The 0.069 chance floor quoted beside the 0.400 same-patient Jaccard is the cohort MEDIAN's null, and this pair's own floor is 0.267
An earlier analysis said one patient's two leukaemias shared six times more target genes than chance. The 'chance' number it used belonged to a different, much smaller comparison; using the right one, the excess is about one and a half times, not six.
The same thing, for a scientist
q144 reported J = 0.400 for p279/p318 and, in the same entry, an independence expectation of 0.069 with the cohort median at 1.63x chance. The 0.069 is `median(ab/N)` over all 351 pairs. It is not the floor for this pair. p279 calls 2,045 and p318 calls 1,883 promoters in q144's 4,646-promoter universe - 44% and 41% of it - so the independence expectation for THIS pair is **0.267**, and the activity-stratified floor is **0.272**. The same-patient excess is therefore **1.46x**, not the 5.8x that 0.400/0.069 implies. The published figure was not wrong about anything it computed; it was placed beside a number describing a different pair, and the adjudication that followed treated 5.8x-over-chance as the reason to prefer it.
The arithmetic, in q144's own universe with q144's own rule
| quantity | value |
|---|---|
| universe (q144's 27-way intersection) | 4,646 |
| p279 calls / p318 calls (depth-matched) | 2,045 / 1,883 |
| fraction of the universe each calls | 44.0% / 40.5% |
| observed Jaccard | 0.397 |
| independence floor for THIS pair | 0.267 |
| activity-stratified floor for THIS pair | 0.272 |
| excess | 1.46x |
| cohort-median independence floor (what was quoted) | 0.069 |
| excess implied by 0.400 / 0.069 | 5.8x |
| overstatement | 3.97-fold |
The reason the two floors differ so much is that they describe sets of different sizes. A 1,000-call pair and a 2,000-call pair in the same universe have chance Jaccards of 0.12 and 0.27. A median over 351 pairs of wildly differing call counts - 0 to 2,282 - is a floor for none of them.
The same correction applied to q142, which computed no null at all
q142's quantile rule fixes both sets at 971 in a universe of 19,404, so its implied independence floor is 0.0256 and its implied excess 4.57x - a figure the beat never stated. Re-implemented on fragment counts the same cell gives J = 0.164, independence floor 0.0255 (6.44x) and an activity-stratified floor of 0.0611, i.e. 2.69x. So q142's unstated excess was overstated by 2.4-fold by the same mechanism, in the opposite universe.
Both beats' excess figures were upper bounds, as the coordinator predicted, and the size of the over-estimate is 2.4x for q142 and 3.97x for q144.
The rule this generalises to
A chance floor for a pairwise statistic is a function of that pair's set sizes and universe. A cohort median is a summary of many floors and is the floor of nothing. When an entry reports observed beside expected, both numbers must be indexed by the same pair - and an entry that reports a cohort median for one and a single pair for the other is comparing two different experiments.
| script | analyses/q146-one-universe-jaccard/check146.py |
|---|---|
| status | run |
| result summary | q144's 0.400/0.069 = 5.8x becomes 0.397/0.272 = 1.46x once the floor is computed for the right pair; a 3.97-fold overstatement. q142's unstated 6.44x becomes 2.69x, a 2.4-fold overstatement. |
| q144 published excess | 5.8 |
| q144 corrected excess | 1.459 |
| q144 shrinkage | 3.97 |
| q142 implied excess | 6.441 |
| q142 corrected excess | 2.689 |
| q142 shrinkage | 2.4 |
inputs
[ "analyses/q146-one-universe-jaccard/results.json" ]
- An excess-over-chance figure is how this corpus decides whether a shared target set is real. A systematic 2.4-4x overstatement in that figure would make shared-target arguments for combination therapy look stronger than the data support.
- How many other pairwise excess-over-chance figures in this corpus divide a single pair's observation by a cohort-level expectation? A grep for entries reporting an observed-over-expected ratio and a cohort median in the same payload would find them.
- E1established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.dataset
The call counts, universe and rule that generate these floors are those of the deposited GSE252378 KMT2A N/C arm re-scored under the paper's own criterion.
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
The 2,045 and 1,883 call counts are a re-implementation on refGene promoters after within-pair depth matching, not the paper's own site calls; q144's un-matched figures were 2,045 and 1,894.
Evidence that argues against this (1)
- Depth matching changes p318's count from 1,894 to 1,883 - a 0.6% change - so depth matching is not what drives any of the differences reported here, and the earlier beat's counts were not depth-confounded at this pair. Janssens DH et al. 2024
- E2established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.paper
Uniform-placement expectations are the standard failure mode in interval-overlap testing and the reason simulation frameworks restrict the randomisation space per comparison.
GAT: a simulation framework for testing the association of genomic intervals · Heger A, Webber C, Goodson M, Ponting CP, Lunter G · Bioinformatics 29:2046-2048 2013What this evidence cannot show
GAT addresses genomic placement rather than set-size mismatch across a cohort; the set-size point here is elementary probability, not a result from that paper.
Evidence that argues against this (1)
- A sufficiently large cohort median could be defended as a summary of a null DISTRIBUTION rather than a floor; but it was used as a divisor for a single pair, which is the misuse regardless. Gel B et al. 2015
the corrected floors are the activity-stratified ones
0.069 is that entry's cohort-median floor and is not the floor for the same-patient pair
the null correction behind the headline's 1.33x
zero-call pairs in the 351-pair median are part of why 0.069 is so low
the worked case of a cohort median used as a pair's floor
the worked case: 5.8x becomes 1.46x