CellRank and Palantir fate workflow

This workflow keeps the Palantir pseudotime direction and the CellRank fate model as separate, inspectable result layers. Pseudotime is a relative ordering and fate probabilities are model-based transition probabilities; neither is a direct experimental time course.

The input overview below is executed on all bundled cells. The Python analysis sections are code-only recipes and do not claim fitted Palantir or CellRank results from this website build. Run them in a prepared Python environment in the order shown; the overview embedding alone is not evidence of cell fate.

Inspect the input

library(scop)
data(pancreas_sub)
input_cells <- colnames(pancreas_sub)
pancreas_sub <- RunStandardWorkflow(pancreas_sub,
  assay = "RNA", seed = 11,
  cores = 1, verbose = FALSE
)
stopifnot(identical(colnames(pancreas_sub), input_cells))
table(pancreas_sub$SubCellType)

## 
##         Alpha          Beta         Delta        Ductal       Epsilon 
##           121           174            20           253            40 
##  Ngn3-high-EP   Ngn3-low-EP Pre-endocrine 
##           169            67           156

CellDimPlot(pancreas_sub, group.by = "SubCellType", reduction = "UMAP", label = TRUE)

FeatureDimPlot(pancreas_sub,
  features = c("Sox9", "Neurog3", "Gcg", "Ins2"),
  reduction = "UMAP", ncol = 2
)

Palantir direction and MAGIC layer

The Python reference backend reuses the SCOP environment selected by PrepareEnv(). Existing Python 3.10/3.11 environments receive the compatible legacy pins, while Python 3.12 environments receive the current pins. MAGIC is optional and should be used for visualization or trend fitting, never as a replacement for raw counts in differential testing.

pancreas_sub <- RunPalantir(
  pancreas_sub,
  group.by = "SubCellType",
  linear_reduction = "PCA",
  nonlinear_reduction = "UMAP",
  early_group = "Ductal",
  terminal_groups = c("Alpha", "Beta", "Delta", "Epsilon"),
  magic_impute = TRUE,
  backend = "python"
)
FeatureDimPlot(
  pancreas_sub,
  features = c("palantir_pseudotime", "palantir_diff_potential")
)
PalantirTrajectoryPlot(pancreas_sub, reduction = "UMAP")

CellRank fate mapping

Use the stored Palantir pseudotime to construct a directed PseudotimeKernel. The explicit Schur size makes the small tutorial run reproducible. The computed macrostate names are printed before any optional manual terminal-state rerun; unknown terminal states stop before the Seurat object is mutated.

pancreas_sub <- RunCellRank(
  pancreas_sub,
  group.by = "SubCellType",
  kernel_type = "pseudotime",
  time_key = "palantir_pseudotime",
  # `brandts` is portable for small tutorial data; use `krylov` only after
  # installing PETSc/SLEPc in the selected environment.
  schur_method = "brandts",
  schur_n_components = 10L,
  n_macrostates = 8L,
  n_cells_terminal = 10L,
  backend = "python"
)
names(pancreas_sub@tools$CellRank)
pancreas_sub@tools$CellRank$states$macrostate_names
CellRankPlot(pancreas_sub, plot_type = "fate", reduction = "UMAP")
CellRankPlot(pancreas_sub, plot_type = "circular")

Drivers, trends and enrichment

Drivers are statistical associations with fate probabilities. Trend modules are similar expression-shape clusters, not causal regulatory modules.

Use the common dynamic-feature workflow to fit and display expression along Palantir pseudotime. These fits use RunDynamicFeatures(); the separate CellRank GAM workflow below uses fate-probability weights.

pancreas_sub <- RunDynamicFeatures(
  pancreas_sub,
  lineages = "palantir_pseudotime",
  features = c("Rbp4", "Sst", "Pyy"),
  assay = "RNA"
)
DynamicHeatmap(
  pancreas_sub,
  lineages = "palantir_pseudotime",
  features = c("Rbp4", "Sst", "Pyy"),
  use_fitted = TRUE,
  assay = "RNA",
  cell_annotation = "SubCellType"
)
DynamicPlot(
  pancreas_sub,
  lineages = "palantir_pseudotime",
  features = c("Rbp4", "Sst", "Pyy"),
  assay = "RNA",
  group.by = "SubCellType"
)

pancreas_sub <- RunCellRankTrends(
  pancreas_sub,
  lineage = "Alpha",
  assay = "MAGIC_imputed_data",
  layer = "counts",
  top_n = 500L,
  heatmap_n = 50L,
  n_points = 200L,
  output_dir = "cellrank_objects"
)
CellRankPlot(pancreas_sub, plot_type = "drivers", lineage = "Alpha")

pancreas_sub <- RunCellRankEnrichment(
  pancreas_sub,
  lineage = "Alpha",
  species = "Mus_musculus",
  output_dir = "cellrank_objects"
)
pancreas_sub@tools$CellRank$enrichment$Alpha$manifest
EnrichmentPlot(
  res = list(enrichment = pancreas_sub@tools$CellRank$enrichment$Alpha$databases$GO_BP),
  db = "GO_BP",
  plot_type = "comparison"
)

After executing the Python sections, the complete payload is stored in srt@tools$Palantir and srt@tools$CellRank, including the actual backend, versions, cell/lineage names, transition matrix, fate probabilities, drivers, trend curves, modules, enrichment tables and run parameters.