GSE252378 carries an institution-level batch effect: unrelated pairs from the same hospital overlap more than cross-hospital pairs, and a cohort-median comparator absorbs that into the same-patient claim
Samples processed at the same hospital look more like each other than samples from different hospitals, whatever patient they came from. Because the one patient with two samples was from a single hospital, comparing them to a whole-cohort average made them look unusually similar when they were not.
The same thing, for a scientist
Restricting to unrelated pairs only (same-patient pairs excluded) and using the excess over the activity-stratified null as the statistic, pairs whose two samples come from the same institution have a higher median excess than cross-institution pairs: 1.456 (n=83) versus 1.372 (n=236), Mann-Whitney one-sided p = 0.0016. The four institutions named in the KMT2A N/C arm's GEO !Sample_description fields are Texas Children's Hospital (10 samples), Fred Hutch Cancer Center (8), St Jude Children's Hospital (7) and Seattle Children's Hospital (2). p279 and p318 are both Texas Children's. So any comparison of the same-patient pair to a cohort median charges the same-patient pair with a batch advantage it did not earn, and that is a sufficient explanation for why q144's 0.400 looked large against a 0.132 median.
The measurement
| n pairs | median excess over the matched null | |
|---|---|---|
| same institution | 83 | 1.456 |
| different institutions | 236 | 1.372 |
Mann-Whitney U = 11,930.5, one-sided p = 0.0016. Same-patient pairs excluded. Primary cell (U_pair, paper rule).
The effect is modest in size and clearly non-zero. It is the right size to explain the discrepancy: 1.456/1.372 = 1.06x on the excess scale, and the same-patient pair's excess (1.329) is below even the cross-institution median, which is why it fails on both comparators once the null is right.
Where it shows up most visibly
Anchored on p279 or p318 and swept across the six other KMT2A::AFF1 samples, the excess splits almost perfectly by hospital:
Texas Children's comparators p179 1.532 / 1.560 p247 1.469 / 1.406 p454 1.397 / 1.361
St Jude comparators TB051939 1.189 / 1.165 TB920488 1.200 / 1.160 TB986292 1.333 / 1.353
the same-patient pair 1.329
Every Texas comparator is above the same-patient value and five of six St Jude comparators are below it. On raw Jaccard the same-patient pair beats 8 of 12 anchored comparisons; on excess it beats 4 of 12. The raw-Jaccard 'advantage' is the St Jude samples pairing badly with everything.
Why this is not a curiosity
Batch structure correlated with the outcome of interest is the canonical way a high-throughput comparison goes wrong 2. Here it is correlated with the outcome by construction: a same-patient pair is, necessarily, a same-institution pair. Every same-patient comparison in a multi-site deposit is confounded with batch, and the confound always runs in the direction that flatters the same-patient hypothesis. The only honest comparators are institution-matched ones.
| script | analyses/q146-one-universe-jaccard/texas146.py |
|---|---|
| status | run |
| result summary | Unrelated pairs only: median excess over the matched null is 1.456 within institution (n=83) versus 1.372 across (n=236), Mann-Whitney one-sided p=0.0016. The same-patient pair is same-institution by construction, so a cohort-median comparator credits it with a batch advantage. |
| n within | 83 |
| n across | 236 |
| median within | 1.456 |
| median across | 1.372 |
| mannwhitney p one sided | 0.00156 |
inputs
[ "analyses/q146-one-universe-jaccard/results.json", "analyses/q144-unbounded-magnitude/manifest.json" ]
institution counts — the numbers behind the chart
{
"Texas Children's": 10,
"Fred Hutch": 8,
"St Jude": 7,
"Seattle Children's": 2
}- Any cross-sample target list assembled from this deposit inherits an institution structure. A target that looks recurrent because it was called in three samples from one hospital is not a recurrent target.
- Is the split institution or sequencing date? The deposit's GSM ranges cluster (GSM8000xxx, GSM8006xxx, GSM8371xxx), and if those ranges track submission batches the confound may be date rather than site - which would change which comparators are matched.
- Does the batch effect survive restricting to samples of the same fusion partner and similar depth, or is it a depth effect wearing an institution label? The St Jude samples are the ones this corpus already recorded as having the shallowest libraries.
- E1established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.dataset
The institution of every sample in GSE252378's KMT2A N/C arm is stated in the GEO sample description; p279, p318, p179, p247 and p454 are Texas Children's Hospital and TB-05-1939, TB-92-0488 and TB-98-6292 are St Jude Children's Hospital.
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
Institution was parsed from a free-text description field, not a structured characteristic. The paper's author list spans Fred Hutch, St Jude, Texas Children's and Seattle Children's, which is consistent, but the field could name the referring hospital rather than the processing site - and it is the processing site that produces a batch effect.
Evidence that argues against this (1)
- If the description names the referring hospital and all libraries were prepared at Fred Hutch, this would not be an institution batch effect but a patient-population effect - which would be a stronger, not weaker, reason not to compare a same-institution pair to a cohort median. Janssens DH et al. 2024
- E2established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.review
Batch effects in high-throughput data are widespread, are caused by processing conditions, reagent lots and personnel, and become a serious problem specifically when they are correlated with the outcome of interest.
Tackling the widespread and critical impact of batch effects in high-throughput data · Leek JT, Scharpf RB, Bravo HC, Simcha D, Langmead B, Johnson WE, Geman D, Baggerly K, Irizarry RA · Nature Reviews Genetics 11:733-739 2010What this evidence cannot show
A review, cited as a statement of the general principle rather than as evidence about this deposit. The measurement above is this beat's own.
Evidence that argues against this (1)
- Some of the same-institution excess could be patient-population similarity rather than processing similarity - referral patterns concentrate genotypes - in which case the correct comparator set is still institution-matched but the mechanism is different. Leek JT et al. 2010
the confound that explains why a cohort-median comparator flattered the same-patient pair
the cohort median comparator is confounded with institution and the same-patient pair is same-institution by construction
institution parsed from the same deposit's sample descriptions
the institution match is justified by the measured batch effect, not by its result
the interleaved-batch requirement follows from the measured institution effect
the nuisance variable that had to be matched