Biology or handling? The deposit already contained the control arm we said it did not have.

The strongest non-lineage mover of our headline marker programme is FOS, with JUNB behind it. Both are immediate-early genes. The textbook objection is that immediate-early expression in single-cell data is largely a tissue-handling artefact: cells sense dissociation and mount a stress response on the way to the sequencer. If that is what our comparator is made of, it cannot adjudicate anything.

We retired the specificity claim on that basis in early August, which was the right verdict reached the wrong way. A retraction justified by a suspect comparator is no better founded than the claim it replaced. Both were arguments from the literature about a dataset neither had measured.

The arm was in the metadata

An earlier audit concluded that this could not be attributed to handling on the metadata to hand, because the reference had no protocol split. It had one. The per-sample technology characteristic in the public deposit splits the 47 RNA records into 11 single-cell libraries and 36 nuclei libraries, the latter being 26 dedicated single-nucleus preparations plus 10 multiome. The processed object we had been reading collapses single-cell and single-nucleus into one label, which is exactly the distinction the experiment turns on, so the split looked unavailable while sitting one field away.

Nuclei are the right negative arm for this question. A nucleus cannot mount a cytoplasmic-signalling-driven transcriptional response. The arm labels also check out physically without trusting the annotation at all: median mitochondrial content is 9.88% in the cells arm against 0.73% in nuclei, and ribosomal content 14.64% against 0.78%.

The result

Scoring an 11-gene immediate-early panel per library and comparing the arms, the panel falls in nuclei at Cliff's delta 0.959596, a mean difference of 0.628486, Cohen's d 3.46 and Mann-Whitney p 1.9e-06. Those first two numbers are worth keeping apart from each other, and we will come back to why.

The control that makes this interpretable is the identity panel: the same three lineage marker programmes, scored the same way, on the same libraries. If the arms differed on those as much as on the stress panel, the whole thing would be global platform deflation and would say nothing about dissociation.

programme            Cliff's delta        p
immediate-early           0.9596    1.9e-06
OPC-like markers          0.2778      0.171
AC-like markers          -0.0859      0.678
OC-like markers          -0.3535      0.081

None of the three lineage programmes separates the arms, and one of them goes the other way. Whatever is different between cells and nuclei here, it is not a uniform loss of signal.

The confound, and the two libraries that are free of it

Two donors were assayed on both protocols. Within them, the same direction holds at Cliff's delta 0.885 and 0.828. Those are the unconfounded figures and they are the ones to quote, with the obvious caveat attached: two pairs is two pairs. No across-pair test is computed, because the smallest two-sided sign-test p at n=2 is 0.5 and would be uninformative whichever way the data fell.

Four ways it could still be something else

• Composition. The arms hold different proportions of each cell state. Standardising to a common composition gives 0.632 against the raw 0.628, and the effect is positive in all four states separately, so composition explains none of it.
• Depth. Nuclei carry less RNA. Restricting to the overlapping library-size window leaves the cell-level effect at 0.783 against 0.771 unrestricted. Note the unit: that pair is cell-level and is not comparable to the 0.96, which is one mean per library.
• Cohort vintage. Dropping the oddly-dated samples leaves Cliff's delta 0.951 at p 1.3e-05.
• Multiome rather than nuclei. The two nuclei technologies are indistinguishable on this panel, at Cliff's delta 0.146 and p 0.51, while the same contrast does find a significant platform difference on one of the identity programmes. So the null is not an absence of power.

There is also a negative control: nuclei against nuclei, same donor, crossing a platform boundary without crossing the cells-versus-nuclei boundary. Its floor is not zero. One of the two pairs separates at Cliff's delta -0.548. It separates in the opposite direction to what the primary comparison needs, the two control pairs disagree with each other on direction, and both primary effects are larger than the largest control effect.

What it does not settle

Nuclei preparation is not a dissociation-free control. The tissue is still dissociated before nuclei are isolated. What the nuclei arm removes is the cell's ability to mount and retain the response, and any transcript that had already left the nucleus. It does not remove the insult. So this shows that the signal needs intact cells, not that gentler handling would abolish it.

It also has an unmeasured counter-argument pointing the other way: newly induced transcription is a nuclear event, so if anything it should be more visible in nuclei. Measuring the intronic fraction would settle that, and we have not done it.

One number for scale

Detection rates on the same reference are a useful sanity check on any argument built from this panel. JUN is detected in 43.95% of cells, FOS in 40.70%, EGR1 in 28.21%, JUNB in 25.34%. Our headline lineage regulator, OLIG2, is detected in 15.31%. The immediate-early genes are among the best-detected things on the panel, which is part of why they turn up at the top of an ordering that is sensitive to how many edges a gene has.