The methylome co-mutation table, recomputed: IDH2 up at p = 0.0004 and not p < 0.0001, IDH1 and DNMT3A flat, and the abstract disagrees with its own Table 2
also called IDH2 enrichment recomputation · FoundationOne Heme KMT2A co-mutation table
The claim that IDH2 mutations are commoner in KMT2A-rearranged leukaemia was checked by recalculating it from the paper's own numbers. The enrichment is real. The p-value quoted everywhere - including in this project's own notes - came from the paper's abstract and is ten times stronger than the paper's own table says.
The same thing, for a scientist
**Four methylome genes, one assay, one cohort, recomputed independently.** The source is the 3,863-case FoundationOne Heme AML series comparing 521 KMT2A-altered with 3,342 KMT2A-wild-type cases [E1]. Counts were back-calculated from the published percentages and two-sided Fisher exact tests and Wilson intervals computed from scratch [E2]: | gene | KMT2Ar | KMT2A-wt | OR | recomputed p | paper's Table 2 | paper's abstract | |---|---|---|---:|---:|---|---| | **IDH2** | 83/521, 16.0% (13.0-19.3) | 348/3342, 10.4% (9.4-11.5) | **1.630** | **4.1e-4** | **0.0004** | ***p < 0.0001*** | | TET2 | 53/521, 10.2% (7.9-13.1) | 548/3342, 16.4% (15.2-17.7) | 0.577 | 1.6e-4 | 0.0002 | 0.0002 | | IDH1 | 35/521, 6.7% (4.9-9.2) | 254/3342, 7.6% (6.7-8.5) | 0.876 | **0.53** | ns | ns | | DNMT3A | 83/521, 16.0% (13.0-19.3) | 575/3342, 17.2% (16.0-18.5) | 0.912 | **0.49** | ns | ns | **Three results, in descending order of how much they should change what the graph says.** **1. The IDH2 enrichment is real and its p-value is wrong in this project's own records.** The recomputation reproduces Table 2 exactly (0.00041 vs 0.0004) for IDH2 and reproduces it for TET2 and DNMT3A too, which is the check that the back-calculation is sound. It does **not** reproduce the abstract's *p < 0.0001*. **The dispatch brief for this beat, the cycle-44 king log, and a counter-evidence note inside `gene-tet2` all carry the abstract's figure.** They should read **p = 0.0004**. **2. IDH1 is flat and that is the informative row** (rule 87). Same oncometabolite, same mechanism class, same instrument, odds ratio 0.876, p = 0.53. A pure 2-HG explanation of the IDH2 row predicts IDH1 to move and it does not. **3. DNMT3A is flat too**, at 16.0% against 17.2%. So of four methylome genes in one table: **one enriched, one depleted, two flat** - and the enriched and the depleted are mutually exclusive lesions whose sum is 26.1% against 26.8%. **The null, as rule 84 requires, is supplied by the study itself** - the KMT2A-wild-type arm of the same assay - which is the best form available. **The multiplicity null is not supplied by anybody:** the paper compares roughly seventeen genes and reports no adjustment. A Bonferroni threshold of 0.05/17 = 0.0029 is cleared by IDH2 (4.1e-4) and TET2 (1.6e-4) and by nothing else.
Why the abstract and the table disagree, and which to believe
The abstract of PMID 41514668 states "IDH2 (16.0% vs. 10.4%; p < 0.0001)". Table 2 of the same paper gives the same two percentages with p = 0.0004. The Results section repeats the abstract's p < 0.0001, so the discrepancy is between the table and the prose, twice.
The recomputation adjudicates it. Back-calculated counts give Fisher p = 4.1e-4, which rounds to the table's 0.0004. The same procedure applied to the other rows reproduces the table's 0.0002 for TET2 (1.6e-4) and its 'ns' for IDH1 (0.53) and DNMT3A (0.49). An arithmetic procedure that reproduces three of the table's four values and the fourth to one significant figure is reproducing the table, so the table is right and the prose is a transcription error.
Why this matters more than a decimal place. p < 0.0001 reads as a strong, essentially unassailable finding. p = 0.0004 in a retrospective commercial-panel series with seventeen unadjusted comparisons, no outcome data, and a 43%-duplication comparison arm reads as one finding that needs replication. This project has been quoting the first and should quote the second, and the difference changes how much weight a future beat is entitled to put on the methylome threat.
And the process failure is the generalisable part. The figure entered this project through an abstract. Nobody opened the table until this beat. An abstract is a summary written by the authors and is not the paper's own data, and this project has a rule about not trusting an identifier from a listing for exactly the same reason. The rule extends to numbers.
What the recomputation cannot do
Stated plainly, because a recomputation from marginals is easy to over-trust:
- It cannot detect an error in the authors' counting. If the percentages themselves are wrong, every figure here is wrong in the same way.
- It adds rounding error of about one case per cell. 16.0% of 521 could be 83 or 84; both give p in the 4e-4 range, so this row is robust, but a borderline row would not be.
- It cannot stratify by lesion type. 43.1% of the KMT2A-altered arm are KMT2A duplications rather than fusions and the co-mutation table is not broken down. This is the single largest unresolved confound in the beat, because KMT2A-PTD is an older, myelodysplasia-associated disease with a different co-mutation landscape from a KMT2A fusion.
- It cannot test exclusivity. IDH and TET2 mutations are mutually exclusive, so the IDH2 and TET2 rows are not independent, and the paper publishes no joint table.
- It says nothing about infants. Median age 62. No paediatric or infant KMT2A-rearranged cohort reporting IDH status was found anywhere in this beat.
One thing it does do that no re-reading could: it caught an internal inconsistency in a published paper that four secondary sources had propagated without noticing.
| script | none committed - recomputation run inline in Python 2026-09-21; the inputs are the four percentage pairs and the two denominators, so it is re-derivable in three lines from the table below |
|---|---|
| status | run |
| method | counts back-calculated as round(pct x n); two-sided Fisher exact by exact hypergeometric enumeration; Wilson score 95% intervals at z = 1.959964 |
| null | The KMT2A-wild-type arm of the same assay in the same cohort - the best available form of a null for this question (rule 84). |
| multiplicity | ~17 genes compared, no adjustment reported by the paper. Bonferroni 0.05/17 = 0.0029, cleared by IDH2 (4.1e-4) and TET2 (1.6e-4) only. |
| largest unresolved confound | 43.1% of the 521 KMT2A-altered cases are KMT2A duplications (PTD), not fusions, and the co-mutation table is not stratified by lesion type. |
inputs
[ "PMID 41514668 Table 2 percentages, read from PMC12784994 full text", "n_kmt2a_altered = 521, n_kmt2a_wildtype = 3342, n_total = 3863" ]
results — the numbers behind the chart
{
"IDH2": {
"k_r": 83,
"n_r": 521,
"pct_r": 16,
"wilson_r": [
13,
19.3
],
"k_w": 348,
"n_w": 3342,
"pct_w": 10.4,
"wilson_w": [
9.4,
11.5
],
"odds_ratio": 1.63,
"fisher_p": 0.000411,
"paper_table2_p": "0.0004",
"paper_abstract_p": "<0.0001",
"verdict": "ENRICHED; abstract overstates by an order of magnitude"
},
"TET2": {
"k_r": 53,
"n_r": 521,
"pct_r": 10.2,
"wilson_r": [
7.9,
13.1
],
"k_w": 548,
"n_w": 3342,
"pct_w": 16.4,
"wilson_w": [
15.2,
17.7
],
"odds_ratio": 0.577,
"fisher_p": 0.00016,
"paper_table2_p": "0.0002",
"verdict": "DEPLETED; reproduces"
},
"IDH1": {
"k_r": 35,
"n_r": 521,
"pct_r": 6.7,
"wilson_r": [
4.9,
9.2
],
"k_w": 254,
"n_w": 3342,
"pct_w": 7.6,
"wilson_w": [
6.7,
8.5
],
"odds_ratio": 0.876,
"fisher_p": 0.531,
"paper_table2_p": "ns",
"verdict": "FLAT; the internal negative control"
},
"DNMT3A": {
"k_r": 83,
"n_r": 521,
"pct_r": 16,
"wilson_r": [
13,
19.3
],
"k_w": 575,
"n_w": 3342,
"pct_w": 17.2,
"wilson_w": [
16,
18.5
],
"odds_ratio": 0.912,
"fisher_p": 0.491,
"paper_table2_p": "ns",
"verdict": "FLAT"
}
}correction owed outside this file set
[ "data/graph/nodes/cxxc-readers/gene-tet2.json - evidence E5 counter_evidence note and the first open_question both state 'p < 0.0001' for the IDH2 enrichment; should be 'p = 0.0004 (Table 2; the abstract's p < 0.0001 does not match the paper's own table)'", "docs/king-log/2026-09-21-cycle-44.md - 'IDH2 is ENRICHED in KMT2A-rearranged AML (16.0% vs 10.4%, p < 0.0001)' appears twice and carries the same error" ]
- The corrected p-value weakens, but does not remove, the live threat that a hypermethylating co-mutation antagonises the fusion's unmethylated-CpG address. A weakened threat is more dangerous than a strong one, because it is easier to dismiss - and the experiment that would settle it (a fusion-only cohort) still has not been run.
- Does the IDH2 enrichment survive restriction to KMT2A FUSIONS? This is the one analysis that would convert a mixed-arm association into a statement about the index lesion, and it needs only a lesion-type-annotated cohort.
- Has anyone else measured IDH2 frequency in a KMT2A-rearranged cohort? One paper is one paper, and its abstract disagrees with its own table.
- What is the IDH1/IDH2 frequency in PAEDIATRIC and INFANT KMT2A-rearranged leukaemia? Not measured anywhere found in this beat - an absence of measurement, not a zero.
- Should this graph adopt a standing practice of recomputing any published proportion it intends to quote more than once? This recomputation took minutes and caught an error four secondary sources had carried.
- E1strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.paper
The FoundationOne Heme series of 3,863 AML cases reports, in Table 2, IDH2 at 16.0% in 521 KMT2A-altered versus 10.4% in 3,342 KMT2A-wild-type cases with p = 0.0004, TET2 at 10.1% versus 16.4% with p = 0.0002, IDH1 at 6.7% versus 7.6% (ns) and DNMT3A at 16.0% versus 17.2% (ns) - while the paper's abstract and Results text give p < 0.0001 for the IDH2 row.
Characterizing KMT2A Rearrangement in Acute Myeloid Leukemia: A Comprehensive Genomic Study. · Batayneh O, Moein M, Naji NS, et al. · Cancers (Basel) 2026What this evidence cannot show
Read in full text from PMC12784994 on 2026-09-21 - the Table 2 values were extracted from the tag-stripped full text rather than from a rendered table, so a column transposition in the extraction is conceivable; it was checked against the abstract's directional statements (TP53 and NPM1 higher in wild-type, FLT3 and KRAS higher in rearranged) and the column order is consistent. Adult AML, median age 62; commercial targeted panel; germline filtering inferred; no outcome data; no multiplicity adjustment across ~17 genes; 43.1% of the KMT2A-altered arm are duplications rather than fusions and the table is not stratified by lesion type.
Evidence that argues against this (1)
- A paper that disagrees with itself on one p-value may have other internal inconsistencies that a single full-text read does not surface, so this table should be treated as a source needing replication rather than as a settled reference - which is the opposite of how the figure has been used in this project so far. Batayneh O et al. 2026
- E2strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.dataset
Independent recomputation from counts back-calculated as round(percentage x n) gives, by two-sided Fisher exact test, IDH2 83/521 vs 348/3342 odds ratio 1.630 p = 4.1e-4; TET2 53/521 vs 548/3342 odds ratio 0.577 p = 1.6e-4; IDH1 35/521 vs 254/3342 odds ratio 0.876 p = 0.53; DNMT3A 83/521 vs 575/3342 odds ratio 0.912 p = 0.49 - reproducing the paper's Table 2 in every row and contradicting its abstract's p < 0.0001 for IDH2; Wilson 95% intervals are IDH2 13.0-19.3 vs 9.4-11.5, IDH1 4.9-9.2 vs 6.7-8.5.
Independent Fisher exact and Wilson-interval recomputation of the KMT2A-altered versus KMT2A-wild-type methylome co-mutation table of PMID 41514668 · Oncofusion knowledge-graph project, Q81-methylome-axis · Recomputed 2026-09-21 in Python from percentages read out of PMC12784994 Table 2; two-sided Fisher exact by exact hypergeometric enumeration, Wilson score intervals at z = 1.959964 2026What this evidence cannot show
A recomputation from marginals, not from patient-level data: it inherits every ascertainment property of the original assay and adds about one case of rounding error per cell. It can detect an internal inconsistency in the paper - which it did - but not an error in the authors' own counting. No multiplicity adjustment is applied; at 17 genes a Bonferroni threshold is 0.0029, cleared only by IDH2 and TET2. The exclusivity between IDH and TET2 means the IDH2 and TET2 rows are not independent and no joint test is possible from published data.
Evidence that argues against this (1)
- Agreement between a back-calculation and the published table could in principle be coincidental at one significant figure rather than a validation of the method; the definitive check is the authors' count table, which is not published. Batayneh O et al. 2026
Supplies the verified figure: OR 1.630, Fisher p = 4.1e-4, Wilson 13.0-19.3 vs 9.4-11.5.
Supplies the internal negative control: OR 0.876, Fisher p = 0.53.
Corrects the p-value in the gene-tet2 E5 counter-note and its first open question: the IDH2 enrichment is p = 0.0004 per Table 2, not the abstract's p < 0.0001. The TET2 depletion figure itself reproduces exactly.
Fourth methylome gene in the same table and also flat: 16.0% vs 17.2%, recomputed p = 0.49.
Of four methylome genes in one assay in this disease: one enriched, one depleted, two flat.