Cycle 110's 'D10 GAINED 66%' rescored on the absolute scale: D10 lost 76.5% of its H3K79me2, no stratum gained anything, and on the genuine RS4;11 rows the gradient spans 41-fold against the contaminated table's 30-fold
also called nothing gained · D10 did not gain 66% · the eviction gradient rescored
An earlier cycle of this project concluded that the least-marked tenth of genes gained chromatin signal while the most-marked lost it; with the fruit-fly yardstick properly applied, that tenth had lost three-quarters of its signal and nothing gained anything. The shape of the earlier result was right and the line between gaining and losing was in the wrong place.
The same thing, for a scientist
**Statement.** Cycle 110 (`analyses/what-a-share-does-to-a-known-collapse`) reported, for RS4;11 under EPZ5676 at day 9, that *'TOP kept 5.6% of its signal, D10 GAINED 66%'* and that the global absolute change was 5.84x down. Those numbers are reproduced here exactly - TOP 0.0556, D10 1.6620, global raw ratio 0.1348 - which confirms the reproduction is faithful before it is corrected. **Two independent errors compounded.** 1. The deposited columns are raw, not spike-in scaled, so cycle 110's 'absolute_ratio' was a raw count-sum ratio, short by the factor `K = 0.14149` for this contrast. 2. It used all 57,794 rows of the RS4;11 file, of which 38,580 are MOLM13. **Corrected, on the same 57,794 rows so the two errors can be separated:** | stratum | cycle 110 called this 'absolute' | share ratio | TRUE absolute | gained? | |---|---:|---:|---:|---| | TOP | 0.0556 | 0.324 | **0.0079** | no | | D1 | 0.0617 | 0.360 | 0.0087 | no | | D5 | 0.6612 | 3.859 | 0.0936 | no | | D9 | 1.0941 | 6.386 | 0.1548 | no | | D10 | **1.6620** | 9.700 | **0.2352** | **no** | **D10 retained 23.5% of its H3K79me2. It did not gain 66%; it lost 76.5%.** Not one of the eleven strata has a true absolute ratio above 1. **And on the genuine RS4;11 rows only**, the gradient is much steeper: TOP 0.0075 to D10 0.3088, a span of **41x** against the contaminated table's 30x, with D6-D10 all above 0.16 where the contaminated table put D6 at 0.11.
What survives from cycle 110, exactly
Survives. The identity share(s) = absolute(s)/G to a maximum deviation of 0.002 across 66 strata - because a share is invariant to any per-sample constant, and K is a per-sample constant. So cycle 110's central methodological finding, that the share profile faithfully reproduces the shape of the underlying profile, is untouched and is if anything strengthened: it now has a second constant multiplier stacked on it and still holds.
Survives. The warning that a monotone share gradient rising from the peaks into the sparse bulk is the signature of a global decrease rather than evidence of redistribution.
Does not survive. Every statement of the form 'stratum X gained'. Cycle 110's own correction of cycle 110 - W4, which said 'the sparse deciles genuinely did fare better than the peaks (D10 GAINED 66%)' - reversed a false claim in the right direction and then over-corrected into a second false claim. The sparse deciles did fare better than the peaks. None of them gained. Both halves matter and the corpus has now held each one alone. 1
The general rule this is the third instance of
Cycle 110 established that a share is absolute/G, and that 'above 1' means 'lost less than average', not 'gained'. This beat establishes the dual, that an unscaled ratio is absolute/K, and that the same sentence applies to it.
Every per-sample normalisation is a constant multiplier. It never changes the shape. It only moves the zero line - and the sign of every gain/loss call is read off the zero line.
This is why a number labelled 'absolute' must name which constant put it there. Cycle 110's field was called global_absolute_fold_down and contained neither a share nor an absolute quantity. The name was doing work the arithmetic had not done. 12
| script | analyses/absolute-h3k79me2-in-the-index-fusion/run.py |
|---|---|
| status | run |
| result summary | Cycle 110's TOP 0.0556 and D10 1.6620 reproduced exactly; with K=0.14149 the true absolute ratios are 0.0079 and 0.2352; no stratum above 1. On genuine RS4;11 rows only, TOP 0.0075 to D10 0.3088. |
| K applied | 0.14149 |
| strata with true absolute ratio above 1 | 0 |
| d10 true PCT retained contaminated table | 23.52 |
| d10 true PCT retained genuine rows | 30.88 |
| top true PCT retained genuine rows | 0.75 |
inputs
[ "GSE149183 RS4_11 sitepro table", "analyses/what-a-share-does-to-a-known-collapse/results.json" ]
- A 'gain' that is really a 76% loss, if it had propagated into a target-selection argument, would have nominated the least-marked genes as the ones a DOT1L inhibitor spares - and they are in fact removed almost as completely as everything else.
- How many other entries in this corpus contain a field named 'absolute' that holds a ratio of two un-normalised sums? A grep on field names would bound it in minutes and nobody has run it.
- The genuine-rows gradient spans 41x from TOP to D10. Is that gradient a property of H3K79me2, of the sitepro window, or of the baseline-dependent background floor? Only a no-primary-antibody control library would separate them, and none is deposited.
- Cycle 110's cycle-103 comparison - that a 0.81-to-1.14 share span is milder than a 1.25x global change would produce - needs redoing with the correct global changes, which are 37-91x rather than 1.25-5.84x. That arithmetic has not been redone here.
- E1strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.dataset
Reproducing cycle 110's stratified table exactly and then applying the deposited dm6 scale factor shows every stratum, including D10, with a true absolute ratio below 1.
GSE149183 series supplementary files: RS4_11 H3K79me2 TSS-ChIPseq signal + dm6 spike-in sitepro table · Perner F, Gadrey JY, Xiong Y, et al. · NCBI Gene Expression Omnibus 2020What this evidence cannot show
The correction is arithmetic once the scale factor is accepted; the uncertainty is entirely in the scale factor, which is a single unreplicated dm6 count per library. Under the depositors' own %-fly convention K is 0.16854 rather than 0.14149, which puts D10 at 0.280 instead of 0.235 - still a 72% loss, so the conclusion is insensitive to the convention. A gain at D10 would require K above 0.60, four times either estimate.
Evidence that argues against this (1)
- If the Drosophila spike-in is recovered less efficiently as the human signal collapses, K is overstated and the true absolute ratios are higher. Even so, D10 would need the spike-in to be overstated four-fold to reach 1.0, which is far outside the 1.59x spread between the two conventions. Orlando DA et al. 2014
- E2established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.paper
The dataset is the DOT1L-inhibitor ChIP-seq series of Perner et al., Blood 2020, with H3K79me2 measured at days 3, 6 and 9 in MOLM13 and RS4;11 with Drosophila spike-in chromatin.
Novel inhibitors of the histone methyltransferase DOT1L show potent antileukemic activity in patient-derived xenografts. · Perner F, Gadrey JY, Xiong Y, Hatton C, Eschle BK, Weiss A, et al. · Blood 2020What this evidence cannot show
PMID and DOI fetched from NCBI eutils 2026-09-23. The paper itself makes no claim about decile-stratified gains; the claim being corrected is this project's own.
Evidence that argues against this (1)
- None found. This is a correction of an internal analysis, not of the source publication. NCBI GEO 2020
its W4 correction over-corrected: the sparse deciles did fare better than the peaks, but D10 lost 76.5% rather than gaining 66%
the shape finding survives the rescaling; only the gain/loss line moves
error 1 of 2: the field named absolute held a raw count-sum ratio
error 2 of 2: the RS4;11 rows behind it were two-thirds MOLM13
eleven strata of TSS-averaged H3K79me2 rescored on an external ruler
the same strata, the same conclusion: the shape is real and nothing gains