Oncofusion

start / graph / ana-the-gse149183-sitepro-columns-are-raw-counts-not-spike-in-scaled

Analysisestablished confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.Q138-absolute-h3k79me2

The GSE149183 per-gene columns are RAW, not spike-in scaled: every one of 1,040,292 values in both files sits exactly on a 0.04 grid, which no per-library scale factor survives - so two entries in this graph called an unscaled number absolute, and the true collapse is 37-91x, not 5.8x

also called the columns were never scaled · the 0.04 grid test · +_dm6_spike-in means alongside, not applied

A public data file's name promised that a fruit-fly yardstick had already been applied to its numbers. A one-line arithmetic check shows it had not: every number is an exact multiple of 0.04, which is what raw counts look like and what scaled counts never look like. Applying the yardstick properly makes the measured drug effect about ten times larger than this project had recorded.

The same thing, for a scientist

**Statement.** Two entries in this graph assert that the per-gene sitepro columns of GSE149183 are already dm6 spike-in scaled, on the strength of the GEO processing line *"Global differences in fly-human chromatin ratios were found, so H3K79me2 TSS signals were normalized accordingly"* and of the `+_dm6_spike-in` in the filename. **Both are wrong.** **The test.** A sitepro average of integer read counts over a fixed number of bins is quantised. Every value in every one of the nine H3K79me2 columns, in both files - 9 x 57,794 x 2 = 1,040,292 values - is an exact multiple of **0.04** (integer/25). Dividing by a per-library factor such as 9.912811459 would destroy that grid instantly. **Off-grid values: zero, in every column of both files, with one exception - `MT-TT_17888`**, a chrM tRNA whose 4 kb window runs off the end of the chromosome so the average is taken over fewer bins. That single row is named in the output rather than dropped. **What the filename actually means.** `+_dm6_spike-in` says the dm6 counts are supplied **alongside** the signal - and they are, in the right-hand columns of the same sheet. The GEO processing line describes what was done for the paper's figures, not for this table. **What it costs.** The absolute scale has to be applied by the reader, and it is one constant per contrast: `K = dm6_reads(DMSO)/dm6_reads(treated)`. With K applied, the true global collapse of H3K79me2 at TSS windows by day 9 is **37x to 91x**, not the **5.8x** cycle 110 recorded - and cycle 110's number was additionally computed on a table that is two-thirds MOLM13.

Go deeper
The two scale-factor definitions, and why they agree

Because the deposited signal is a raw count, the ChIP-Rx convention applies directly: absolute = raw / dm6_reads, so K_chiprx = dm6(DMSO)/dm6(treated). The depositors instead report % fly reads/total reads and its ratio to matched DMSO, which corresponds to K_dep = (human(DMSO)/human(treated)) / flyratio applied to depth-normalised signal. The two are algebraically identical only when the fly fraction is small, and here it reaches 34%.

Measured across all twelve contrasts, the two definitions disagree by 1.04x to 1.59x, worst case MOLM13 EPZ d9. The pre-registration declared that magnitudes would become ranges if any contrast exceeded 2x; none does, so magnitudes are reported as points with the range stated. The recomputed % fly ratios reproduce the depositors' own Ratio to DMSO control column to six decimal places, which is an independent check that the spike-in table was parsed correctly. 1

The identity this adds, and its dual

This thread already owns share(s) = absolute(s)/G. The dual is:

absolute_ratio(g) = raw_ratio(g) x K, with K one constant per contrast.

So a normalisation is a constant multiplier. It cannot change the shape, which means every rank, correlation and permutation p-value in this beat is a RELATIVE statement that would hold on the raw table. What K buys, and the only thing it buys, is where the line between gained and lost sits - and that is exactly the thing no other chromatin dataset in this project has.

This is why calling an unscaled ratio 'absolute' is not a small error: it puts the zero line in the wrong place, and the sign of every gain/loss call follows the zero line. 1

Which entries this corrects, and how
  • ana-the-spike-in-ruler-is-deposited-in-plain-text-and-the-global-change-is-5-to-23-fold - its headline 4.9x-23.4x is computed from the dm6 ratios alone and stands. Its statement that the deposited per-gene columns are spike-in normalised does not.
  • analyses/what-a-share-does-to-a-known-collapse/results.json and the entries drawing on it - the field global_absolute_fold_down there is a raw count-sum ratio, neither a share nor an absolute quantity. Its share arm and its identity result (share = absolute/G, max deviation 0.002) are unaffected, because a share is invariant to any per-sample constant. Its absolute arm is short by the factor K.

The practical rule: a filename is not a processing record, and a processing line describes the figures, not necessarily the table. Check the arithmetic of the numbers themselves. 1

Facts and figures
scriptanalyses/absolute-h3k79me2-in-the-index-fusion/run.py
statusrun
result summary0 off-grid values in 1,040,292 deposited values except the single row MT-TT_17888; K ranges 0.067 to 0.229 across twelve contrasts; the two scale-factor conventions disagree by at most 1.59x.
grid0.04
values tested1040292
worst convention disagreement fold1.59
identityabsolute_ratio(g) = raw_ratio(g) x K, K constant per contrast
inputs
[
  "GSE149183 series supplementary sitepro tables"
]
off grid rows
[
  "MT-TT_17888"
]
K chiprx range
[
  0.06711,
  0.22864
]
How this could bite us back
  • A pharmacodynamic target for a DOT1L inhibitor set from the uncorrected table would be set about ten-fold too lenient: 5.8x of global H3K79me2 removed rather than 37-91x.
  • Conversely, if the fly spike-in is itself IP'd less efficiently once human H3K79me2 collapses, the 37-91x is an overestimate and the true figure lies between it and the raw 5.8x. Nothing in this deposit resolves that, which is why both conventions are carried.
Still open
  1. Does any other deposit reused by this project carry a '+spike-in' filename whose columns are likewise unscaled? The grid test is one line and should be run on every quantitative table this graph imports.
  2. The depositors' %-fly convention and ChIP-Rx diverge as the fly fraction grows; at 34% fly they differ by 1.5x. Which is right depends on whether the fly chromatin is IP'd with the same efficiency as human chromatin at collapsed H3K79me2 - and nothing in this deposit answers that.
  3. Is there a public dataset with BOTH spike-in counts and replicates for H3K79me2 under a DOT1L inhibitor? That would put an error bar on the 37-91x figure, which currently has none.
Every source behind this entry (2)
  1. E1established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.dataset

    Every value in every H3K79me2 column of both GSE149183 sitepro tables is an exact multiple of 0.04, except one chrM tRNA row, which establishes that no per-library scale factor was applied to the deposited columns.

    GSE149183 series supplementary files: MOLM13 and RS4_11 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

    The test is negative evidence of a transformation: it shows no factor was applied that would break the quantisation. A scale factor of exactly 1, or one that is an exact multiple of 25 in the denominator, would be invisible to it - but the deposited ratios (9.312..., 19.274..., 23.406...) are not of that form, so the exclusion is complete for this deposit. The single off-grid row, MT-TT_17888, is named rather than dropped.

    Evidence that argues against this (1)
    • The GEO processing line explicitly says the H3K79me2 TSS signals WERE normalized to the fly-human ratio. Taken at face value it contradicts this result. The reconciliation is that the line describes the paper's analysis, not the deposited summary sheet; the arithmetic of the deposited numbers is the harder evidence and it is unambiguous. NCBI GEO 2020
  2. E2established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.paper

    GSE149183's H3K79me2 samples carry Drosophila spike-in chromatin and were additionally aligned to dm6, and the per-library human and dm6 aligned-read counts are deposited in the same supplementary sheets.

    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 2020
    PMID 32575123DOI
    What this evidence cannot show

    PMID and DOI fetched from NCBI eutils on 2026-09-23. The spike-in counts are the depositors' own; this beat did not re-align anything.

    Evidence that argues against this (1)
    • A spike-in ratio is itself a single measurement per library with no replicate, so the absolute scale inherits an unquantified error. This is why both scale-factor conventions are computed and both reported. Orlando DA et al. 2014
Connected to