It it possible that slingshot sets different starting clusters for different conditions?

[LilianGomes]

Hi all,
I’m running Slingshot on two experimental conditions and encountering an interesting issue:
depending on the condition, my lineage curves appear to start from different clusters.

Specifically:

Control (ctrl) cells start at cluster “vRG”

Infected cells start at cluster “oRG”

However, when I check the computed lineages, they all seem to start at vRG, regardless of condition.
This raises the question — is this a biological effect or a possible confusion in the Slingshot algorithm due to differences in cell composition between conditions?

$Lineage1
[1] "vRG"           "tRG"           "oRG"           "RG_astrocytic"

$Lineage2
[1] "vRG"          "oRG_dividing" "IPC_IAO"      "vRG_dividing"

$Lineage3
[1] "vRG"              "oRG_dividing"     "IPC_IAO"          "IPC_IAO_dividing"

$Lineage4
[1] "vRG"             "oRG_dividing"    "IPC_IAO"         "IPC_EN_dividing"

$Lineage5
[1] "vRG"          "oRG_dividing" "Astrocytes"  

$Lineage6
[1] "vRG"    "IPC_EN"

$Lineage7
[1] "IPC_EN_late"

here is my code

 # Convert Seurat object to SingleCellExperiment
All_integrated_prog_sce <- as.SingleCellExperiment(All_integrated_prog, assay = "SCT")

# Transfer reductions to SCE
reducedDims(All_integrated_prog_sce) <- list(
  PCA = Embeddings(All_integrated_prog, "pca"),
  UMAP = Embeddings(All_integrated_prog, "umap.rpca")
)

# Run Slingshot with both possible starting clusters
All_integrated_slingshot <- slingshot(
  All_integrated_prog_sce,
  reducedDim = "UMAP",
  clusterLabels = All_integrated_prog_sce$cell_type,
  stretch = 0,
  extend = "n",
  omega = TRUE,
  start.clus = c("vRG", "oRG")
)

# Combine reduced dimensions and metadata
df <- bind_cols(
  reducedDim(All_integrated_slingshot, "UMAP") %>% as.data.frame(),
  colData(All_integrated_slingshot)[, 1:14] %>% as.data.frame()
)

# Run Slingshot by condition
sdss <- slingshot_conditions(
  SlingshotDataSet(All_integrated_slingshot),
  All_integrated_slingshot$exp_group
)

# Extract curve data
curves <- dplyr::bind_rows(
  lapply(sdss, slingCurves, as.df = TRUE),
  .id = "exp_group"
)

# Compute pseudotime and weights
pt <- slingPseudotime(All_integrated_slingshot, na = TRUE) %>%
  as.data.frame() %>%
  rename_with(~ paste0("Lineage", seq_along(.), "_pst"))

w <- slingCurveWeights(All_integrated_slingshot) %>%
  as.data.frame() %>%
  rename_with(~ paste0("Lineage", seq_along(.)))

# Combine everything
df1 <- bind_cols(
  reducedDim(All_integrated_slingshot, "UMAP") %>%
    as.data.frame() %>%
    setNames(c("UMAP_1","UMAP_2")),
  pt, w,
  tibble(
    exp_group = All_integrated_slingshot$exp_group,
    cell_type = All_integrated_slingshot$cell_type
  )
) %>%
  rowwise() %>%
  mutate(
    # Select lineage based on weights only
    max_weight_index = which.max(c_across(matches("^Lineage\\d+$"))),
    # Assign pseudotime from the matching lineage
    pst = c_across(matches("^Lineage\\d+_pst$"))[max_weight_index]
  ) %>%
  ungroup()

Is the observation that control lineages start at “vRG” while infected ones start at “oRG” biologically plausible, or could this be a consequence of Slingshot’s behavior when the starting cluster varies across conditions (e.g., due to differences in cluster composition or ordering in exp_group)? dashed line is the infected condition.