Oncofusion

start / graph / ana-the-canonical-kmt2a-targets-lose-more-only-because-they-are-top-decile-genes

Analysismoderate confidence 3 of 5: One strong study, or several smaller consistent ones.Q138-absolute-h3k79me2

Under EPZ5676 the canonical KMT2A-fusion targets look like they lose H3K79me2 far more than the genome (p=0.017 and p=0.0004), and against a baseline-matched null the entire effect vanishes (p=0.34 and p=0.22) - eight of the ten are simply top-decile genes

also called the HOXA effect is a baseline effect · matched null kills it · B3 fires

The genes a KMT2A fusion is famous for switching on appear to lose their chromatin mark much faster than average under a DOT1L inhibitor. They do not. They start with far more of the mark than average, and once you compare them with other genes that also start high, they behave like any of them.

The same thing, for a scientist

**Statement, RELATIVE.** Ten canonical KMT2A-fusion target genes - HOXA5, HOXA7, HOXA9, HOXA10, MEIS1, PBX3, CDK6, MEF2C, MYB, JMJD1C - were fixed in the pre-registration before any number was computed. All ten are present in both cell lines' tables, including in the 19,214 genuinely-RS4;11 rows, so none had to be dropped. At day 9 under EPZ5676, with 20,000 permutations and seed 20260923: | line | observed median log2 ratio | unmatched null | p | **baseline-matched null** | **p** | |---|---:|---:|---:|---:|---:| | RS4;11 | -3.874 | -1.499 | **0.017** | **-4.044** | **0.338** | | MOLM13 | -3.444 | -0.992 | **0.0004** | **-3.732** | **0.218** | **The difference between the two nulls is the result.** Against a naive random draw the targets look dramatically more sensitive. Against a draw matched on DMSO-signal decile the effect is not merely weakened, it is **absent and slightly reversed in sign** - the targets lose marginally *less* than comparable genes, well inside the permutation distribution. **Why: eight of the ten sit in decile 10.** In RS4;11, PBX3 (baseline 4,083), CDK6 (4,845), MYB (3,845), JMJD1C (3,332), HOXA7 (2,651) and HOXA9 (2,439) are all top-decile; MEIS1 is decile 9 and HOXA5 decile 8. Only HOXA10 (72, decile 4) and MEF2C (123, decile 5) are not. In MOLM13 nine of ten are decile 10. **Pre-registered branch B3 fires**, and it is the branch that exists because two beats in the last four cycles had results their two-option pre-registration could not express.

Go deeper
The absolute numbers, which are a different statement

Everything above is RELATIVE. The ABSOLUTE statement, which needs the spike-in, is that these genes are near-completely stripped. RS4;11, EPZ5676, day 9, percentage of H3K79me2 retained:

genebaselinedecileretained
CDK64,845100.33%
PBX34,083100.68%
MEIS11,80590.70%
HOXA51,35880.87%
MYB3,845100.88%
HOXA72,651101.06%
HOXA92,439101.40%
JMJD1C3,332101.68%
HOXA107245.78%
MEF2C12357.03%

Genome median retention in the same library: 4.72%. So HOXA9 holds less than a third of the genome-median fraction - which is the observation that makes the unmatched null fire - and yet the matched null says that is what a gene of HOXA9's starting level does. Both sentences are true and only one of them is about HOXA9. 1

The housekeeping anchor, which is not a null

Five housekeeping genes were fixed at the same moment as the targets. In RS4;11 under EPZ5676 at day 9 the decile-10 ones retain GAPDH 0.22%, RPLP0 0.44%, TUBB 0.44%, RPL15 0.61% - lower than every decile-10 canonical target. ACTB, which sits at decile 3, retains 9.38%.

That is a five-gene anchor, not a statistical null, and it is reported as such. It says the direction of any residual target effect is towards sparing, not towards extra sensitivity - consistent with the matched null's slightly reversed sign and with the compound-10 result in the sibling entry. It does not rescue a significance claim and is not used as one. 1

What this kills and what it leaves standing

Killed: any argument that a DOT1L inhibitor is target-selective at the chromatin level in the sense of stripping the HOXA/MEIS1 programme preferentially. It does not; it strips everything, and the appearance of preference is the base rate of high-signal genes.

Left standing: the therapeutic logic does not depend on chromatin selectivity. It depends on differential dependence on the mark - which genes cannot function without it - and that is a transcriptional and phenotypic question this measurement cannot address.

And a methodological point this project has now met four cycles running: the proportion that looks like an observation is the one that has no null. Here the unmatched p-values, 0.017 and 0.0004, are the ones that would have been quoted. 13

Facts and figures
scriptanalyses/absolute-h3k79me2-in-the-index-fusion/run.py
statusrun
result summaryEPZ5676 d9: RS4;11 observed -3.874, unmatched null -1.499 p=0.017, matched null -4.044 p=0.338; MOLM13 observed -3.444, unmatched -0.992 p=0.0004, matched -3.732 p=0.218.
n permutations20000
seed20260923
prereg branchB3
hoxa9 absolute PCT retained rs411 epz d91.4
genome median absolute PCT retained rs411 epz d94.72
inputs
[
  "GSE149183 RS4_11 and MOLM13 sitepro tables"
]
targets
[
  "HOXA5",
  "HOXA7",
  "HOXA9",
  "HOXA10",
  "MEIS1",
  "PBX3",
  "CDK6",
  "MEF2C",
  "MYB",
  "JMJD1C"
]
targets absent from table
[]
How this could bite us back
  • If a DOT1L inhibitor strips the fusion's target loci no faster than it strips baseline-matched genes, then a pharmacodynamic assay reading H3K79me2 at HOXA9 measures drug exposure, not target engagement in any selective sense - and a HOXA9 ChIP readout would give the same answer in a cell with no KMT2A rearrangement at all.
Still open
  1. Does matching on gene length as well as baseline change the answer? H3K79me2 spreads far past the 3 kb window, so a fixed window systematically under-samples long genes, and the HOXA cluster genes are short.
  2. gene-hoxa5 has no entry in this graph although HOXA5 is named in 29 entries, 25 of them written before this beat. It is a clean gap.
  3. Does the same matched-null test kill the apparent target selectivity of a MENIN inhibitor in the deposits this project already holds? That test has never been run with a baseline-matched null, and the three entries that assert target selectivity all use unmatched comparisons.
  4. A paediatric pinometostat trial exists on the registry but a full paper is not retrievable by PubMed title/abstract search on 'pinometostat' (23 records, none paediatric) - so the paediatric pharmacodynamics may exist only as a conference abstract. Worth a dedicated search before any paediatric claim is made.
Every source behind this entry (3)
  1. E1strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.dataset

    In both RS4;11 and MOLM13, the apparent extra sensitivity of canonical KMT2A-fusion target genes to EPZ5676 disappears against a DMSO-baseline-decile-matched permutation null.

    GSE149183 series supplementary files: RS4_11 and MOLM13 H3K79me2 TSS-ChIPseq signal + dm6 spike-in sitepro tables · Perner F, Gadrey JY, Xiong Y, et al. · NCBI Gene Expression Omnibus 2020
    What this evidence cannot show

    Ten genes, one library per condition, no replicates: a null result on a ten-gene set from single libraries has limited power, and 'no difference from matched null' is not 'no difference'. The matching is on DMSO-signal decile only; gene length, transcription rate and the fixed 4 kb window are not controlled, and H3K79me2 spreads into gene bodies well beyond 3 kb, so a long-gene bias is not excluded. The measurable set is genes with DMSO day-9 signal at or above 25 (17,341 in RS4;11, 54,958 in MOLM13); 1,873 and 2,836 rows respectively fall below it and are excluded, and that is printed by the script.

    Evidence that argues against this (1)
    • The founding work on DOT1L in KMT2A-rearranged leukaemia reports that H3K79me2 is specifically enriched across MLL-fusion target loci including HOXA9 and MEIS1 and that its loss there tracks with transcriptional shutdown - which is a statement about enrichment and consequence, not about differential rate of removal. This analysis does not contradict it: the targets do start far higher, which is the enrichment, and they are stripped at the rate their starting level predicts. Bernt KM et al. 2011
  2. E2established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.database

    JMJD1C and KDM3C are the same gene, so the canonical-target entry for it in this graph is gene-kdm3c.

    What this evidence cannot show

    Fetched live from the HGNC REST API on 2026-09-23, which returns hgnc_id HGNC:12313, alias_symbol including KDM3C, prev_symbol TRIP8 and uniprot_ids Q15652. This graph records that a previous beat logged gene-jmjd1c as missing twice while gene-kdm3c existed both times; the id was resolved against disk before use here rather than assumed. A first draft of this entry cited HGNC:12776 from memory and it was wrong - the id was fetched only after that.

    Evidence that argues against this (1)
  3. E3established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.paper

    EPZ-5676 (pinometostat) is a selective DOT1L inhibitor whose clinical activity in KMT2A-rearranged leukaemia was modest despite substantial H3K79 methylation reduction.

    PMID 29724899DOI
    What this evidence cannot show

    Adult trial; PMID and title fetched from NCBI eutils 2026-09-23.

    Evidence that argues against this (1)
    • Responses to pinometostat in the adult trial were followed by relapse, and preclinical work identified treatment-emergent resistance mechanisms - so near-complete removal of H3K79me2 is compatible with escape, which bounds how much any chromatin-depth measurement can predict. Campbell CT et al. 2017
Connected to