split-purification

Run SPLIT::purify() on your Xenium count matrix together with an RCTD object from spacexr. Match cell barcodes between the two inputs first, then create a purified Seurat object from the resulting purified counts and process it with standard normalization and UMAP steps.

RCTD performs cell type deconvolution and annotation, while SPLIT purifies the count data by removing transcripts contaminated by neighboring cells. SPLIT requires RCTD results as input, so RCTD must be run first before purification.

No, SPLIT requires the spacexr package installed from GitHub (dmcable/spacexr), not the Bioconductor version. If you encounter RCTD compatibility errors, remove the Bioconductor package and reinstall from GitHub.

Cell barcode formats often differ between the spatial object and RCTD results, such as added prefixes. Inspect the barcode formats of both, strip prefixes with pattern matching, and build a mapping table before intersecting the cell sets.

A Xenium dataset of about 20K cells takes roughly 4 minutes and 21 GB of memory, while VisiumHD with about 100K cells takes around 10 minutes and 52 GB. Reduce chunk_size and n_workers or raise future.globals.maxSize if you hit memory errors.

Yes, retaining 40-70% of cells is typical because SPLIT discards contaminated cells to improve data quality. The trade-off yields improved cell type specificity, clearer UMAP separation, and sharper spatial boundaries for downstream analysis.