Eighteen libraries, no replicate: what the only absolute-scale chromatin dataset in this project is structurally unable to say, and why this project's own floors may not be borrowed to fill the gap
also called no replicates, no error bar · the floor may not be imported · what one library per condition forbids
The one dataset here that can measure how much of a chromatin mark is really lost was run once per condition - nine samples per cell line, no repeats. That means no result from it has an error bar, and borrowing an error bar from a different experiment would be worse than having none, because this project's own repeat measurements disagree with each other by up to twentyfold.
The same thing, for a scientist
**The design.** GSE149183's H3K79me2 arm is nine libraries per cell line - DMSO, EPZ5676 1 uM, compound 10 100 nM at days 3, 6 and 9 - eighteen in total, **with no replicate of any condition.** This is not an inference: the DOT1L and MEN1 occupancy samples in the same series **are** labelled `REP1` (GSM4491616-GSM4491625), so the depositors used replicate labelling where replicates existed and did not on the H3K79me2 samples. **What follows, stated as prohibitions rather than caveats.** 1. **No per-gene claim has an error bar.** 'Gene X is spared' and 'gene X differs between the lines' are unscoreable at the single-gene level. Where a single gene is named in this beat it is named as an illustration of a set-level result, never as a result. 2. **The count of spared genes is not an estimate.** It has no sampling distribution, and at low baseline it is dominated by counting noise - which is why it is reported at four baseline thresholds and watched to fall from 164 to zero. 3. **The between-line comparison is a difference of two single measurements.** It cannot be tested, and the fact that it reverses sign between the two drugs is the practical demonstration of that. 4. **The permutation p-values are over genes, not over experiments.** They answer 'is this gene set unusual within this library' and cannot answer 'would this replicate'. 5. **What survives is effect size.** A 37-91x global collapse is larger than any noise floor this project has ever measured by two orders of magnitude, and that is the only reason the headline is reportable at all. **And the borrowed-floor prohibition, which is specific to this corpus.** This project has measured same-condition noise floors and found them spanning roughly **20-fold across arms of a single deposit** - 15.97% at a top stratum against 3.03% in the bulk in one arm, and a four-times-tighter figure for H3K79me2 in another. A floor that varies that much between arms of one experiment is not a constant of the assay and **may not be transported into a different deposit, a different antibody lot, a different depth or a different cell line.** Importing one here would manufacture significance out of an unrelated experiment's variance.
What it would take to fix, and how cheap it is
Three replicate DMSO libraries and three replicate day-9 EPZ5676 libraries in RS4;11, with the same dm6 spike-in, would give:
- a measured same-condition floor on the absolute scale, which this project has never had - all its floors are floors on shares
- an interval on the 52x figure
- the ability to test whether the compound-10 target sparing replicates
- the ability to test whether the EPZ day-6 heavy tail is real
That is six libraries. The rest of the design already exists and is public. This is the smallest experiment in this project's queue that would convert four 'moderate' entries into testable ones. 1
The asymmetry worth keeping
Absence of replicates does not make a large effect unreliable and does not make a small effect reliable. It removes the ability to tell which you have. So the correct reading discipline for this deposit is:
- effects of 10x and above: report, with the design named
- effects between 2x and 10x: report as observations, never as findings
- effects below 2x: do not report as biology at all
Every number in this beat has been placed against that scale, and the one result that sits in the middle band - the 2.5x between-line difference - is the one that failed to replicate across drugs.
| libraries total | 18 |
|---|---|
| libraries per line | 9 |
| replicates | 0 |
| evidence that absence is real | the DOT1L and MEN1 occupancy samples in the same series carry REP1 labels |
conditions
[ "DMSO", "EPZ5676 1 uM", "compound 10 100 nM" ]
timepoints
[ "d3", "d6", "d9" ]
forbidden
[ "per-gene significance", "an interval on any absolute magnitude", "a same-condition noise floor", "testing the between-line difference", "importing a floor measured on a different deposit" ]
permitted
[ "effect sizes far above any plausible floor", "set-level permutation tests within a library", "pattern replication across the four drug-by-line arms" ]
- A pharmacodynamic claim built on an unreplicated chromatin measurement carries no confidence interval into a dose-selection decision, and the temptation is to supply one from an unrelated experiment - which is how a 20-fold-variable floor becomes a fixed threshold in a protocol.
- Does any public H3K79me2 dataset combine a spike-in with replicates? If one exists it would supply the first measured absolute-scale noise floor this project could legitimately use.
- This project's measured floors are all floors on SHARES. What is the corresponding floor on an absolute scale? It is not the same number and nobody has derived the relationship, which the identity absolute = raw x K makes tractable on paper.
- Would a leave-one-timepoint-out consistency check across the three timepoints serve as a weak pseudo-replicate here? It would not give a floor, but it would show which results depend on a single library.
- E1established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.dataset
The H3K79me2 arm of GSE149183 comprises nine libraries per cell line with no replicate of any condition, while the DOT1L and MEN1 occupancy samples in the same series carry REP1 labels.
GSE149183 supplementary filelist.txt and SOFT family file (sample list as served 2026-09-23) · Perner F, Gadrey JY, Xiong Y, et al. · NCBI Gene Expression Omnibus 2020What this evidence cannot show
The absence of a REP label is strong but not conclusive evidence of a single library; a second run could exist unlabelled or undeposited. The paper's methods were not read in full for this point - it rests on the deposited sample list.
Evidence that argues against this (1)
- A nine-point design with three timepoints and two drugs is itself a form of internal replication: the monotone decline is observed in four independent drug-by-line arms, and the target-sparing result in four independent libraries. That is replication of a pattern, not of a measurement, and it is the strongest thing this design offers. Perner F et al. 2020
- E2established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.paper
Spike-in ChIP-seq scale factors are derived from a single exogenous read count per library, so an unreplicated design gives an absolute scale with no measurable uncertainty.
Quantitative ChIP-Seq normalization reveals global modulation of the epigenome · Orlando DA, Chen MW, Brown VE, et al. · Cell Reports 2014What this evidence cannot show
General methodological point, not specific to this deposit.
Evidence that argues against this (1)
- Automated ChIP-Rx pipelines now offer several normalisation-factor definitions, and the spread between them is a partial substitute for a replicate-derived interval - here ChIP-Rx and the depositors' fraction-based factor differ by up to 1.59x, which bounds the technical component of the scale but not the biological one. Bressan D et al. 2024
the measured 20-fold spread is why no floor may be imported into this deposit
what survives the design: effect sizes two orders of magnitude above any plausible floor
the project H3K79me2 floor is a floor on shares and cannot be carried to an absolute scale
the six replicate libraries are what the prohibition list asks for