SALVAGED from a beat killed mid-write: with direct DOT1L inhibition as the positive control that bounds the achievable effect, VTP-50469 achieves only F = 0.09-0.25 of it. SGC0946 strips H3K79me2 to 6.7-14.8% of DMSO; the menin inhibitor leaves 77-91%. And the bin-wise responses do not correlate (rho = -0.026)
also called the fraction of achievable H3K79me2 redistribution · how much of DOT1L inhibition does a menin inhibitor buy · Q122 salvage, arm 1
Two drugs can both be said to reduce a chemical mark on DNA packaging, but until you measure how much is achievable you cannot say whether one of them does much at all - so this compared the menin inhibitor against a drug that blocks the enzyme writing the mark directly, in the same cells and the same experiment. Blocking the writing enzyme removes almost all of the mark, leaving only about 7 to 15 per cent of the untreated level, whereas the menin inhibitor leaves 77 to 91 per cent, meaning it buys between a tenth and a quarter of what is actually achievable. The two drugs also do not shift the mark at the same places in the genome, which argues that whatever the menin inhibitor does to this mark does not work through the writing enzyme's pathway at all.
The same thing, for a scientist
**PROVENANCE FIRST: this entry is a coordinator SALVAGE.** Q122 was killed by the account session limit after completing its computation and writing `PREREGISTRATION.md`, `run.py` and a 73 KB `results.json`, but **before it wrote any node file**. The standing instruction forbids retrying on a session limit, so the coordinator recovered the numbers from the beat's own committed `results.json`. **No figure here was computed by me and none was re-derived from raw data; they are read out of the beat's output and cited to it.** **The measurement my own brief asked for, now made.** SGC0946 is direct DOT1L inhibition, so it bounds the H3K79me2 redistribution achievable at these sites. Defining V = VTP50469/DMSO and G = SGC0946/DMSO as within-sample share ratios, and F = (1−V)/(1−G) as the share of the **achievable redistribution**: | arm | population | V (menin inh.) | G (DOT1L inh.) | **F** | |---|---|---:|---:|---:| | murine MLL-AF9, F3 | S_FLAG | 0.7734 | 0.0937 | **0.2500** | | murine MLL-AF9, F3_NT | S_FLAG | 0.8318 | 0.1499 | **0.1979** | | murine MLL-AF9, F3 | S_K79 | 0.9117 | 0.0675 | **0.0947** | | murine MLL-AF9, F3_NT | S_K79 | 0.8462 | 0.1482 | **0.1806** | | K562 (no KMT2A fusion) | S_K79 | 0.8291 | 0.0666 | **0.1831** | **So a menin inhibitor buys between a tenth and a quarter of what DOT1L inhibition buys on this mark.** The positive control is emphatic — SGC0946 takes the mark to **6.7-14.8%** of DMSO in the murine arms and 6.7% in K562, and the pre-registered instrument check `I2` passed in all four arms. Until this, every statement in this corpus about "the mark falls under a menin inhibitor" was a bare number with no scale. **And the two drugs do not move the same bins.** Pre-registered refutation **R4 FIRES**: the bin-wise correlation between the VTP response and the SGC0946 response at S_K79_F3 is **rho = −0.026**, against a firing threshold of rho < 0.10. Whatever the menin inhibitor does to H3K79me2, it is not a graded version of DOT1L-pathway inhibition. **Arm 1's F is a LOWER BOUND and the beat said so.** Its numerator is measured at 48 h and its denominator at 72 h, so VTP's 72-hour effect could be larger. The K562 arm (both at 7 d) and the MOLM13 arm (96 h) are not subject to that asymmetry.
Why a fraction and not a number - and why this one is a fraction of REDISTRIBUTION, not of loss
The beat is explicit about something worth preserving, because it is the kind of wording that decays into a stronger claim on being quoted.
These are within-sample share ratios, not absolute levels. GSE260742 has no usable spike-in — Q117 verified zero hits for the whole spike-in vocabulary across 419,824 characters of sample records — so every cross-condition figure in this series is a share or a rank. A share ratio of 0.0675 under SGC0946 means the mark's share of that sample's signal at those sites fell to 6.75% of its DMSO share. It does not mean 93% of the mark was removed from the cell.
Hence F is the share of achievable REDISTRIBUTION, in the beat's own words, not the share of achievable loss. An absolute-loss version of this statement needs a spike-in or an orthogonal quantitative measurement, and neither exists in this series. Any future quotation of "the menin inhibitor achieves a quarter of DOT1L inhibition" must carry that qualifier or it becomes a claim the data cannot support.
What the beat added that the brief never asked for
My brief asked for three arms of one series. The beat built four, and two of the additions are the most informative parts of the result.
A K562 negative control — a line with no KMT2A fusion at all — which is the arm that turns a measurement about a drug into a measurement about a mechanism. It has its own entry: ana-q122-salvaged-the-menin-inhibitors-h3k79me2-effect-is-indistinguishable-in-a-line-with-no-kmt2a-fusion.
An H2AK119ub drug-engagement arm, which gives an independent read that the compound was doing something in the cells at all — a control this corpus has wanted and never had for this series.
And build verification from each file's own chromosome extents in all four arms, which caught that the four K562 files declare an mm10 container while their coordinates are hg38 (the beat's DEVIATION 1). That is exactly the class of silent error that has cost this project cycles before, and it was found by a check the brief did not request.
| SALVAGE PROVENANCE | Q122 completed its computation and wrote PREREGISTRATION.md, run.py and results.json, then was killed by the account session limit before writing any node file. The standing instruction forbids retrying on a session limit. The coordinator read the figures out of the beat's committed results.json and wrote this entry; no figure was re-derived from raw data by the coordinator. |
|---|---|
| series | GSE260742 |
| normalisation | no usable spike-in; every cross-condition figure is a within-sample share ratio |
| V is | VTP50469 / DMSO within-sample share ratio |
| G is | SGC0946 / DMSO within-sample share ratio |
| F is | (1-V)/(1-G), the share of the achievable REDISTRIBUTION, not of achievable loss |
| I0 no track artefact dominated | PASS in all four arms; worst largest-bin share 0.00031 |
| what this supplies that the corpus lacked | a SCALE. Every prior statement here that 'the mark falls under a menin inhibitor' was a bare share ratio with no reference to what was achievable. |
| absence class | measured-not-curated |
| absence class note | The measurement exists, in this repository, in analyses/q122-h3k79me2-three-arms-with-replicates/results.json, computed and pre-registered by a beat that died before it could curate any of it into the graph. This entry curates one part of it. The remaining scored criteria in that file - C2, C3, C6, C7, C8 and R1/R3/R5 - are measured, committed and STILL UNCURATED, and are named in the open questions so a future cycle does not have to rediscover them. |
arm1 murine mllaf9
{
"F3_at_S_FLAG": {
"V": 0.7734,
"G": 0.0937,
"F": 0.25
},
"F3NT_at_S_FLAG": {
"V": 0.8318,
"G": 0.1499,
"F": 0.1979
},
"F3_at_S_K79": {
"V": 0.9117,
"G": 0.0675,
"F": 0.0947
},
"F3NT_at_S_K79": {
"V": 0.8462,
"G": 0.1482,
"F": 0.1806
},
"lower_bound_note": "numerator 48 h, denominator 72 h - F is a LOWER BOUND on VTP's 72-hour effect"
}arm2 k562 no kmt2a fusion — the numbers behind the chart
{
"V": 0.8291,
"G": 0.0666,
"F": 0.1831
}instrument check I2 passed all four arms — the numbers behind the chart
{
"arm1_F3_at_S_K79_G": 0.0675,
"arm1_F3NT_at_S_K79_G": 0.1482,
"arm2_K562_G": 0.0666,
"arm3_MOLM13_G": 0.4243
}R4 not a dot1l pathway response
{
"rho": -0.026,
"fires_if": "rho < 0.10",
"verdict": "FIRES"
}beat added beyond the brief
[ "a K562 negative control with no KMT2A fusion - the arm that makes this mechanistic rather than descriptive", "an H2AK119ub drug-engagement arm, an independent read that the compound acted in the cells", "build verification from each file's own chromosome extents in all four arms, which caught that the four K562 files declare an mm10 container while their coordinates are hg38" ]
- Q122's results.json holds SIX MORE scored criteria that no entry yet records: C2 and C3 (the mark's own domains, and between-culture spread 0.0584/0.0655 against a cross-partner delta of 0.2089), C6 (MOLM13 replicates at 0.9063 in GSE260742 against q120's 0.8667 in GSE127507 - an independent series), C7 (the MLL1 shared-epitope channel IS moved, 0.3012 under SGC0946 and 0.3298 under VTP, which sits awkwardly beside q120's 'inert' framing at a different population), C8 (flatness is not an aggregation artefact, firing across eight strata) and R1/R3/R5. Curating those is a reading task, not a compute task.
- The 48 h versus 72 h asymmetry in arm 1 is the single biggest threat to the headline. GSE260742 may contain a 72 h VTP H3K79me2 sample; if it does, one download closes it.
- R4's rho = -0.026 should be recomputed restricted to bins with high DMSO signal, to separate mechanistic independence from bin-level noise. Not run.
- E1moderate confidence 3 of 5: One strong study, or several smaller consistent ones.analysis
In murine MLL-AF9 (GSE260742, mm10), direct DOT1L inhibition with 5 uM SGC0946 reduces the H3K79me2 within-sample share ratio to G = 0.0937 and 0.1499 at the FLAG-defined site set and 0.0675 and 0.1482 at the mark's own domains across two culture families, while 500 nM VTP50469 leaves V = 0.7734, 0.8318, 0.9117 and 0.8462 at the same sets. The resulting fraction of achievable redistribution F = (1-V)/(1-G) is 0.2500, 0.1979, 0.0947 and 0.1806. In K562, which carries no KMT2A fusion, G = 0.0666, V = 0.8291 and F = 0.1831. The pre-registered instrument check I2 (SGC0946 must visibly redistribute the mark) passed in all four arms, including MOLM13 at G = 0.4243. The pre-registered refutation R4 fires: bin-wise correlation between the VTP and SGC0946 responses at S_K79_F3 is rho = -0.026 against a threshold of rho < 0.10.
analyses/q122-h3k79me2-three-arms-with-replicates/results.json and PREREGISTRATION.md, this repository; GSE260742 .tdf files binned at 2 kb; pre-registration written before any bin file existed · Oncofusion knowledge graph, Q122 beat (computation) and coordinator (salvage of the numbers into this entry) · internal analysis, reproducible from the repository 2026What this evidence cannot show
CONFIDENCE IS moderate AND THE REASON IS PROVENANCE, NOT THE NUMBERS. THE BEAT THAT COMPUTED THESE FIGURES DIED BEFORE WRITING ANY ENTRY, SO NO AGENT EVER WROTE PROSE AGAINST THEM, AND I HAVE NOT RE-DERIVED A SINGLE FIGURE FROM RAW DATA - every value is read out of the beat's own results.json. Its scripts and pre-registration are committed and the computation is reproducible, but it has had no second pass by anyone. SPECIFIC LIMITS THE BEAT ITSELF RECORDED: (1) ARM 1's F is a LOWER BOUND because its numerator is 48 h and its denominator 72 h; (2) there is no usable spike-in in GSE260742, so every figure is a within-sample SHARE ratio and F is the fraction of achievable REDISTRIBUTION, not of achievable loss; (3) the two murine culture families F3 and F3_NT are two cultures, NOT replicates, and nothing in 158 sample records establishes they are one culture; (4) the four K562 files declare an mm10 container while their coordinates are hg38. Doses and durations differ between arms (500 nM VTP 48 h/7 d/96 h; 5 uM SGC0946 72 h/7 d), so F is not strictly comparable across arms.
Evidence that argues against this (3)
- The strongest objection is that F's denominator flatters the numerator's smallness. SGC0946 inhibits the enzyme that writes H3K79me2 directly, so G near zero is close to the maximum any intervention could achieve and F is therefore a comparison against a near-ceiling. A menin inhibitor is not a DOT1L inhibitor and was never expected to match it; reporting F = 0.09-0.25 as though it were a disappointing efficacy figure would be a category error. What F legitimately establishes is a SCALE for statements this corpus has been making without one - not that the menin inhibitor underperforms at a job it was not doing.
- R4's rho = -0.026 is a bin-wise correlation over a 2 kb-binned genome, and near-zero correlation between two noisy within-sample ratios is also what you would see if BOTH responses were dominated by measurement noise at the bin level rather than if they were mechanistically independent. The beat's I0 instrument check (no track dominated by one artefact bin, largest bin share 0.00026) argues against gross artefact but does not establish per-bin signal-to-noise. A stronger version would correlate at coarser resolution or restrict to bins with high DMSO signal, and was not run.
- Arm 1's 48 h vs 72 h asymmetry cuts directly at the headline. If VTP's effect on this mark continues to develop between 48 and 72 hours - which is plausible for a mark whose turnover is slow - then F at matched timepoints would be larger, and the 0.0947 figure in particular could understate substantially. The K562 (7 d both arms) and MOLM13 (96 h) arms are not subject to it, and K562's F = 0.1831 sits inside the murine range, which is weak reassurance rather than a control for the asymmetry.
supplies the SCALE this corpus lacked: direct DOT1L inhibition takes the mark to 6.7-14.8% of DMSO, a menin inhibitor leaves 77-91%, so VTP buys F = 0.09-0.25 of the achievable redistribution
- related to →Pinometostat (EPZ-5676) — the DOT1L inhibitor, and the most instructive failure in this field
SGC0946 is the tool DOT1L inhibitor that bounds the achievable effect here; the clinical DOT1L inhibitor's disappointment is the other end of the same axis
the K562 no-fusion arm of the same salvaged beat; its F = 0.1831 sits inside the fusion-bearing murine range of 0.0947-0.2500
the same beat's C1/C2/C3/C6 fire, giving partner-dependence a species-independent leg (murine MLLT3 V 0.7734/0.8318, MOLM13 replicating at 0.9063 in an independent series against q120's 0.8667) - those criteria are measured, committed and still uncurated
F replicates in a third and fourth murine culture at 0.2148 and 0.2314, inside Q122's 0.0947-0.2500 range, with the same 48 h/72 h lower-bound qualifier
qualifies rather than refutes the arithmetic: F stays valid as a ratio of two share-ratio deficits, but the word ACHIEVABLE presumes the two drugs act on one axis, and the floor moves 0.69-0.78 under VTP against 1.89-2.01 under SGC0946 in four cultures with no overlap
the arm dispatched to remove F's 48 h/72 h caveat strengthens it: the kinetic series is non-monotone, five statistics disagree about the sign of its last step, and its 24 h point shows more redistribution than the 48 h main-arm point, so F keeps the words lower bound
F = (1-V)/(1-G) has the same shape and goes negative when V > 1; Q122's own committed JSON contains a V of 1.0829, which is not among the six F values it reported, so this is a near-miss rather than an error
why that beat's 48 h/72 h asymmetry cannot be closed by re-analysis: the series holds no 72 h VTP H3K79me2 sample and no timed SGC0946 sample, so neither side can be moved. The cheapest fix is one SGC0946 48 h ChIP at the bench