SPLIT: Spatial Purification of Layered Intracellular Transcripts

🚧 This package is under active development. ❗ Make sure you use the latest version (i.e., v0.2.0).

⚡ Use the Quick Start guide below to get up and running quickly.

🆕🔥 A comprehensive tutorial of running SPLIT on ATERA (full-transcriptome spatial) data and its comparison to Xenium is available as .Rmd and .html (SPLIT v0.2.0 runs in minutes on full-transcriptome data). ❗ Requires SPLIT v0.2.0 or later.

📖 A comprehensive tutorial of running SPLIT on Xenium data is available as .Rmd and .html (<30 min total runtime on a standard PC, incl. 4 min for SPLIT with a peak memory usage of ~21 GB).

🆕🔥 A comprehensive tutorial of running SPLIT on VisiumHD data is available as .Rmd and .html (~30 min total runtime on a standard PC, incl. 10 min for SPLIT with a peak memory usage of ~52 GB). ❗ Requires SPLIT v0.1.2 or later.

---

What's new in v0.2.0

• Annotation-method agnostic: SPLIT no longer depends on RCTD output. It now accepts any deconvolution result — all you need is a cells x cell-types weight matrix, a genes x cell-types reference, and optionally a primary cell-type vector (otherwise inferred as the argmax of the weights).
• Full-transcriptome compatible: SPLIT now scales to large full-transcriptome platforms such as ATERA (~18,000 genes) and runs to completion in minutes.
• VisiumHD and Xenium support remains fully intact and backward compatible.
• Faster and leaner: improved chunked sparse-matrix computation reduces peak memory usage and runtime across all platforms.

---

📦 Installation

To install SPLIT from GitHub:

remotes::install_github("bdsc-tds/SPLIT")

---

🚀 Quick Start

RCTD-based (legacy, backward compatible)

If you already have your dataset as a Seurat object (xe) and RCTD results from doublet-mode decomposition, you can run SPLIT as before:

library(SPLIT)
library(Seurat)

# Post-process RCTD output
RCTD <- SPLIT::run_post_process_RCTD(RCTD)

# Run SPLIT purification
res_split <- SPLIT::purify(
  counts = GetAssayData(xe, assay = "Xenium", layer = "counts"),
  rctd   = RCTD,
  DO_purify_singlets = TRUE
)

# Create a purified Seurat object
xe_purified <- CreateSeuratObject(
  counts    = res_split$purified_counts,
  meta.data = res_split$cell_meta,
  assay     = "Xenium"
)

# Optional: filter, normalize and visualize
xe_purified <- subset(xe_purified, subset = nCount_Xenium > 5)
xe_purified <- xe_purified %>%
  SCTransform(assay = "Xenium") %>%
  RunPCA() %>%
  RunUMAP(dims = 1:20)

UMAPPlot(xe_purified, group.by = "first_type", label = TRUE, repel = TRUE) +
  theme(aspect.ratio = 1)

Annotation-method agnostic (v0.2.0+)

Provide deconvolution weights, a reference matrix, and primary cell-type labels from any annotation tool:

library(SPLIT)
library(Seurat)

# Extract deconvolution weights, primary cell-type vector and reference matrix
# from any deconvolution result (shown here with RCTD for illustration)
new_input <- SPLIT::convert_rctd_result_to_purify_input(rctd = RCTD)

# Run SPLIT purification — returns a SingleCellExperiment object
res_split <- SPLIT::purify(
  counts                = GetAssayData(xe, assay = "Xenium", layer = "counts"),
  reference             = t(new_input$reference),          # genes x cell-types
  primary_cell_type     = new_input$primary_cell_type,     # named character vector
  deconvolution_weights = new_input$deconvolution_weights, # cells x cell-types
  DO_output_sce         = TRUE   # set FALSE to get a plain list instead
)

# Create a purified Seurat object from the SCE output
xe_purified <- CreateSeuratObject(
  counts    = assay(res_split, "purified_counts"),
  meta.data = as.data.frame(colData(res_split)),
  assay     = "Xenium"
)

# Optional: filter, normalize and visualize
xe_purified <- subset(xe_purified, subset = nCount_Xenium > 5)
xe_purified <- xe_purified %>%
  SCTransform(assay = "Xenium") %>%
  RunPCA() %>%
  RunUMAP(dims = 1:20)

UMAPPlot(xe_purified, group.by = "first_type", label = TRUE, repel = TRUE) +
  theme(aspect.ratio = 1)

---

Citation

If you use SPLIT in your work, please cite:

Resolving sensitivity, specificity and signal contamination in Xenium spatial transcriptomics
Mariia Bilous, Daria Buszta, Jonathan Bac, Senbai Kang, Yixing Dong, Stephanie Tissot, Sylvie Andre, Marina Alexandre-Gaveta, Christel Voize, Solange Peters, Krisztian Homicsko, Raphael Gottardo
Nature Methods (2026). https://doi.org/10.1038/s41592-026-03089-8

---

Contact

If you have any questions about the package, feel free to open an issue or contact Mariia Bilous at Mariia.Bilous@chuv.ch.