Oncofusion

start / graph / ana-q181-the-mark-regex-has-no-token-for-enhancer-nucleosome-or-accessibility

Analysismoderate confidence 3 of 5: One strong study, or several smaller consistent ones.Q181-epigenetics-dimension

Chromatin accessibility is named in the PI's own dimension and the detector has no token for it - nor for enhancer, nucleosome, Polycomb or remodelling - so 121 of the 195 entries scored as having nothing do name a chromatin concept

also called q181 vocabulary gap · _EPI_MARK coverage

The tool looks for the names of specific chemical tags on histones, like H3K4me3. It has no word for 'enhancer', 'nucleosome' or 'open chromatin' - even though the project's own definition of this topic includes chromatin accessibility. So entries that describe a protein's chromatin biology without naming a tag score as empty.

The same thing, for a scientist

_EPI_MARK enumerates specific histone modifications plus 5mC/5hmC, CpG island, DNA methylation and bivalent. The dispatch definition of this dimension explicitly lists chromatin accessibility alongside the marks, and there is no token for it - nor for enhancer, super-enhancer, nucleosome, Polycomb/PRC, heterochromatin, remodelling, or the assay names ChIP-seq and CUT&RUN. Measured over the 195 entries the detector scores as having neither a field nor a mark, 121 (62.1%) name at least one such concept. The largest single contributor is bare 'chromatin' (97 entries), which is also the least reliable: hand-reading 25 of the 121 found 12 substantive statements about the subject, 5 explicit NOT-APPLICABLE or bounded-absence statements, 7 method or paper-title mentions, and 1 genuine gap. The finding is therefore two-sided: the vocabulary IS too narrow to describe the dimension as the PI defined it, and the obvious widening - adding bare 'chromatin' - would import exactly the noise the detector's authors deliberately excluded 'acetylation' and 'methylation' to avoid.

Go deeper
What is in the regex and what the dimension says

_EPI_MARK covers H3K4/9/14/18/23/27/36/56/79, H4K5/8/12/16, H2A.Z, H2AK119, H2BK120, the PRMT arginine marks, 5mC, 5hmC, CpG island, DNA methylation and bivalent 1. That is a good and deliberately curated list of marks.

The dimension it is measuring, as stated in the dispatch for this beat, is "H3K4me1/2/3, H3K79me2/3, H3K27ac, H3K27me3, H3K9me1/2/3, CpG methylation, chromatin accessibility". The last item has no token. Neither do the structural and machinery concepts that carry chromatin meaning without naming a residue.

121 of 195, and the honest breakdown of what that is worth

Over the 195 entries with no field and no mark 1:

conceptentries% of 195
chromatin (bare)9749.7%
enhancer / super-enhancer2512.8%
histone (bare)2311.8%
remodelling2010.3%
Polycomb / PRC1 / PRC2157.7%
ChIP-seq / CUT&RUN / CUT&Tag136.7%
heterochromatin / euchromatin94.6%
nucleosome84.1%
chromatin accessibility / ATAC / DNase HS63.1%
CpG / hypermethylation21.0%
any of the above12162.1%

Drop bare chromatin and the stratum collapses to at most 24 entries on the specific terms. The hand-read says that is closer to the truth than the 121: of 25 sampled, only 12 carried a substantive statement about the subject 2.

What the hand-read actually found, including the cases that are neither coverage nor gap

The 25-entry sample separated into four kinds, and two of them are not on the coverage/gap axis at all 2:

  • Substantive (12) -- gene-smarcc1 "one of the two scaffold proteins that hold every version of the SWI/SNF nucleosome-remodelling machine together"; gene-smarcd1 BAF60a occupancy at the E1-E5 Myc enhancer cluster in MLL-AF9 RN2 cells; gene-ell3 "marks enhancers and recruits SEC rather than sitting inside it"; gene-ube2t reaching the chromatin fraction independently of the FA core complex; gene-atm phosphorylating ENL and switching it from elongation to PRC1-mediated repression; gene-asf1a, gene-dpf2, gene-chek1, gene-zc3h18, gene-zeb2, gene-tert, gene-brd2.
  • Explicit NOT-APPLICABLE (4) -- gene-golga2 "nowhere in the nucleus"; gene-nup214 "the FG repeats are not a transcription-factor domain and not a chromatin reader"; gene-traf2, gene-ap2b1. These are correctly annotated. They are not gaps.
  • Explicit bounded absence (1) -- gene-irx1: "Entirely transcriptional: no chromatin assay of any kind". Also correctly annotated, and a different state again: the assay has not been done, as against there being nothing to assay.
  • Method or title mention (7) -- gene-foxm1's hit describes the source study's design; gene-gpkow's compares BioID with chromatin purification; gene-ap2b1's matched "cell-specific remodeling of the human interactome", which is not chromatin remodelling. My own concept regex produced this false-positive class, and it is the same failure the detector's authors avoided by excluding generic words.

The five NOT-APPLICABLE and bounded-absence cases are the important ones. A metric with two states cannot express either, which is the subject of hyp-q181-a-two-state-coverage-metric-cannot-express-not-applicable.

Proposed fix, NOT APPLIED

scripts/check-pi-dimensions.py is unmodified. The recommendation is deliberately narrow:

  • Add only the specific, low-ambiguity terms: chromatin accessib, ATAC-seq, DNase hypersensit, nucleosom, super-enhancer, heterochromatin. Together these touch ~24 entries and carry little noise.
  • Do NOT add bare chromatin, bare histone, bare enhancer or remodel. They account for the bulk of the 121 and for every false positive in the hand-read. Adding them would raise the printed number by ~18 points while making it mean less, which is the failure this beat exists to avoid.
  • If a fuller accounting is wanted, print the concept counts as a separate, clearly-labelled generous line, the way the reporter already prints its excluded topic-link and mechanism-link counts for localization.
Facts and figures
scriptanalyses/q181-is-the-epigenetics-gap-real/{vocab,sample121}.py
statusrun
generated2026-09-24
tool modifiedfalse
proposed fix appliedfalse
measurements
sampled concept-stratum e…5.000sampled concept-stratum entries that are correctly-annotated NOT-APPLICABLE or bounded absence: 5entries naming chromatin …6.000entries naming chromatin accessibility, ATAC-seq or DNase hypersensitivity: 6entries naming a chromati…121.0entries naming a chromatin concept the mark regex has no token for: 121
recommended additions
[
  "chromatin accessib",
  "ATAC-seq",
  "DNase hypersensit",
  "nucleosom",
  "super-enhancer",
  "heterochromatin"
]
recommended against
[
  "chromatin (bare)",
  "histone (bare)",
  "enhancer (bare)",
  "remodel"
]
measurements — the numbers behind the chart
[
  {
    "quantity": "entries naming a chromatin concept the mark regex has no token for",
    "value": 121,
    "units": "entries",
    "n": 195,
    "subject": "gene/protein entries with no epigenetics-named field and no specific mark token",
    "universe": "541 gene/protein entries at commit ac2aeab; 195 score as having neither",
    "null": null,
    "note": "62.1% of the stratum; falls to at most 24 if bare 'chromatin' is excluded"
  },
  {
    "quantity": "entries naming chromatin accessibility, ATAC-seq or DNase hypersensitivity",
    "value": 6,
    "units": "entries",
    "n": 195,
    "subject": "the one concept the PI's own dimension statement names explicitly and the regex lacks entirely",
    "universe": "same 195-entry stratum",
    "null": null,
    "note": "3.1% - small, but it is the dimension's own vocabulary scoring zero by construction"
  },
  {
    "quantity": "sampled concept-stratum entries that are correctly-annotated NOT-APPLICABLE or bounded absence",
    "value": 5,
    "units": "of 25 sampled",
    "n": 25,
    "subject": "entries a two-state metric must score as uncovered although nothing is missing from them",
    "universe": "seed-181 sample of 25 from the 121-entry concept stratum",
    "null": null,
    "note": "20%, Wilson 8.9-39.1; scales to 11-47 of the 121"
  }
]
Still open
  1. How many of the 97 bare-'chromatin' hits are self-referential - the word appearing only inside the protein's own family name, as in 'histone deacetylase' for gene-hdac2? That number sets how badly the 121 is inflated and is a one-line check nobody has run.
  2. Should chromatin accessibility be part of this dimension's metric at all, given it is a property of a locus rather than of a protein? The PI's statement lists it; the detector measures per-protein entries; those two may not be compatible.
  3. Only the first matched sentence per entry was read. How many of the 7 scored as method mentions carry a substantive statement further down?
Every source behind this entry (2)
  1. E1established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.dataset

    check-pi-dimensions.py's _EPI_MARK regex contains no token for chromatin accessibility, ATAC-seq, DNase hypersensitivity, enhancer, super-enhancer, nucleosome, Polycomb, PRC1, PRC2, heterochromatin or remodelling; of the 195 gene/protein entries scored as having neither an epigenetics-named field nor any mark token, 121 (62.1%) name at least one of these concepts, with bare 'chromatin' in 97, enhancer in 25, remodelling in 20, Polycomb/PRC in 15, heterochromatin in 9, nucleosome in 8 and chromatin accessibility or ATAC in 6.

    What this evidence cannot show

    expected_absent_terms here names tokens that must NOT appear in the _EPI_MARK source line for the claim to hold - the claim is about what the regex lacks, so its support is an absence in the cited source. The concept term list is mine; a chromatin concept expressed in wording neither regex anticipated is missed by this audit as well.

    Evidence that argues against this (1)
    • The detector's narrowness is defensible and its source says why: 'the generic words acetylation and methylation are deliberately NOT added, because they would match any passing mention and turn the measure into noise.' Bare 'chromatin' is 97 of my 121 and is exactly such a word. On the strongest reading of the counter-argument this analysis has found a 24-entry vocabulary gap, not a 121-entry one, and the headline number should be read as 24.
  2. E2moderate confidence 3 of 5: One strong study, or several smaller consistent ones.dataset

    In a seed-181 sample of 25 of the 121 concept-naming entries, 12 carry a substantive chromatin statement about the entry's own subject, 4 carry an explicit not-a-chromatin-protein category statement (gene-golga2 'nowhere in the nucleus', gene-nup214 'not a chromatin reader', gene-traf2, gene-ap2b1), 1 carries an explicit bounded absence (gene-irx1 'Entirely transcriptional: no chromatin assay of any kind'), 7 are method or paper-title mentions, and 1 (gene-hdac2) is a genuine content gap.

    What this evidence cannot show

    n=25 against a stratum of 121; the 12/25 substantive rate has a Wilson 95% interval of 30.0-66.5%, which is too wide to distinguish 'a third' from 'two thirds'. One scorer, who also designed the categories. Only the FIRST matched sentence per entry was printed and read, so an entry whose substantive statement sits later in the file could be scored as a method mention on the strength of an earlier incidental one - an error that runs against the substantive count.

    Evidence that argues against this (1)
    • gene-hdac2 was scored as the sample's one genuine gap on the grounds that no chromatin sentence was found in it at all, yet it entered the 121 stratum in the first place - meaning it matched a concept term, almost certainly bare 'histone' inside the expansion of 'histone deacetylase' in its own name. If a protein's family name alone can put it in the concept stratum, the stratum is contaminated by self-reference and 121 is an overcount by an amount this analysis did not measure.
Connected to