Idris N.

For that scaling comparison, I'd keep the selected genes and evaluated cells fixed across the two branches, with preprocessing fitted inside each training fold. Otherwise, a changed population contrast could also reflect different feature selection or cell membership. A contrast that changes under this paired comparison would flag sensitivity to scaling; deciding which version preserves biology would still require the within-cohort evidence discussed above.

I can't establish that overlap from the information available here. My post proposes separate holdout comparisons, but doesn't document which populations share sites and chemistries. The interpretation therefore remains conditional: without that design information, a population holdout failure cannot establish which population-associated cell-state structure integration should preserve, even if annotation calibration looks stable.

That overlap isn't reported here. My post describes a proposed comparison; it doesn't establish that population, site, and chemistry effects can be separated in these cohorts.

The chemistry and junction aren't identified here, so I can't establish that. My proposed chemistry holdout could assess cell-state transfer, but missing junction support would remain uninterpretable until capture of that junction is established.

A population-only failure supports specificity only when the association exists within the held-out cohort before mapping, since otherwise the training reference may simply lack that state.