Sketch-based Method Questions

[majetyneha]

Hi all - we have a very large single-cell RNA dataset (~1.9 mill cells) and have decided to utilize Seurat's Sketch-based workflow to analyze the data. With limited literature using the new technique, we wanted post on here to clarify one part of the pipeline. Our understanding of the standard sketch workflow (per the vignette): LogNormalizeData -> SketchData -> Clustering/integration -> project data back on full dataset. We noticed that the Sketch Vignette does not apply ScTransform and we couldn’t find documentation or discussion addressing why.

Our questions:

1. Was the omission of ScTransform in the Sketch Vignette an intentional design decision or simply not documented/tested yet?
2. Is applying ScTransform to raw counts of the sketch assay (post-sketch - created with leverage score sampling computed from log-normalized representation, pre-clustering) a methodologically sound approach for deriving the normalized values used in downstream clustering (in place of using the sketch assay’s inherited log-normalized data layer)?
3. Since the leverage score sampling computed from log-normalized representation deliberately oversamples rare populations, could this skew the ScTransform’s per-gene regularized model in a way that wouldn’t occur if if ScTransform was run on a full dataset? Is this a relevant concern, or is ScTransform’s per-gene fitting robust to non-uniform sampling?
4. If we wanted to validate our 100k-cell ScTransform on sketch results against full-dataset ScTransform is there a recommended approach if we have had scalability issues (we hit the 2^31 entry limit on R if we try to ScTransform on full dataset, but resolved this with doing ScTransform on sketched data) or is this not possible with current sketch infrastructure?

Appreciate the help!