Eleven hand-picked intervals are not a null: the CTCF loss on KDM3B depletion scales monotonically with baseline anchor strength (decile medians 0.78 to 0.62), so a strong anchor compared against gene bodies looks unique and compared against its own strength class is the 58th of 1,625
also called match on the quantity that varies · why the eleven controls could not answer the question · the strength confound, measured
If you want to know whether one place in the genome is special, you have to compare it with places of the same kind and the same strength. Comparing a very strong site against ten ordinary stretches of DNA will always make the strong site look special, because strength itself is what drives the effect you are measuring.
The same thing, for a scientist
`mechanism-kdm3b-depletion-lowers-ctcf-only-at-the-kmt2a-anchor` compared CTCF loss at the *KMT2A* exon-11 anchor (0.30x) against **eleven hand-picked intervals** (0.89-1.21x) and concluded locus specificity. Its own caveat named the defect precisely: *"The control loci are GENE BODIES ... not CTCF anchors. A gene body averages over tens of kilobases of mostly non-anchor sequence, so an anchor-level CTCF change inside one would be heavily diluted."* **The caveat was right and the measurement that closes it also shows the confound is real and quantitative, not merely possible** [E1]. Ranking all 12,443 CTCF anchors called from the same control tracks, **the siKDM3B/control ratio falls monotonically with baseline anchor strength**: | baseline-strength decile | 1 (weakest) | 2 | 3 | 4 | 5 | 6 | 7 | 8 | 9 | 10 (strongest) | |---|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:| | median ratio | 0.779 | 0.751 | 0.731 | 0.731 | 0.714 | 0.694 | 0.672 | 0.653 | 0.640 | **0.615** | **So strength is a confounder with a measured gradient**, and the *KMT2A* anchor sits at the 91.7th percentile of strength. Against non-anchors it is unique; against its own strength class it is rank **58 of 1,625** [E1]. **The generalisable rule.** A control set answers a specificity question only if it is matched on **the quantity that drives the effect**, and the way to find that quantity is to measure the effect across the whole population first and look for a trend. Here the trend was available from the same four files in twenty minutes, and it converts the published comparison from *"one locus moved and eleven did not"* into *"one locus moved by 0.40x and the median of its strength class moved by 0.63x"* - which is a different claim with a different conclusion. **This is rule 84's shape with a twist.** Rule 84 says a proportion is not a result until its null is written down. **A CONTRAST is not a result until its comparator is matched on the covariate**, and the covariate is discoverable only by measuring the population. An unmatched contrast does not merely lack a null: it has one, and it is the wrong one.
Four things the matched design had to get right, each of which could have manufactured the answer
- Call the comparators from the SAME FILES, not from an external peak list. An ENCODE RPE CTCF peak set would have introduced a second antibody, a second pipeline and a second definition of a peak, and baseline strength - the matching variable - would then be measured in files other than the ones the ratio comes from. The whole point of matching is that the covariate and the outcome share a denominator.
- Fix the window width. Anchors called as runs above a threshold have widths that correlate with strength, so measuring strength over the called peak would make strength partly a width statistic. A fixed 1,000 bp window centred on the summit bin - chosen to match the 1,048 bp KMT2A element - removes that.
- Declare the matched sets before looking. Three were pre-declared (+/-10% of strength, nearest 500, strength decile) and all three are reported; they agree at 3.6%, 4.4% and 4.9%. Had only one been declared, a reader could not tell whether it was the one that gave the nicest answer.
- Expect regression to the mean and control for it. Anchors are selected on high control signal and that same signal is the ratio's denominator, so the selected set drifts below 1.0 by construction. Two defences were pre-declared: quote the matched-set MEDIAN beside every rank as the null (0.615-0.633), and re-run with anchor calling on control replicate 1 and the denominator on replicate 2 only. The second one moved the percentile from 3.6% to 31.9%, which is a large effect and is the reason the honest answer here is a bracket rather than a number 1.
The first defence is the cheaper one and it is the one to copy. A rank with no matched median beside it is uninterpretable: rank 58 of 1,625 sounds decisive until the matched median is 0.633, at which point the reader can see that the typical comparator also lost a third of its signal.
Where else this project is exposed
The pattern is: a locus-specificity claim whose comparator set was chosen for convenience. This is not a criticism of the beats that wrote them - a hand-picked comparator is the right first move and is often all the data support - but the exposure is worth naming, because the fix is usually the same twenty minutes.
Candidate exposures in this corpus, none of them checked by this beat:
- Any claim of the form "X changes at the KMT2A BCR and not at MLLT1 / MLLT3 / TCF3 / AFF3". Those four are whole gene bodies spanning 43 kb to 597 kb. For a mark measured at 1 kb bins the dilution argument applies to them as much as to CTCF, and the matched comparator would be intervals of the same width and the same baseline mark level.
- The chr11 percentile ranking of the H3K9me1 response (
ana-q108-chr11-rank-of-the-kmt2a-bcr-h3k9-response, rank 72 of 18,082 chr11 bins). A bin ranking is a genuine null and a good one - and it is not matched on baseline mark level, so if the H3K9me1 response also scales with baseline the same correction would apply. This is directly testable from data already read and is the obvious next check. - Any DepMap or dependency contrast against "other leukaemia lines" where the index arm differs systematically in a covariate - lineage, which this project already documents (rule 83's lineage split), but also expression level or copy number of the gene in question.
The test to apply before believing any specificity claim: name the covariate that could drive the effect, measure the effect across the population, and see whether it has a gradient in that covariate. If it does, the contrast must be matched. If it does not, say so - a flat gradient is itself the licence the unmatched comparison needed.
| trap class | inference trap - a specificity contrast whose comparator set is not matched on the covariate that drives the effect |
|---|---|
| the covariate | baseline CTCF occupancy at the anchor, measured in the same control tracks the ratio comes from |
| the cheapest defence | quote the matched-set median beside the rank. Rank 58 of 1,625 sounds decisive; rank 58 of 1,625 with a matched median of 0.633 does not. |
| relation to rule 84 | rule 84 says a proportion needs its null written down. This is the contrast version: an unmatched contrast does not lack a null, it has the WRONG one, and the covariate that makes it wrong is discoverable only by measuring the effect across the population first. |
| localization | not applicable - this entry is methodological and asserts nothing about a cellular location |
| subnuclear localization | not applicable - a design principle has no subnuclear coordinate and none is claimed |
| cell type | the worked example is hTERT-immortalised human retinal pigment epithelial (RPE) cells; the principle is cell-type independent, and the example's own cell type is the reason its conclusions license nothing about a blood cell |
| cell state | asynchronous proliferating culture 72 h after control or KDM3B siRNA transfection, in the worked example |
| epigenetic readout | CTCF ChIP-over-control occupancy at 200 bp across 12,443 anchors, used to expose a baseline-strength gradient in the response to KDM3B depletion |
| shown for fusion | Not shown for any fusion. The worked example is in unrearranged RPE and the principle is about study design, not fusion biology. |
measured gradient decile medians of ratio
[ 0.779, 0.751, 0.731, 0.731, 0.714, 0.694, 0.672, 0.653, 0.64, 0.615 ]
what changes
{
"against eleven hand-picked non-anchor intervals": "0.30x versus 0.89-1.21x - reads as locus-specific",
"against 1,625 strength-matched anchors": "0.40x versus a matched median of 0.633x, rank 58, 3.57th percentile - reads as the strong end of a continuum",
"under the pre-declared bias control": "0.586x versus a matched median of 0.685x, 31.9th percentile - reads as unremarkable"
}four design requirements
[ "call comparators from the SAME files, so the covariate and the outcome share a denominator", "use a FIXED window width, so strength is not partly a width statistic", "pre-declare every matched set and report all of them", "expect regression to the mean: quote the matched MEDIAN beside every rank, and run a split-replicate version" ]
exposures named and not checked
[ "any claim that a mark changes at the KMT2A BCR and not at the MLLT1, MLLT3, TCF3 or AFF3 gene bodies - the same dilution argument applies", "ana-q108-chr11-rank-of-the-kmt2a-bcr-h3k9-response - a genuine bin-level null, but not matched on baseline mark level; directly testable from data already read", "any dependency-screen contrast where the index arm differs systematically in a covariate such as lineage, expression or copy number" ]
- Applied too enthusiastically this becomes a licence to dismiss every focused experiment, which would be wrong. Most published mechanism work compares a locus of interest against a handful of controls because that is what a ChIP-qPCR can do, and demanding a genome-wide matched null before believing anything would discard almost the entire chromatin literature. The rule is narrower: a claim of SPECIFICITY - the word 'only', or 'and nowhere else' - needs a matched comparator. A claim that something happens at a locus does not.
- For an intervention the direction of the correction matters. Matching made the KMT2A anchor LESS special, which makes any drug acting on this axis less locus-selective and its off-target burden larger. A beat that applied this rule and found the opposite - that a locus was more special than an unmatched comparison suggested - would be strengthening an intervention case on the same reasoning, and should be held to the same standard of pre-declaration.
- Does the H3K9me1 response to KDM3B depletion also scale with baseline H3K9me1 level? If it does, the chr11 bin ranking that places the KMT2A BCR at 72 of 18,082 needs the same matching correction, and the data for the test have already been read - the 1 kb H3K9me1 control and siKDM3B tracks of the same series. This is the single cheapest follow-up in this beat's whole territory.
- Is the strength gradient regression to the mean, biology, or both? Separating them needs an independent-replicate design - call anchors and measure baseline in one pair of replicates and the ratio in another - which requires n >= 4 per arm and therefore new data. Until then the gradient licenses matching and nothing more.
- How many specificity claims in this corpus rest on a comparator set chosen for convenience rather than matched on a covariate? The question is answerable by a sweep for entries whose evidence contrasts one interval or gene against a short hand-picked list, and it has never been run.
- E1strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.dataset
Across 12,443 CTCF anchors called from the two siCT control CTCF tracks of GSE210480 in RPE cells, the median siKDM3B/control occupancy ratio falls monotonically from 0.779 in the weakest baseline-strength decile to 0.615 in the strongest, so baseline anchor strength is a measured confounder of the effect; the KMT2A exon-11 anchor sits at the 91.7th percentile of baseline strength with a ratio of 0.396, which is rank 58 of 1,625 anchors matched within +/-10% of its strength (3.57th percentile, matched median 0.633), rank 22 of the nearest 500 (4.40th), rank 61 of 1,245 in its decile (4.90th) and rank 258 of all 12,443 (2.07th); and a pre-declared split-replicate control that removes the shared-noise component of the selection moves it to the 31.9th percentile of 185 matched anchors.
GEO GSE210480 CTCF bigWigs read genome-wide at 200 bp with pybigtools 0.3.0, anchors called from the control tracks only, three pre-declared matched sets and a split-replicate bias control, 2026-09-22; analyses/q111-ctcf-anchor-matched-ranking/design.py, run.py, results.json and anchors.tsv · Gray ZH, Chakraborty D, Duttweiler RR, Alekbaeva GD, Murphy SE, Chetal K, Ji F, Ferman BI, Honer MA, Wang Z, Myers C, Sun R, Kaniskan HU, Toma MM, Bondarenko EA, Santoro JN, Miranda C, Dillingham ME, Tang R, Gozani O, Jin J, Skorski T, Duy C, Lee H, Sadreyev RI, Whetstine JR · NCBI GEO (data accompanying Cell 186:4528-4545.e18) 2023What this evidence cannot show
The strength gradient is measured in ONE dataset, in ONE cell type, for ONE perturbation, at n = 2 versus n = 2, with no input file and no spike-in. Part of the gradient is expected from regression to the mean: anchors are selected and matched on the same control signal that forms the ratio's denominator, so a downward slope in the strongest deciles is partly statistical. The split-replicate control shows that component is substantial. So 'CTCF loss scales with baseline strength' is asserted as a property of THIS measurement, sufficient to disqualify an unmatched contrast, and NOT as an established biological law - that would need the independent-replicate version, more datasets and an absolute-occupancy readout.
Evidence that argues against this (1)
- If the gradient were purely regression to the mean, then matching on strength would be necessary for a fair rank and the gradient itself would carry no biology - and the conclusion about the eleven hand-picked intervals would be unchanged, because a statistical confounder disqualifies an unmatched contrast exactly as a biological one does. The place the distinction matters is the opposite direction: a real biological gradient would make 'strong anchors are intrinsically more CTCF-labile' a finding, and this dataset cannot separate the two. Reporting the gradient as a reason to match, rather than as a result, is the position these data support. Iqbal Lab oncofusion knowledge graph 2026
- E2established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.database
The entry under test named this defect in its own caveats and open questions before it was measured: its E1 caveat states that the control loci are gene bodies and 1 kb promoter windows rather than CTCF anchors, that an anchor-level change inside a gene body 'would be heavily diluted and could be missed entirely', and that this 'is the weakest part of the design'; its first open question asks whether hundreds of other CTCF anchors lose CTCF to the same degree and calls it 'the one check that decides whether this entry is a real mechanistic correction or an artefact of comparing an anchor with gene bodies'; and its data block records 'an anchor-matched genome-wide comparison ... NOT RUN by this beat'.
mechanism-kdm3b-depletion-lowers-ctcf-only-at-the-kmt2a-anchor, this graph's own entry written cycle 59, read verbatim 2026-09-22 · Iqbal Lab oncofusion knowledge graph 2026What this evidence cannot show
Quoting an entry's own caveat as evidence that the caveat was correct is close to circular, and it is recorded here for a specific reason rather than as support for a biological claim: this project's standing rule is that a caveat naming an uncovered case is a TASK, and this is a worked instance where the task was named, costed as 'one afternoon', left undone for a cycle, and then turned out to reverse the entry's headline. The value is procedural.
Evidence that argues against this (1)
- The entry could be said to have done everything right - it measured what it could, named the defect, stated the refuting experiment, predicted the outcome correctly and set confidence to moderate. On that reading there is no error to correct, only a follow-up to run, and framing this as a refutation overstates the fault. The part that does need changing is narrow and specific: the word 'only' in the title and the sentence placing the locus address on the CTCF arrow, both of which assert more than the eleven intervals could support and both of which a reader will quote without reaching the caveat. Iqbal Lab oncofusion knowledge graph 2026
its eleven comparators are unmatched on the covariate that drives the effect, and the gradient is measured: decile medians 0.78 to 0.62
the design principle the ranking implements, with the four requirements it had to meet
a genuine bin-level null, but not matched on baseline mark level - the same correction may apply and is testable from data already read
the quantity that varies here is call-set size, and it is what the rule decides
here the quantity that varies is promoter coverage