The 3.7% was mostly the query: DOT1L recall is 19.6% once you ask about leukaemia
also called narrowed DOT1L literature recall · DOT1L/H3K79 recall re-measurement
Last cycle a check said we had read only 3.7% of recent DOT1L papers, the worst of any topic we measured. Re-running it with the word "leukaemia" added to the search puts the number at 16-20%, roughly in line with the rest of the project, because most DOT1L papers are about other diseases entirely.
The same thing, for a scientist
`ana-literature-recall` reported DOT1L/H3K79 at **13 of 354 primary papers, 3.7%**, the worst of its five queries by a factor of four, and flagged in its own counter-evidence that the query `DOT1L[tiab] OR "H3K79"[tiab]` carried no disease restriction. This re-runs the identical machinery over a ladder of five narrowings. T0 reproduces the parent number exactly (354 primary, 13 cited, 3.7%), which validates the reimplementation. Adding leukaemia words takes the denominator from 354 to **67 primary papers** and the recall to **16.4%**; restricting to KMT2A/MLL specifically gives **11 of 56, 19.6%**. So the headline was roughly **five-fold pessimistic**, and DOT1L is not an outlier: at 19.6% it sits above the KMT2A-rearranged query (15.2%) and the elongation-partner query (12.8%), and below menin in leukaemia (31.7%). Expanding the aliases in the other direction (DOT1, KMT4, KIAA1814) adds 54 papers and **zero** cited ones, which is the honest residual gap. The analysis also emits the thing that is actually usable: **62 uncited primary papers on the narrowed beat**, 45 of them KMT2A-specific, and this cycle's entries are written from that list.
The ladder, and what each rung costs
| tier | in window | primary | cited | primary recall |
|---|---|---|---|---|
| T0 original, unrestricted | 452 | 354 | 13 | 3.7% |
| T1 alias-expanded, unrestricted | 506 | 404 | 13 | 3.2% |
| T2 alias + leukaemia words only | 103 | 67 | 11 | 16.4% |
| T3 alias + leukaemia vocabulary | 116 | 77 | 11 | 14.3% |
| T4 alias + KMT2A/MLL specifically | 91 | 56 | 11 | 19.6% |
Two things are worth separating 1. The narrowing direction raises recall by deleting papers a leukaemia graph should never have been scored against — DOT1L in prostate cancer, in multiple myeloma, in nephron progenitors, in naive pluripotency, in C. elegans. The alias direction lowers it: T1 adds 54 papers over T0 and not one of them is cited, so every symbol this project was not searching on is a paper it has not read. Both directions are real and they are not the same measurement.
T3 scores lower than T2 despite being a superset, which is the expected arithmetic — widening to AML, myeloid, lymphoblastic adds 10 primary papers and no cited ones. The tightest and the loosest leukaemia tiers bracket the answer at 14-20%.
What survives the correction
The parent analysis's conclusion does not survive in the form it was written. "DOT1L is the worst-covered by a wide margin" was a statement about a contaminated denominator; on a comparable denominator DOT1L is mid-field, better read than the elongation partners and the KMT2A-rearranged literature generally.
What does survive is smaller and more specific: 11 papers is a thin base for a node the graph treats as central, the alias gap is real, and the 62-paper reading queue contains at least four things the graph had no record of at all — the menin-dependent partitioning of H3K79me2 into two regulatory units (PMID 37679565), DOT1L PROTAC degraders with in vivo regression (PMID 41685902), DOT1L antagonism of PRC1.1 as transcriptional memory (PMID 41634341), and catalysis-independent DOT1L function in cell-fate determination (PMID 33077595). None of those are visible from a recall percentage; they are visible from reading the list the percentage produces.
The transferable lesson is about the shape of the error. A recall number computed against an unrestricted gene-symbol query reads as a fact about the project and is largely a fact about how many diseases the gene is studied in. Any future recall measurement on a gene with a life outside leukaemia should carry the disease restriction into the query, not into the caveat.
What this cannot tell you
No tier distinguishes not read from read and judged off-beat, which is the same blind spot the parent analysis named and did not fix 1. A query-level recall also cannot see whether the 11 cited papers are the right 11; 19.6% recall built entirely on medicinal-chemistry papers would be worse than 10% built on the mechanistic ones.
The title-level leukaemia filter reported in the results file is a lower bound on on-topicness, because a leukaemia paper can have a title with no disease word in it, and it should not be read as a precision estimate.
| script | analyses/dot1l-literature-recall/run.py |
|---|---|
| status | run |
| run date | 2026-09-19 |
| graph cited pmids at run | 1459 |
| n uncited primary on beat | 62 |
| result summary | 3.7% was a query artefact by roughly five-fold; the leukaemia-restricted recall is 16.4-19.6%, mid-field for this graph. The residual gap is real but small, and the actionable output is a 62-paper reading queue. |
inputs
[ "PubMed E-utilities esearch/esummary", "data/graph/nodes/*/*.json cited PMIDs" ]
date window
[ "2020/01/01", "2026/12/31" ]
tiers — the numbers behind the chart
[
{
"tier": "T0 original (unrestricted)",
"primary": 354,
"cited": 13,
"primary_recall_pct": 3.7
},
{
"tier": "T1 alias-expanded (unrestricted)",
"primary": 404,
"cited": 13,
"primary_recall_pct": 3.2
},
{
"tier": "T2 alias + leukaemia words only",
"primary": 67,
"cited": 11,
"primary_recall_pct": 16.4
},
{
"tier": "T3 alias + leukaemia vocabulary",
"primary": 77,
"cited": 11,
"primary_recall_pct": 14.3
},
{
"tier": "T4 alias + KMT2A/MLL specifically",
"primary": 56,
"cited": 11,
"primary_recall_pct": 19.6
}
]aliases searched
[ "DOT1L", "H3K79", "DOT1", "KMT4", "KIAA1814", "H3 K79", "lysine 79" ]
- Acting on a corrected recall number can produce the opposite error: if 19.6% is read as 'coverage is fine', the 62-paper queue stops being worked and the alias gap that T1 exposed stays open.
- Narrowing a query to raise a metric is the classic way to make an audit stop reporting. The ladder is published in full precisely so that the narrowing is auditable rather than a single replacement number.
- Are the 11 cited papers the right 11? A recall percentage is blind to whether coverage is concentrated in medicinal chemistry rather than mechanism, and this beat's cited set leans clinical.
- The four other queries in the parent analysis were never checked for the same contamination. 'elongation-complex fusion partners' uses bare symbols with no disease restriction and its 12.8% is probably pessimistic by a similar mechanism.
- Should a recall instrument record bounded REJECTIONS - papers an agent read and judged off-beat - so that 'not cited' stops being indistinguishable from 'considered and declined'?
- E1established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.analysis
Re-running the parent recall machinery with the DOT1L/H3K79 query restricted to leukaemia raises primary-paper recall from 3.7% (13/354) to 16.4% (11/67) with leukaemia words and 19.6% (11/56) restricted to KMT2A/MLL; alias expansion without disease restriction adds 54 papers and zero cited ones; 62 uncited primary papers remain on the narrowed beat.
DOT1L/H3K79 literature recall, re-measured after narrowing the query to leukaemia · Oncofusion knowledge graph (internal analysis) 2026What this evidence cannot show
Run 2026-09-19 by Q3b. PubMed esearch, publication date 2020/01/01-2026/12/31, retmax 1200, sorted by publication date; esummary for publication types and titles; a paper counts as READ if any graph entry cites it in support or in counter-evidence; 'primary' excludes Review, Systematic Review, Editorial, Comment, Letter, Case Reports, Erratum, Meta-Analysis. Tier T0 reproduces the parent analysis's numbers exactly, which is the only check that the reimplementation is faithful. The graph's cited-PMID set was 1,459 at run time against the parent's 1,453, so the two runs are six PMIDs apart and the tiny drift is not material. Every number is recall against a query someone chose.
Evidence that argues against this (2)
- A narrowed query flatters the measurement as reliably as a broad one deflates it. Requiring a leukaemia word in title or abstract excludes basic chromatin papers that are genuinely load-bearing for this beat - the non-processivity kinetics, the methylation-state zoning, the H2Bub structural work - none of which mention leukaemia and all of which this project should read. 19.6% is recall against the clinical literature, not against the biology. Frederiks F et al. 2008
- The parent analysis stated this correction in advance, in its own counter-evidence on E4, and its open_questions asked for exactly this re-measurement. So this is the instrument working as designed rather than an error being caught, and the honest reading is that the 3.7% figure should never have been quoted without the caveat attached. Oncofusion knowledge graph (internal analysis) 2026
- E2established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.analysis
The parent measurement placed DOT1L/H3K79 at 3.7% primary recall against menin in leukaemia at 31.7%, KMT2A-rearranged at 15.2%, infant ALL at 16.0% and elongation-complex partners at 12.8%.
Literature recall of the Oncofusion knowledge graph against five stated PubMed queries, and the laboratories it never cites · Oncofusion knowledge graph (internal analysis) 2026What this evidence cannot show
The comparator queries were not re-measured here, so the cross-query comparison mixes one narrowed denominator against four unnarrowed ones. The menin query is already disease-restricted by construction ('menin AND leukaemia'), so that comparison is fair; the KMT2A-rearranged and infant-ALL queries are also disease-restricted. The elongation-partner query is not fully restricted and its 12.8% is probably also pessimistic.
Evidence that argues against this (1)
- Comparing a re-measured number against four numbers measured under a different query discipline is exactly the methodological move this project has warned itself about elsewhere. The defensible claim is the within-analysis one - 3.7% to 19.6% on the same corpus - not the ranking against the other four beats. Oncofusion knowledge graph (internal analysis) 2026
the 3.7% headline was a query artefact by ~5x; leukaemia-restricted recall is 16.4-19.6%
measures how much of the recent DOT1L literature this graph has read