The '0.400, keeps 1,894 of 2,045 calls' in the dispatch brief is two independent per-sample call counts read as a retention ratio; the intersection is 1,116 of a 2,812 union
A summary of an earlier result said one sample kept 1,894 of another's 2,045 target genes. Those are just the two samples' separate totals sitting next to each other in a table; the number they actually share is 1,116.
The same thing, for a scientist
The dispatch brief for this beat described q144's same-patient result as 'Jaccard 0.400 - keeps 1,894 of 2,045 calls'. In q144's `paper_criterion_site_calls.per_sample` table, `p279: 2045` and `p318: 1894` are two independent per-sample call counts. Neither is an intersection and no retention was computed. Re-derived in q144's own universe with its own rule and within-pair depth matching: p279 calls 2,045, p318 calls 1,883, the intersection is **1,116** and the union **2,812**, giving J = 0.397. A 92.6% retention reads as strong conservation of the target set; the true figure is that **45.4% of p279's calls are lost at lineage switch and 40.7% of p318's calls are new**.
The two readings
| reading | intersection | interpretation |
|---|---|---|
| '1,894 of 2,045 kept' | 1,894 | 92.6% of the diagnostic target set survives the lineage switch |
| what the data say | 1,116 | only 54.6% of p279's calls are retained; 40.7% of p318's are new |
Both are consistent with J = 0.40 only in the sense that neither was computed from the other. The Jaccard was computed correctly in q144; the retention was never computed at all.
Why the misreading is the natural one
Two numbers of similar size in adjacent rows of a per-sample table, for two samples that are known to be one patient, invite being read as before-and-after. The table was per-sample and correct; the summary that travelled was not. This is the same shape as a correct total printed above a truncated list: the arithmetic was never wrong, only the sentence built from it, and the sentence is what gets quoted into the next brief.
The number that should be quoted instead
For the lineage switch specifically, the informative pair of numbers is the asymmetric retention, not the symmetric Jaccard: 1,116 / 2,045 = 54.6% of the B-ALL target promoters are still called in the AML, and 1,116 / 1,883 = 59.3% of the AML's calls were already targets in the B-ALL. Both are well above the matched chance floor and both are far from 92.6%.
| script | analyses/q146-one-universe-jaccard/run146.py |
|---|---|
| status | run |
| result summary | The '1,894 of 2,045' retention does not exist in the data. Depth-matched in q144's universe: intersection 1,116, union 2,812, J=0.397. 54.6% of p279's calls are retained, not 92.6%. |
| intersection | 1116 |
| union | 2812 |
| n p279 | 2045 |
| n p318 depth matched | 1883 |
| retention of p279 calls | 0.546 |
| share of p318 calls preexisting | 0.593 |
| briefed retention | 0.926 |
| loss of p279 calls | 0.454 |
inputs
[ "analyses/q144-unbounded-magnitude/results.json", "analyses/q146-one-universe-jaccard/results.json" ]
- A 92.6% retention figure would support carrying a diagnostic target list into salvage therapy. The real 54.6% does not, and the difference is the whole clinical question.
- Is the asymmetric retention (54.6% out, 59.3% in) different from what unrelated Texas KMT2A::AFF1 pairs show? The symmetric Jaccard says no; the asymmetric pair of figures has not been compared to a matched null and should be.
- E1established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.dataset
The per-sample call counts (p279 2,045; p318 1,894 undepth-matched) and the 0.400 Jaccard are q144's own outputs on the deposited GSE252378 fragment data under the paper's calling rule.
GSE252378: AutoCUT&RUN of KMT2A-rearranged primary leukemia samples (KMT2A N- and C-terminal antibodies, ENL, DOT1L, H3K4me3, IgG) · Janssens DH, et al. · NCBI Gene Expression Omnibus 2024What this evidence cannot show
This entry re-derives the intersection from the same inputs with within-pair depth matching added, which moves p318's count from 1,894 to 1,883; the intersection of 1,116 is therefore from the depth-matched run, not from q144's exact arrays.
Evidence that argues against this (1)
- Without depth matching the intersection would differ slightly; the 92.6% reading is wrong by a factor of 1.7 either way, so depth matching is not what settles it. Janssens DH et al. 2024
- E2established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.paper
The paper itself frames the p279-to-p318 transition as a redistribution of oncoprotein targets rather than a retention of them, identifying B-ALL-specific and GMP-like-specific target sets that differ between the two samples.
MLL oncoprotein levels influence leukemia lineage identities · Janssens DH, Duran M, Otto DJ, Wu W, Xu Y, Kirkey D, Mullighan CG, Yi JS, Meshinchi S, Sarthy JF, Ahmad K, Henikoff S · Nature Communications 15:9341 2024What this evidence cannot show
The paper's set difference is computed on its own peak intervals with its own calibrator, so its magnitudes are not directly comparable to the 54.6% here.
Evidence that argues against this (1)
- The paper's own framing is that oncoprotein LEVELS fall at lineage switch, which is a magnitude claim; a falling level with an unchanged target set would also produce fewer calls at a fixed threshold, so a retention figure cannot by itself distinguish redistribution from dilution. Janssens DH et al. 2024
the per-sample table that was misread; the table itself is correct
the retention figure the headline replaces
54.6% of the B-ALL target promoters are still called in this AML state
the 54.6% retained / 40.7% new asymmetry
- ← depends onp279 and p318: one infant, two diseases, and an oncoprotein target map that does not carry over
the 54.6% / 40.7% asymmetric retention