I'm trying to replicate an analysis from a paper ("Single-cell RNA-seq unravels alterations of the human spermatogonial stem cell compartment in patients with impaired spermatogenesis") in which they compare a factor with 2 levels and 3 replicates per level.

Specially, and trying to replicate the RNA velocity results they produced (although I didnt understand well how they integration step) in order to see if I can get similar results before moving onto a dataset never analysed with scvelo or velocyto. I made the 6 .loom files from fastq files using Cellranger and velocytoCLI. I loaded the 6 resulting loom files with Seurat, then integrated three replicates of one condition (and same with the other condition) with anchors-based method and performed the typical downstream analysis in order to get the UMAP embedding. Finally I saved it as a h5 file and loaded in a python script in order to calculate the moments, splicing constants (dynamical model), velocity graph... and finally I ended up with a somehow different low embedding and velocities, despite I tried to replicate the same components, variable genes and clustering resolution used.

So this raised me some points to wonder: 1- It is conceptually ok to run a velocity analysis with several integrated samples? (because althought the dataset is "batch-corrected", the unspliced/spliced RNA ratios of each sample might confound the actual calculations of the velocities. 2 - If so, it is ok to calculate the low embedding in the integrated (or batch-corrected dataset) data but calculate the velocities with the raw quantities of spliced and unspliced reads? (because in the end, velocities, that are calculated with raw data, are integrated in the UMAP embedding, which was calculated with the batch-corrected/integrated data)

Thank you so much in advance.