A counted PubMed absence is bounded by ABSTRACTS, not by papers - and the measurement you are looking for is usually a supporting figure
also called the corpus bound · abstract-bounded absence · rule 67's blind spot
Searching PubMed only looks at titles and summaries, not at the figures inside papers. So a search can correctly return zero results while the number you wanted is sitting in Figure 6 of the very first paper on your list.
The same thing, for a scientist
This project has rules for three ways a counted absence goes wrong. **Rule 67**: the query may not be able to return a one - fix it with a positive control through the identical shape. **Rules 70-72**: the *name list* may be wrong - fix it by enumerating spellings, retired symbols, drug names and pathway phrases. **Rule 64**: the *complex form* may be the only reachable phrasing. **This is a fourth, and none of the first three catches it: the CORPUS is wrong.** PubMed indexes titles, abstracts and MeSH headings. It does not index figures, results text, methods or supplementary material. So a measurement that is a **supporting panel rather than a headline** is invisible to every query, no matter how well-built. **Measured, on this beat, 2026-09-20:** | query | records | |---|---:| | `SET-NUP214 AND (H3K79 OR "H3-K79" OR "K79" OR "lysine 79")` | **0** | | `SET-NUP214 AND (H3K79 OR H3K4)` | **0** | | `SET-NUP214 AND DOT1L` (positive control, identical shape) | **1** | | `SET-NUP214 AND histone` | 21 | Every defence fires correctly. The control returns a one, so rule 67 passes. Four spellings of the mark were tried, so rules 70-72 pass. **And the conclusion 'no histone mark has been measured on SET::NUP214' is false.** PMID 18299449 - the paper the DOT1L control returns - contains an anti-dimethyl-H3K79 ChIP across five HOXA promoters, in two cell lines, before and after knockdown. It is in **Figure 6A**. The abstract never names the mark. **The failure has a signature worth memorising: the query that returns zero is the same query that returns the paper containing the answer, under a different operand.**
Why this class is systematically biased against exactly the data this project wants
Abstracts carry a paper's claim. Figures carry its measurements. The two differ most for the kinds of data a knowledge graph needs:
- Histone marks are almost always a supporting panel. A paper's abstract says 'activates HOXA'; the H3K79 ChIP is Figure 6.
- Negative controls and comparator cell lines are almost never in an abstract. Whether a menin inhibitor was tried on a line is exactly the sort of thing that appears once, in a supplementary table, unremarked.
- Co-IP panels. Which proteins were blotted in an immunoprecipitate is a methods-and-figure fact.
So the systematic bias is: a counted absence over PubMed will over-report absence for mechanism-level measurements and under-report it for headline claims. That is the opposite of what the reader of a knowledge graph assumes.
What to do instead, in cost order
- Grep the repo first. This project has previously declared a literature empty in one entry while citing the disproving paper in four others (rule 64). Do this before anything else; it is free.
- Read the full text of the two or three papers your subject query DOES return. On this beat the whole correction came from reading the one paper
SET-NUP214 AND DOT1Lreturned. If your subject has fewer than ~60 records, the subject query plus reading is cheaper and better than any number of intersection queries. - Use a full-text index for the negative. Europe PMC's search covers open-access full text and is a different corpus from PubMed; it will not cover paywalled figures but it moves the bound.
- State the corpus in the caveat, not just the query and the name list. The correct wording is '0 records in PubMed titles, abstracts and MeSH', never 'nobody has measured'.
- When the subject is small, invert the query. Ask 'what does the SET-NUP214 literature contain' (54 records, readable in an afternoon) rather than 'does SET-NUP214 intersect X' (one number, unbounded).
How this differs from rule 71, which it resembles
Rule 71 is the PHGDH case: a symbol query returned 113 records and title-checked to zero, and the answer was in a paper that used the inhibitor's name and a pathway phrase instead of the symbol. That is a vocabulary failure - the paper's abstract did contain the finding, in words the query did not try.
This is different. The vocabulary was right. There is no word that would have found it, because the abstract does not contain the finding at all. Widening the name list cannot fix a corpus bound, and that is why it needs its own entry rather than a line in rule 70's name-list discipline.
| concept class | method / evidence discipline |
|---|---|
| family | counted absences |
| correct wording | 0 records in PubMed titles, abstracts and MeSH as of <date>, with <query> and <name list> - NOT 'nobody has measured' |
| cheapest mitigation | if the subject has fewer than a few hundred records, read the subject literature instead of intersecting it |
distinguishes from
{
"rule_67": "the query cannot return a one - a mechanics failure",
"rules_70_to_72": "the name list is wrong - a vocabulary failure",
"rule_64": "the complex form is the only reachable phrasing - also a vocabulary failure",
"this_concept": "the corpus does not contain the measurement - the abstract does, and it is not in the abstract"
}worked case
{
"subject": "SET-NUP214",
"question": "has a histone mark ever been measured on this fusion?",
"pubmed_answer": 0,
"true_answer": "yes - H3K79me2 ChIP at five HOXA promoters, PMID 18299449, Figure 6A",
"why_invisible": "the measurement is a supporting figure and the abstract never names the mark",
"controls_that_passed": "positive control through the identical shape returned 1; four spellings of the mark were tried"
}- This concept can be over-applied into paralysis: if no counted absence can ever be trusted, the project loses a tool it genuinely needs. The correct response is to state the corpus, not to stop counting.
- Reading full texts instead of counting is much more expensive per question and does not scale to a genome-wide sweep, so the mitigation only works for small, well-defined subjects.
- How many of this graph's 1,878 absence assertions are of the intersection kind, where the corpus bound applies? The existing bounded-or-not analysis classifies them by whether a bound is stated, not by what kind of bound it is.
- Would a Europe PMC full-text sweep over this graph's existing counted absences change any of them? It is one script and it has never been run.
- Is there a threshold subject size below which 'read the whole subject literature' should simply be mandatory in this project? 54 records was comfortably readable; 220 would not have been.
- E1strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.database
PubMed esearch on 2026-09-20 returns 0 records for `SET-NUP214 AND (H3K79 OR "H3-K79" OR "K79" OR "lysine 79")` and 0 for `SET-NUP214 AND (H3K79 OR H3K4)`, with a positive control of 1 for `SET-NUP214 AND DOT1L` through the identical query shape and 21 for `SET-NUP214 AND histone`.
PubMed E-utilities esearch counts for SET-NUP214 against histone-mark vocabularies, query translations read back · NCBI PubMed 2026What this evidence cannot show
A single worked case. The general claim that PubMed indexes titles, abstracts and MeSH rather than full text is standard and is not established by this measurement; what this measurement establishes is that the consequence bites on a real question in this project's own subject matter.
Evidence that argues against this (1)
- The project already holds a rule of similar shape - brief rule 71, whose worked case is this paper: a PHGDH symbol query returned 113 records and title-checked to zero, while this paper tests serine biosynthesis directly in MLL-ENL leukaemia under the inhibitor name CBR5884 and a Ser/Gly-free diet. A reader could argue the present concept is the same rule. It is not: there the finding IS in the abstract, under a vocabulary the query did not try, so widening the name list fixes it. Here no name list can fix it, because the abstract does not contain the finding. The two have different remedies, which is why they need separate entries. Garcia-Cuellar MP et al. 2020
- E2strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.paper
PMID 18299449 reports chromatin immunoprecipitation with an antidimethyl histone H3K79-specific antibody showing the mark at HOXA1, HOXA3, HOXA9, HOXA10 and HOXA11 promoters in LOUCY, absent in SKW3, and partially lost upon siRNA-mediated SET-NUP214 inactivation - and its abstract does not mention H3K79, histone methylation or any chromatin mark.
The recurrent SET-NUP214 fusion as a new HOXA activation mechanism in pediatric T-cell acute lymphoblastic leukemia. · Van Vlierberghe P, et al. · Blood 2008What this evidence cannot show
Established by reading the full text via the PMC HTML route (PMC2343598) after Europe PMC fullTextXML returned HTTP 500, and by comparing it against the abstract returned by efetch. The abstract's text was checked for the strings H3K79, H3-K79, K79 and 'methylation'; none occurs.
Evidence that argues against this (1)
- METHODOLOGICAL DEFEATER AGAINST THE REMEDY, not against the finding. The rule's prescribed fix - open the likely papers and read the figures - moves the bound rather than removing it, and this beat's own reading log shows where it stops. PMID 34320268, the one paper that maps the SET-NUP214/MLL contact and therefore the highest-value full text in this layer, is not in PMC: Europe PMC fullTextXML, efetch db=pmc and the PMC HTML route are all structurally unavailable for it, and every entry in this beat that cites it records ABSTRACT LEVEL ONLY. For a paywalled paper the corpus bound is identical to the PubMed bound. So a claim of the form 'the supporting figures were checked' has to name which papers could actually be opened, and this entry establishes that an abstract index is narrower than the paper - not that the paper is reachable. Cigdem S et al. 2021
the corpus bound that limits this beats menin absence, worked on the H3K79 case
the H3K79 ChIP recovered by reading the paper the DOT1L query returned