These bigWigs carry explicit zero-valued records, the trap does not fire in this deposit, and 0.0000 means *measured and empty*.

Go deeper
The controls, including the two that failed and were not re-tuned

C1 (dynamic range) - DEGENERATE as written, passes on a disclosed repair. The gate divides by the median of random 147 bp tiles, which is exactly 0.0 in three of four tracks. top/0 = inf >= 5 compares true while measuring nothing: the gate was degenerate, not passed, and results_bigwig.json still records the inf. Recomputed against the tile mean, the top 0.1% of CTCF-anchor windows is 22.2x / 21.3x / 16.5x / 18.2x - the tracks have real dynamic range.

C1b (heterochromatin positive control) - FAILS in all four tracks at every window tried. Pre-registered: >= 3x over the chr19q13 KRAB-ZNF cluster. The 1 Mb window gives 1.08x / 1.26x / 1.01x / 1.05x over the tile mean. Post-hoc and labelled post-hoc, at finer scale: the 4 kb chr19 ZNF-cluster window this project already uses gives 1.22x / 0.47x / 0.27x / 1.13x, the ZNF274 gene body 1.36x / 1.31x / 1.14x / 1.28x, and the ZNF274 3' 10 kb window 2.20x / 1.64x / 1.62x / 1.81x. The best number obtained anywhere is 2.20x against a required 3x. The threshold was not moved.

AND THE SAME CONTROL PASSES IN THE OTHER DEPOSIT, WHICH CHANGES WHAT THE FAILURE MEANS. The MCF-7 arm of this beat ran the identical control on GSE139196's H3K9me3 pair and got 3.05x and 4.15x at the 1 Mb chr19q13 window and 3.79x and 5.41x at ZNF274 3' - comfortably past the pre-registered 3x. So the control is not wrong for H3K9me3; it works. I had reasoned, before the MCF-7 H3K9me3 tracks were read, that two failures in two deposits made the control more likely to be at fault than the data. My own second arm refuted that. The honest reading is now: the GSE225028 H3K9me3 CUT&RUN libraries do not resolve the canonical constitutive H3K9me3 domain that a matched ChIP in another cell type resolves at 3-5x, and the sensitivity of this deposit at a 147 bp element is correspondingly unestablished. The SUV39H1 tracks fail the same control in BOTH deposits (1.22x/0.47x here, 0.93-1.15x in MCF-7), which is a separate and unsurprising fact - the KRAB-ZNF cluster is TRIM28/SETDB1's address, not SUV39H1's, whose canonical address is pericentromeric satellite. This is the single reason this entry is graded moderate rather than strong.

C3 (gene desert) - behaves. chr2:35,000,000-36,000,000, verified live against Ensembl on 2026-09-22 to contain no protein-coding gene, reads 0.95x / 0.70x / 1.42x / 0.62x of the tile mean - and so, notably, does most of the KMT2A locus in these tracks.

C4 - could not be run, declared so in advance.

A mapping-artefact signature that fell out of a housekeeping fix

The query universe was restricted to the primary assembly midway through, for a mechanical reason: a bigWig written from a bedGraph carries only the chromosomes its data touched, so an unlocalised-scaffold anchor raised a KeyError on one track and not another. That dropped 71 of 59,892 SK-N-SH CTCF anchors - 0.12% of the set.

It also halved the C1 numerator. The mean of the top 0.1% of anchor windows fell from 42.8 to 21.5 in SU-DIPG36 and from 27.6 to 22.2 in SU-DIPG13. A tenth of a percent of the anchors, all on unplaced scaffolds, were carrying a large share of the strongest SUV39H1 values in the whole set. That is the signature of repeat-driven mismapping, and it is consistent with the C1b failure: SUV39H1's canonical address is pericentromeric satellite, which Bowtie2 with one copy per position cannot place, so what survives into the primary assembly may be the least informative part of this enzyme's distribution. Both C1 numbers are reported above.

What this result cannot reach

Lineage. SU-DIPG13 and SU-DIPG36 are patient-derived H3K27M diffuse midline glioma neurospheres. Infant KMT2A::AFF1 B-ALL is a B-lymphoid disease of fetal-liver-derived progenitors. Nothing transfers. Cycle 65 established the exact shape of the risk at this very element: BRCA1 sits at the 93.7th and 92.3rd percentile of same-cell CTCF anchors at the exon-11 anchor in HeLa-S3 and H1, and at the 34.0th and 45.1st in the only two blood lines, where ENCODE's 300,000-peak pre-IDR lists carry no BRCA1 peak at all - occupancy here is cell-type-restricted, so a negative in two glioma lines bounds two glioma lines.

State. Both are resting, unirradiated cultures. hyp-the-writer-reaches-the-kmt2a-anchor-only-through-ctcf-and-only-after-the-break predicts damage-conditional recruitment; a resting-state zero is what that hypothesis predicts and therefore does not touch it.

Replication. One sample per antibody per line. The difference between 0.0000 and 1.6660 at exon11 cannot be decomposed into line difference and technical variation from these files at all, and no entry should treat either number as replicated.

Sensitivity. The C1b failure means the deposit's sensitivity at a 147 bp element is not established. The negative is reported with that weight and no more.

Relationship to existing entries, and the open questions it closes
  • absence-no-h3k9-writer-or-reader-localised-at-the-kmt2a-locus (Q110) recorded SUV39H1: never-measured. Corrected in place this cycle: it is now measured in two human cell types and negative in both.
  • ana-q113-geo-holds-human-suv39h1-occupancy-data-that-the-encode-only-census-could-not-see asked "Is SUV39H1 present over the 90 kb KMT2A gene body in either deposited system? ... This beat deliberately did not run it." Answered for the DIPG system: no - 0.94x and 1.11x the random-tile mean, 39.7th and 49.7th percentile of width-matched CTCF anchors.
  • exp-suv39h1-and-h3k9me3-cutandrun-in-two-patient-derived-diffuse-midline-glioma-lines is the deposit entry; this is its first reading.
  • genome-kmt2a-exon-11-ctcf-anchor-is-a-mid-rank-ctcf-anchor-in-encode-cells (Q112) is the same instrument-control story in six other cell types.
Facts and figures
scriptanalyses/q114-suv39h1-at-the-kmt2a-locus/run_bigwig.py and run_bigwig_posthoc.py
statusrun
run date2026-09-22
run byQ114-suv39h1-at-the-locus
result summarySUV39H1 is NOT OCCUPIED at any of seven KMT2A intervals in either SU-DIPG13 or SU-DIPG36. exon11 reads 0.0000 (an explicit zero record) and 1.6660; the largest fold at any interval in either line is 1.21x the random-tile mean; median exon-11 percentile across the two tracks is 37.06 of 59,821 class-matched CTCF anchors.
genome buildGRCh38/hg38, verified from EACH FILE'S OWN chrom table across chr1, chr2, chr11, chr19 and chrX; no coordinate converted by an offset
cell typeSU-DIPG13 and SU-DIPG36, patient-derived H3K27M diffuse midline glioma neurospheres
cell stateresting, unirradiated, stem-like neurosphere culture
lineagepaediatric glioma - NOT haematopoietic, NOT KMT2A-rearranged
shown for fusionNO - nobody has looked in any KMT2A fusion system; this is a glioma deposit
verdict question A occupancyNOT OCCUPIED at exon11, bcr_core, kmt2a_tss_2kb and kmt2a_gene, in both lines
verdict question B class matchedUNINTERPRETABLE - the arm was declared C4-blind before running. Descriptively the anchor is ordinary: median exon-11 percentile 37.06, 0 of 2 tracks above the 75th.
prior heldtrue
n class matched anchors59821
n class matched anchors for the 90kb interval4998
interval sourceanalyses/q112-signal-level-reread/PREREGISTRATION.md, itself from analyses/q35-brca1-cbx3-encode-at-the-kmt2a-locus - found in the corpus, not invented
nan fillna probeall three modes (default, fillna=0, fillna=None) returned identical values in all four files; records() returns one raw zero-valued interval over exon11 in SU-DIPG13, so the zero is explicit rather than uncovered
inputs
[
  "GSM7037531",
  "GSM7037532",
  "GSM7037529",
  "GSM7037530",
  "ENCFF685KTA"
]
intervals grch38
{
  "exon11": "chr11:118488614-118488760",
  "ctcf_element": "chr11:118488159-118489206",
  "ctcf_summit_1kb": "chr11:118488194-118489194",
  "dcal_cut_1kb": "chr11:118484362-118485361",
  "bcr_core": "chr11:118482422-118489887",
  "kmt2a_tss_2kb": "chr11:118435456-118437456",
  "kmt2a_gene": "chr11:118436456-118526832"
}
exon11 mean
{
  "GSM7037531_SU-DIPG13_SUV39H1": 0,
  "GSM7037532_SU-DIPG36_SUV39H1": 1.6660438171804768
}
exon11 percentile vs class matched ctcf anchors
{
  "GSM7037531": 0,
  "GSM7037532": 74.17
}
controls
{
  "C1_as_preregistered": "DEGENERATE - divides by a random-tile median of exactly 0.0 in 3 of 4 tracks",
  "C1_recomputed_over_tile_mean": {
    "GSM7037531": 22.2,
    "GSM7037532": 21.3,
    "GSM7037529": 16.5,
    "GSM7037530": 18.2
  },
  "C1b_znf_positive_control": "FAILED at all four window scales in this deposit; best value 2.20x (ZNF274 3' 10 kb, SU-DIPG13 SUV39H1) against a pre-registered 3x. The IDENTICAL control PASSES on the MCF-7 H3K9me3 pair of the other deposit (3.05x and 4.15x at 1 Mb; 3.79x and 5.41x at ZNF274 3'), so the failure is a property of these CUT&RUN libraries rather than of the control. The SUV39H1 tracks fail it in both deposits, which is expected - the KRAB-ZNF cluster is not SUV39H1's address.",
  "C3_chr2_desert_over_tile_mean": {
    "GSM7037531": 0.95,
    "GSM7037532": 0.7,
    "GSM7037529": 1.42,
    "GSM7037530": 0.62
  },
  "C4": "could not be run - no CTCF experiment in any DIPG or H3K27M line in any archive"
}
How this could bite us back
  • Chaetocin is the SUV39H1 inhibitor this deposit's parent study advances, and a negative occupancy result at one locus is not a safety argument for it: chaetocin is a promiscuous thiodioxopiperazine with documented redox activity, so no phenotype under chaetocin can be attributed to SUV39H1 catalysis from these data.
  • If SUV39H1 is not at this locus in any cell, an H3K9me3 change measured here under a menin inhibitor could not be attributed to SUV39H1 either. That removes a mechanism this graph has implicitly held in reserve and leaves SETDB1, EHMT2/G9a and the KDM3/KDM4 erasers as the candidates - a shift that matters, because G9a inhibitors and KDM inhibitors have very different toxicity profiles from a SUV39H1 agent.
  • A negative from a non-haematopoietic cell is the easiest result in this graph to over-quote. Stripped of its first clause it becomes 'SUV39H1 is not at KMT2A', which this measurement cannot support and which would wrongly close the CD34+ experiment that is the real next step.
Still open
  1. Does the same read in a haematopoietic cell change the answer? No archive holds a SUV39H1 occupancy experiment in any blood cell, so this needs wet work - a CUT&RUN with CST #8729S in CD34+ cells or a KMT2A-rearranged line is the single experiment that would close the SUV39H1 arm of this graph's H3K9 argument.
  2. Why does the GSE225028 H3K9me3 pair show no KRAB-ZNF enrichment at any of four window scales? Either these libraries have poor signal-to-noise below the megabase, or the ZNF-cluster address is weaker in neural cells than the epithelial and lymphoid literature implies. The raw SRA runs would separate library complexity from biology and were not retrieved.
  3. Is the SU-DIPG13 versus SU-DIPG36 difference at exon11 (0.0000 versus 1.6660) replicate noise or a line difference? One sample per antibody per line means this deposit cannot answer it, and no entry should treat either number as a replicated measurement.
Every source behind this entry (4)
  1. E1established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.dataset

    GSE225028 deposits four hg38 bigWigs from patient-derived H3K27M diffuse midline glioma neurospheres - SUV39H1 (CST #8729S) and H3K9me3 (CST #13969S), each in SU-DIPG13 and SU-DIPG36 - aligned with Bowtie2 v2.3.4.1, and no peak file is deposited.

    What this evidence cannot show

    Per-sample SOFT records fetched 2026-09-22. There is no IgG and no SUV39H1-knockout track in the deposit, so antibody specificity cannot be checked from these files, and each antibody was run once per line - there are no replicates. The two H3K9me3 bigWigs carry a 455-entry chrom table and the two SUV39H1 bigWigs a 195-entry one, in the same deposit.

    Evidence that argues against this (1)
    • GEO metadata states a pipeline; it does not certify that the pipeline ran. The build was therefore re-verified from each file's own chrom table rather than from these lines, and all four are GRCh38 across chr1, chr2, chr11, chr19 and chrX. source
  2. E2established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.paper

    The DIPG CUT&RUN deposit accompanies a study of chaetocin-mediated SUV39H1 inhibition in diffuse midline glioma, which is why SUV39H1 was profiled in these cells at all.

    PMID 38011799DOI
    What this evidence cannot show

    Metadata fetched live from PubMed efetch 2026-09-22, DOI read only from PubmedData/ArticleIdList. The abstract was read; the full text was not, so nothing here is claimed about what the paper's own figures show at any locus.

    Evidence that argues against this (1)
    • A paper whose thesis is that SUV39H1 matters in these cells is the least likely place to find a reported SUV39H1 absence at an unrelated locus. Nothing here rests on the paper's interpretation, only on its deposited tracks. source
  3. E3strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.dataset

    The null, the threshold, the controls and a NEGATIVE prior were written to disk before any signal value at the KMT2A locus was read from either deposit, and every deviation from them is appended to the same file rather than edited into it.

    Q114 pre-registration - SUV39H1 occupancy at the KMT2A locus, in the only two human systems where it has ever been measured · Q114-suv39h1-at-the-locus · oncofusion-knowledge-graph (analyses/q114-suv39h1-at-the-kmt2a-locus/PREREGISTRATION.md) 2026
    What this evidence cannot show

    The pre-registration discloses one prior read - ChIP-Atlas's hg38 SUV39H1 target-gene list, which derives from the MCF-7 samples and does not touch this deposit - and it declares before running that C4 cannot be run in this arm at all. Six deviations (D1-D6) are appended, three of them failures of the author's own controls.

    Evidence that argues against this (1)
    • A pre-registration written by the agent that then runs the analysis is weaker than an external one. What it buys is only that the threshold could not move after the numbers appeared; the appended deviations are the audit trail, and two of them record controls that failed and were not re-tuned. source
  4. E4strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.database

    The class-matched null is 59,821 primary-assembly CTCF point-source summits from ENCODE optimal-IDR CTCF peaks in SK-N-SH (ENCFF685KTA, ENCSR541AMF, GRCh38) - a neural line, because no CTCF experiment exists in any DIPG or H3K27M glioma line in any archive.

    What this evidence cannot show

    SK-N-SH is a neuroblastoma line, not a diffuse midline glioma. 71 of 59,892 peaks fall on unlocalised or alt scaffolds and were dropped so that the two arms query an identical universe; that 0.12% of the set contributed a large share of the highest SUV39H1 values, which is itself reported below.

    Evidence that argues against this (1)
    • A cross-cell-type null may include anchors that are not anchors in DIPG, which would lower the comparison distribution and INFLATE the target's percentile. That bias runs against the negative reported here, not in its favour. source
Connected to