Oncofusion

start / graph / concept-q124-a-within-sample-share-in-a-top-bin-set-is-anti-correlated-with-library-complexity

Explainermoderate confidence 3 of 5: One strong study, or several smaller consistent ones.Q124-the-timecourse-and-the-pcgf1-knockout

Across 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

also called the enrichment confound · a top-set share is a signal-to-noise statistic · why the timecourse's minimum sits at its worst library

This project compares conditions by asking what fraction of a sample's signal sits in its strongest regions - a trick that avoids needing an external yardstick. But a messier experiment spreads its signal out, which lowers that fraction without anything biological happening. In one dataset the ordering of four timepoints matched the ordering of how messy they were, exactly.

The same thing, for a scientist

**This corpus's standard cross-condition instrument is the within-sample share ratio**, `R_T(S) = sum over S of share_drug / sum over S of share_DMSO`, adopted because GSE260742 deposits no spike-in and a share cancels the unknown per-sample scale factor exactly. **It does not cancel a difference in ChIP enrichment, and this beat measured how large that can be** [E1]. Over the four VTP-50469 timepoints of GSE260742's murine H3K79me2 kinetic series: | | DMSO | 4 h | 8 h | 16 h | 24 h | |---|---:|---:|---:|---:|---:| | non-zero bins in the universe | 501,520 | 789,737 | 973,715 | **1,107,059** | 986,555 | | enrichment index `C` (own coverage in own top 0.1%) | 0.0972 | 0.0849 | 0.0838 | **0.0653** | 0.0838 | | **`V` at the fusion's own sites** | - | **0.8996** | **0.8835** | **0.6181** | **0.6790** | **Spearman(`V`, non-zero count) over the four timepoints is exactly -1.000**, and a null built from `C` alone - what `V` would be if the drug did nothing and only the libraries differed - spans **0.6724 to 0.8731** against `V`'s observed **0.6181 to 0.8996**. **The observed range is not wider than the null's.** **The mechanism is not sequencing depth, it is signal-to-noise.** A share is `x_S / total`, and if extra reads were distributed like the existing ones the share would be invariant. What moves it is extra reads distributed differently - background - which is what a lower-efficiency ChIP produces. **So a top-set share ratio is, in part, a ratio of two samples' enrichment.** **And there is no within-sample fix, which is the hard part.** Three partial mitigations were tried and each has a limit. **Common support** - restricting to bins non-zero in every track - removes the fill-in bins from both numerator and denominator and shrank the anomalous step by 80%, but does not correct low enrichment inside the retained bins. **Rank statistics** are invariant to a monotone rescaling but not to background filling in beneath the set. **Subtracting `C`** is circular: a genuine redistribution lowers `C` too, which the SGC0946 contrasts prove by giving residuals of only -0.004 to -0.025 for a drug whose share-ratio effect is 0.09 to 0.15. **What the confound does NOT do, and this is the part that keeps the corpus's other numbers standing.** In the four 48 h VTP contrasts the drug and vehicle tracks differ in non-zero count by only **0.986 to 1.110 fold**, against 2.208 across the timecourse. **The confound is large where the libraries are unmatched and small where they are matched**, and it is a reason to report the non-zero count beside every share ratio rather than a reason to withdraw one.

Go deeper
The reporting rule this implies

Print the non-zero bin count and the enrichment index beside every within-sample share ratio, and print the drug-to-vehicle ratio of the non-zero counts. This beat's results.json carries all three for all fourteen tracks. The numbers that need scepticism are the ones where that ratio is far from 1:

contrastnon-zero count, drug / vehiclereading
the four 48 h VTP contrasts0.986 - 1.110matched; the confound is small
the four 72 h SGC contrasts1.072 - 1.197matched; and the drug's effect is 6-10x larger than the mismatch could produce
the timecourse, drug against its 24 h vehicle1.575 - 2.208unmatched; every ratio in that arm is affected

That table is the practical content of this entry. It is one column added to a diagnostics block this project already computes, and it separates the arms whose numbers can be read from the arm whose cannot.

Why a spike-in is the only real fix, and why it is absent

A spike-in provides an external yardstick, so a drop in enrichment shows up as a drop in the sample's own signal relative to the spike rather than as a redistribution. Q117 verified zero hits for the entire spike-in vocabulary across all 419,824 characters of GSE260742's 158 sample records, and no scale factor is deposited. The sibling series GSE127507 is worse in an instructive way: all 26 of its records state a Drosophila spike-in was run and measured and deliberately not used, and no scale factor is deposited there either.

The published literature has demonstrated the consequence directly for a different mark: standard ChIP-seq normalisation fails to detect the genome-wide H3K27me3 decrease caused by an EZH2 inhibitor, and the decrease appears only with a Drosophila chromatin spike-in normalisation. So this is not a hypothetical limitation of the instrument, it is a documented failure mode of the instrument class, and the reason every measurement entry in this beat says 'redistribution' rather than 'loss'.

Facts and figures
instrumentwithin-sample share ratio over a fixed top-0.1% bin set, 2 kb bins, no spike-in available
confoundper-sample ChIP enrichment / signal-to-noise, which a within-sample share does not cancel
diagnostic to reportthe non-zero bin count and the enrichment index C for every track, and the drug-to-vehicle ratio of the non-zero counts
the only real fixan external spike-in, which GSE260742 does not deposit (zero spike-in vocabulary in 419,824 characters of sample records) and which GSE127507 measured and deliberately did not use
absence classmeasured-not-analysed
absence class noteThe diagnostic that reveals this confound - non-zero bin count per track - has been computed by every beat in this family since Q117 and printed in their track_diagnostics blocks, and was never compared against the share ratios beside it. Closing the wider question needs no download: the counts are in Q117's, Q120's and Q122's committed results.json for every track they binned, and the correlation with their reported share ratios can be computed from those files alone.
absence class assigned2026-09-23 (Q124)
measured exposure by arm
{
  "timecourse_drug_over_vehicle_nonzero_ratio": [
    1.575,
    2.208
  ],
  "48h_VTP_contrasts": [
    0.986,
    1.11
  ],
  "72h_SGC_contrasts": [
    1.072,
    1.197
  ]
}
partial mitigations and their limits
[
  "common support (bins non-zero in every track): removes fill-in bins from numerator and denominator, shrank the anomalous step by 80%, does not correct low enrichment inside retained bins",
  "rank statistics: invariant to a monotone rescaling of a track, not invariant to background filling in beneath the set",
  "subtracting the enrichment index: circular, because a genuine redistribution lowers the index too"
]
How this could bite us back
  • If a top-set share ratio is partly a ratio of two samples' ChIP enrichment, then any conclusion of the form 'the drug moves the mark' drawn from a pair with mismatched libraries is uninterpretable, and this project has built several conclusions on unreplicated pairs. The saving grace here is measurable and is reported: in the four 48 h contrasts the libraries are matched to within 11%.
  • It also cuts at the fusion-free K562 result that cycle 82 called the strongest threat to the drug's mechanism story. That arm's numbers were computed by the same instrument from the same series, and nobody has checked its drug-to-vehicle non-zero count ratio. The check is a file lookup in Q122's committed output and this beat did not run it because K562 was outside its two arms.
Still open
  1. Across every share ratio this project has reported from Q117, Q118, Q120 and Q122, how strongly does the value correlate with the drug-to-vehicle non-zero bin count ratio? Every input is in their committed results.json and no download is needed. If the correlation is strong outside the timecourse arm, more numbers than this one need qualifying.
  2. Is there a within-sample statistic that is genuinely enrichment-invariant? A quantile-normalised per-bin ratio, or a share computed over a fixed-size high-confidence set defined by the input rather than by a ChIP track, are candidates and neither was tried here.
  3. GSE127507's records state a Drosophila spike-in WAS run and measured and deliberately not used. Are the spike-in read counts recoverable from the deposited SRA runs by aligning to dm6? That would retro-fit an external yardstick to the human arms of this project's most-used series, and it is compute rather than wet work.
  4. Why does the timecourse's 24 h vehicle have 501,520 non-zero bins when every drug track in the same experiment has 790,000 to 1,107,000? A vehicle that is the least complex library in its own arm is the worst possible denominator and nothing in the deposit explains it.
Every source behind this entry (1)
  1. E1strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.dataset

    In GSE260742's murine H3K79me2 VTP-50469 kinetic series, 2 kb bins over a 1,227,708-bin universe, the five tracks carry 501,520 (DMSO 24 h), 789,737 (4 h), 973,715 (8 h), 1,107,059 (16 h) and 986,555 (24 h) non-zero bins - a 2.208-fold spread - with within-sample enrichment indices (own coverage inside own top 0.1%) of 0.0972, 0.0849, 0.0838, 0.0653 and 0.0838; the within-sample share ratio at the FLAG-defined top-0.1% set is 0.8996, 0.8835, 0.6181 and 0.6790, whose rank order over the four timepoints is the exact inverse of the non-zero count (Spearman -1.000), and a null of the form C_drug/C_DMSO spans 0.6724-0.8731 against the observed 0.6181-0.8996. In the four 48 h VTP-50469 contrasts of the same series the drug-to-vehicle non-zero count ratio is 0.986-1.110 and in the four 72 h SGC0946 contrasts it is 1.072-1.197.

    What this evidence cannot show

    THE ENRICHMENT INDEX C IS NOT A PURELY TECHNICAL QUANTITY: a genuine redistribution lowers it too, demonstrated in this beat's own data by the SGC0946 contrasts whose residuals obs-minus-null are only -0.0043 to -0.0248 for a drug whose share-ratio effect is 0.09-0.15. So C_drug/C_DMSO is NOT a clean null and the statement 'the observed range is not wider than the null's' is a confound ALARM and not a demonstration that the timecourse's variation is technical. The 16 h minimum survives a common-support restriction and two rank statistics, which argues the confound explains the recovery's size and not its existence. n = 4 timepoints, so the Spearman of -1.000 has a one-sided permutation probability of 1/24 = 0.042 under a random ordering and is suggestive rather than decisive. The non-zero bin count is a property of a 2 kb binning of a deposited .tdf and is a proxy for library complexity, not a measurement of it - no FASTQ, duplicate rate or library-size metric was examined, and the deposit provides none.

Connected to