start / graph / concept-q117-a-within-sample-share-is-the-cross-condition-comparison-that-survives-a-missing-scale-factor

ExplainerestablishedQ117-gained-fusion-occupancy

When a deposit has no recoverable spike-in, two comparisons still work and one does not: a within-sample rank and a within-sample share both cancel the unknown scale factor, a fold change does not - and the share is the one that also has no ceiling

also called Q117 share-not-fold-change rule · what to do when the scale factor is gone

If you cannot recover how two experiments were put on the same scale, you can still ask what fraction of each experiment's own signal sits in a region of interest, and compare those fractions. That comparison is immune to the missing information; a straight ratio of signal heights is not.

The same thing, for a scientist

Let the true occupancy at bin *i* in sample *s* be `x_si`, and let the deposited track be `k_s * x_si` where `k_s` is an unknown, unrecoverable per-sample constant (sequencing depth, IP efficiency, cell number, and whatever a missing spike-in would have corrected). | comparison | carries `k_s`? | ceiling? | usable | |---|---|---|---| | **fold change** `k_v x_vi / (k_d x_di)` | **yes**, as `k_v/k_d` | no | **no** | | **within-sample rank / percentile** | no - ranking is invariant to multiplication by a positive constant | **yes**, saturates at 100 | yes, with care | | **within-sample share** `sum_H(x) / sum_all(x)` | no - `k_s` cancels | **no** | **yes** | **The share is the better of the two survivors and the reason is the ceiling.** In Q117, KDM2B's baseline percentile at the fusion's strongest sites was **99.685** - a rank statistic could not have detected a gain there however large, exactly as a gene depleted in 100% of DepMap lines cannot show selectivity. The share statistic has no such limit and gave a usable answer (ratio 1.009). **Where both were computed, they agreed**, which is the check that licensed reporting the negative rather than reporting the test as uninformative [E1]. **Three things the share does not fix**, stated because the temptation is to treat it as a rescue: 1. **Shares sum to one, so there is no invariant reference region.** If a factor is stripped from its strongest sites, the signal has to land somewhere, and every other region's share rises whether or not anything happened there. In Q117 the housekeeping loci *Gapdh* and *Actb* both fell to about half their DMSO share under a menin inhibitor, and that floor - not 1.0 - is what other loci must be read against [E1]. 2. **It cannot separate 'the protein left the DNA' from 'the antibody pulled down more background'.** Both reduce the share identically. Only a spike-in or an orthogonal readout breaks that. 3. **It is still a single-sample statistic.** With n = 1 per condition it carries no spread, and this project has measured the same 147 bp window at percentiles 51.8, 69.1, 90.8 and 96.3 across four samples of one antibody in one line. **The practical rule: when the scale factor is unrecoverable, report (a) the within-sample share ratio, (b) the enrichment over the track's own genome median, (c) the total-coverage ratio so the reader can see how much depth differs, and (d) a housekeeping or matched-region floor. Never a fold change.**

Go deeper

The enrichment-over-own-median form, which is often the clearest

A third scale-invariant quantity is the region's signal divided by the same track's genome-wide median bin. It is a ratio of two numbers from one sample, so k_s cancels, and unlike a share it does not depend on the size of the region set. Q117's headline used it: the fusion's top-0.1% bins sit at 15.9x the track median in DMSO, 3.0x under a menin inhibitor and 6.1x under a DOT1L inhibitor. Those three numbers are directly comparable and none of them required a scale factor.

Its weakness is the mirror of the share's: it is sensitive to what happens to the median. In Q117 the median moved from 372 to 375 to 400 coverage-bases across the three conditions, which is small enough to ignore and was checked rather than assumed.

When to refuse rather than substitute

A within-sample statistic answers how concentrated is this factor at these sites, relative to its own distribution. It cannot answer how many molecules are on the chromatin. If the question is genuinely about absolute amount - global levels of a histone mark, total chromatin-bound protein, anything where every site could move the same way - no within-sample statistic will do and the honest answer is that the deposit cannot support the question. That is the situation for the global H2AK119ub increase in the depositing paper of Q117's dataset, which is why the authors ran a spike-in for exactly that one panel and for nothing else.

Facts and figures

domain	research method
rule	With an unrecoverable per-sample scale factor, report the within-sample share ratio, the enrichment over the track's own genome median, the total-coverage ratio and a matched-region floor. Never a fold change. Prefer the share over the rank because the rank has a ceiling.

worked numbers

{
  "fusion_enrichment_over_own_genome_median": {
    "DMSO": 15.9,
    "VTP50469": 3,
    "SGC0946": 6.1
  },
  "track_genome_median_coverage_bases": {
    "DMSO": 372,
    "VTP50469": 375,
    "SGC0946": 400
  },
  "housekeeping_floor_share_ratio_VTP_over_DMSO": {
    "Gapdh": 0.508,
    "Actb": 0.571
  },
  "ceiling_example": {
    "factor": "KDM2B",
    "baseline_percentile_at_fusion_sites": 99.685,
    "rank_statistic": "uninformative by construction",
    "share_statistic": 1.009
  }
}

How this could bite us back

The rule makes negative results easier to defend and positive results harder to state, which is the right asymmetry for this project and the wrong one for anybody trying to publish an effect. An analyst under pressure to report a magnitude will reach for the fold change, and the fold change is available, computable and wrong.

Still open

What fraction of the occupancy deposits this project cites ship processed tracks with no recoverable scale factor? Nobody has counted, and the answer decides how much of the archive is only reachable by within-sample statistics.
Is there a defensible matched-region floor that does not rely on housekeeping genes - for example the distribution of the share ratio over expression-matched, factor-unbound promoters? That would replace the weakest step in the rule.

Every source behind this entry (2)

E1established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.dataset
In a deposit with no recoverable spike-in, the within-sample share and the enrichment over the track's own genome median gave consistent answers where a rank statistic was at its ceiling (KDM2B baseline percentile 99.685, share ratio 1.009), and the housekeeping loci Gapdh and Actb fell to 0.508 and 0.571 of their DMSO share, defining the floor rather than sitting at 1.0.
NCBI GEO GSE260742, murine MLLAF9 FLAG / KDM2B / BCOR / RING1B ChIP-seq under DMSO, SGC0946 and VTP50469; analyses/q117-gained-fusion-occupancy/results.json and locus-table.json · NCBI Gene Expression Omnibus 2026
What this evidence cannot show
One dataset, n = 1 per condition, 2 kb bins, mm10. The agreement between the share and the rank statistic for KDM2B is a consistency check on two statistics applied to the same single sample pair, not a replication. The Gapdh and Actb floor rests on two genes.
Evidence that argues against this (1)
- A share ratio can be driven entirely by what happens outside the region of interest. If a drug creates a large artefactual pileup somewhere, every other region's share falls and the analyst reads it as loss. Q117 hit exactly this: a single blacklisted bin carried up to 18% of some tracks, and the share ratios changed by about 10% when it was removed. A share is only as good as the denominator's cleanliness, which has to be inspected before the statistic is trusted. Amemiya HM et al. 2019
E2established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.dataset
The situation the rule addresses is common rather than exotic: in GSE260742, 152 of 158 samples ship only an unnormalised igvtools .tdf and no sample record mentions a spike-in or a scale factor.
NCBI GEO GSE260742 - per-sample SOFT records, all 158 samples · NCBI Gene Expression Omnibus 2026
What this evidence cannot show
A single series. Whether tdf-only deposits without scale factors are typical of the occupancy archive this project draws on has not been measured, and the claim of commonness rests on one example plus the observation that browser-track deposition is standard practice.
Evidence that argues against this (1)
- Raw reads are usually in SRA, so the constraint is a constraint on re-using PROCESSED files, not on the data. An analyst willing to re-align gets depth control for free and, if the experiment included spike-in chromatin, can recover the scale factors from the reads. The rule here is for the common case where compute or disk makes re-alignment impractical - which is the case this beat was in. NCBI Sequence Read Archive 2026

Connected to

depends on →The premise that GSE260742 'carries a ChIP-Rx spike-in' does not survive reading the deposit: zero of 158 sample records mention a spike-in, the only ChIP-Rx experiment in the paper is an H2AK119ub panel that is not identifiable among the deposited samples, and 152 of 158 samples ship only an unnormalised igvtools .tdf
the rule exists because no scale factor is recoverable for this series
supports →The residual fraction this graph says is absent from the entire literature, measured: FLAG-MLLAF9 enrichment at its own top 1,228 sites falls from 15.9x the genome median to 3.0x under VTP50469 - about 18% residual - and 12.8% of those sites stay in the top 0.1%
the residual-fraction measurement is the worked example of the rule
supports →KDM2B, the PRC1.1 CpG-island recruiter, sits at the 99.7th percentile of its own distribution on the fusion's strongest sites in DMSO and does not gain share under either a menin inhibitor or a DOT1L inhibitor - it holds position while comparable bins drift down
KDM2B at the 99.685th percentile is the worked example of the rank statistic's ceiling
← supportsGSE127507 ran a Drosophila spike-in, measured it, and deliberately did not use it - so the VTP-50469 founding dataset carries no recoverable scale factor either, and the depositors say so in the sample record
a second dataset in which the only valid cross-condition statistic is a within-sample share or rank
← supportsA within-sample share cancels an unknown scale factor but it does NOT make two studies comparable - the same eviction reads as 0.20 over a genome-wide denominator and 0.67 over a promoter-only one, in the same cells on the same day
extends that concept with what a share does NOT buy: comparability across universes, and any sight of a uniform global loss
← refutesAcross four VTP timepoints the within-sample share at a fixed top-0.1% set ranks perfectly inversely to the track's non-zero bin count (Spearman -1.000 over a 2.2-fold spread), so this project's standard instrument is confounded with ChIP enrichment - and the confound cannot be removed from inside one sample
narrows it: a share cancels the per-sample SCALE factor exactly, as that entry says, and does NOT cancel the per-sample ENRICHMENT - measured here at Spearman -1.000 between V and non-zero bin count over four timepoints spanning 2.208-fold
← related toThe first replicated share ratio this corpus has ever had: two title-declared replicates of ONE murine culture span 0.2623 on the same quantity - 4.5x the 0.0584 between-culture spread this corpus uses as its indifference threshold - and they disagree in SIGN about whether the PCGF1 knockout blunts the DOT1L-inhibitor response
puts the first empirical resolution bound on the statistic that concept defines
← related toTwo deposit defects in one beat, both invisible to the check in use: a genome-build verdict read from a file's declared extents names its CONTAINER, and the furthest coordinate carrying signal names its DATA - and a clause that cannot return the other answer has been reported as a passing check in four beats
both are properties of this instrument that decide what its numbers mean
← related toEvery per-sample normalisation is a constant multiplier: it never changes the shape, it only moves the line between 'gained' and 'lost' - which is why 'share', 'raw' and 'absolute' give the same ranking and three different answers to 'did this go up?'
a share survives a missing scale factor precisely because the factor is a constant, and loses the zero line for the same reason