Tutorial 03 — scRNA-seq: Markers & Cell Type Annotation

Find cluster markers, visualize, label by hand and check against ground truth

Author

Single Cell RNA-seq Workshop

Running the code in this tutorial

Every code chunk here is tagged with #| eval: false, so the published page shows the code without running it. To run it yourself, open the downloaded .qmd in RStudio:

Run one chunk: click the green ▶ (Run Current Chunk) at the chunk’s top-right, or press Ctrl/Cmd + Enter. This runs the chunk regardless of its eval setting — the easiest way to work through the tutorial interactively.
Run a chunk on render: change that chunk’s #| eval: false to #| eval: true (or delete the line) so it executes when the document is rendered.
Render the whole tutorial: click the Render button in RStudio (or run quarto render in a terminal) to execute the chunks top-to-bottom and knit a finished HTML. To run everything on render, set eval: true once in the YAML header at the top.

About this tutorial

Part three of the series. We start from the unintegrated clustered object saved in Tutorial 02 and walk through:

Note

Companion lecture: Lecture 03 — Markers & Manual Cell-Type Annotation · Companion book chapter: Chapter 3 — Markers & Manual Annotation — the long-form prose treatment of this tutorial’s material, with cross-references to the prerequisite appendices. · HPC version (Talapas): Talapas analysis pipeline — 03_markers_annotation.R

Finding marker genes per cluster
The canonical marker visualizations (Vln / Feature / Heatmap / Ridge / Dot)
Manual cell-type annotation with RenameIdents
Cross-checking your manual labels against the ground-truth seurat_annotations that ship with ifnb
Saving the annotated object so Tutorial 04 (Reference Annotation) can pick it up

Note

We’re annotating the unintegrated object. Why? Because markers are cell-type-driven, not batch-driven — even when an embedding is dominated by a batch effect, the genes that distinguish T cells from B cells are still T-cell genes. Working markers up before integration is a useful sanity check: if a marker gene cleanly identifies a cluster on the unintegrated data, it should still cleanly identify the biologically equivalent cluster after integration in Tutorial 05.

How to use this page

The rendered HTML shows the code but does not execute it. To run it:

Download the .qmd source: Tutorial_03_Markers_Annotation.qmd. If your browser saves the file as Tutorial_03_Markers_Annotation.qmd.txt, drop the trailing .txt so the filename ends in .qmd, then open it in RStudio.
Make sure you have completed Tutorial 02 — it writes the ifnb_clustered.rds that this tutorial loads.
Work through the chunks, flipping eval: true as you go.

Dataset — resuming from Tutorial 02

File	Produced by	What it is
`../data/ifnb_clustered.rds`	Tutorial 02	Seurat object with PCA, SNN graph, clusters at five resolutions, UMAP, plus the `stim` and `seurat_annotations` metadata.

Learning goals

Use FindMarkers / FindAllMarkers appropriately
Read each of the five standard marker-visualization plots and know when to pick which
Assign cell-type labels manually using PBMC canonical markers
Validate against the seurat_annotations ground truth and quantify agreement

Common errors / things that bite

FindAllMarkers takes 5+ minutes — presto (installed in the Tutorial 00 setup) makes it ~10× faster; Seurat detects it automatically.

ISGs (ISG15, IFI6, IFIT3, MX1) top some clusters’ marker lists — expected on the unintegrated ifnb data: the cell types with the strongest interferon response (monocytes especially) split into CTRL vs STIM clusters, so those clusters’ top markers are condition genes rather than cell-type genes. Most clusters are still topped by cell-type markers. Tutorial 05 (Integration) collapses the condition split.

names(new.cluster.ids) <- levels(ifnb) errors with “‘names’ attribute [N] must be the same length as the vector [M]” — your clustering produced more clusters than the template’s label list. This is real biology, not a bug. You’re clustering the unintegrated data, where the interferon-β response is a strong, genome-wide transcriptional signal — so the cell types that respond most (monocytes above all) separate into two clusters, a CTRL one and a STIM one, instead of one. That’s the very same condition/batch split you visualized on the UMAP in Tutorial 02, now showing up as extra clusters: you commonly end up with ~15–20 clusters for ~13 cell types, often with two clusters of the same type. It’s expected, and Tutorial 05’s integration aligns CTRL and STIM so each cell type collapses back to a single cluster. Practically: the Step 4 template no longer uses that brittle line — it maps labels by cluster id and applies only the ids that exist, so it works for any cluster count. Run levels(ifnb) to see how many you have, and label each (giving the CTRL and STIM copies of a type the same label is fine).

Idents() is set to the wrong column — clusters at multiple resolutions all live in meta.data (RNA_snn_res.0.1, _0.3, etc.) but only one is the “active” identity. Idents(seu) <- "RNA_snn_res.0.5" switches; levels(Idents(seu)) confirms.

FindMarkers errors with “data layers are not joined. Please run JoinLayers” — Seurat v5 keeps per-sample split layers (counts.CTRL, counts.STIM, data.CTRL, data.STIM) after merge(), and FindMarkers refuses to run until they’re collapsed. Fix with ifnb <- JoinLayers(ifnb) before calling FindMarkers / FindAllMarkers. The resume chunk above does this automatically. Layers(ifnb) shows the current layer list.

Solutions / instructor copy

The Think about it prompts on this page have inline answers (click “Show answer” / “Click for answer” to expand). For the full runnable instructor copy with every chunk pre-evaluated and outputs shown:

For students: the same .qmd you downloaded already contains every solution inline — just expand the collapsed answers
For instructors: render with quarto render --profile solutions from the Exercise_Folder/ directory. The fully-evaluated HTML is written to docs/Exercise_Folder/_solutions/Tutorial_03_Markers_Annotation.html. See Exercise_Folder/_quarto-solutions.yml for the build profile

Setup

library(Seurat)
library(tidyverse)
library(patchwork)   # combine + annotate multi-panel Seurat plots

set.seed(2026)

# ---------------------------------------------------------------------------
# Output directory for this module's figures and tables.
# Every figure/table chunk below writes a file named  Mod3_C<chunk>_<name>
# into ../output/Mod3/ so it can be cross-referenced from the rest of the site.
#   Mod3 = Module 3 (this tutorial);  C<n> = the nth code chunk.
# ---------------------------------------------------------------------------
out_dir <- "../output/Mod3"
dir.create(out_dir, recursive = TRUE, showWarnings = FALSE)
dir.create("../data",  showWarnings = FALSE)   # shared .rds hand-off folder (sibling of scripts/, created one level up)
message("This module writes its figures & tables to: ", normalizePath(out_dir))

# --- Figure saving --------------------------------------------------------
# save_fig() writes every figure as a .png (for viewing) AND a .svg (vector,
# editable in Illustrator / Inkscape for a manuscript). Pass the .png path; the
# .svg is written alongside with the same basename. (.svg uses the 'svglite'
# package, installed in Tutorial 00.)
save_fig <- function(filename, plot, width, height, dpi = 300, ...) {
  ggplot2::ggsave(filename, plot, width = width, height = height, dpi = dpi, ...)
  svg_path <- paste0(tools::file_path_sans_ext(filename), ".svg")
  tryCatch(ggplot2::ggsave(svg_path, plot, width = width, height = height, ...),
           error = function(e) message("  (could not write ", basename(svg_path),
                                       " - install 'svglite'? ", conditionMessage(e), ")"))
}

ifnb <- readRDS("../data/ifnb_clustered.rds")
Idents(ifnb) <- "RNA_snn_res.0.5"

# In Seurat v5, merging per-sample objects (as Tutorial 01 does after
# scDblFinder) leaves the RNA assay with split layers. FindMarkers requires
# a single joined layer, so rejoin here if it hasn't been done already.
if (length(Layers(ifnb, search = "counts")) > 1) {
  ifnb <- JoinLayers(ifnb)
}

ifnb

Step 1 — Find markers for one cluster

cluster1.markers <- FindMarkers(ifnb, ident.1 = 1, min.pct = 0.25)
head(cluster1.markers, 10)

# --- Table out: top markers for cluster 1 (gene moved from rownames) -------
cluster1.markers |>
  tibble::rownames_to_column("gene") |>
  readr::write_csv(file.path(out_dir, "Mod3_C3_cluster1_markers.csv"))

Reading the output

Each row is a gene; the five columns are: p_val (raw Wilcoxon p-value), avg_log2FC (log₂ fold-change, cluster 1 vs all other cells), pct.1 (fraction of cluster-1 cells expressing the gene), pct.2 (fraction of all other cells expressing it), and p_val_adj (Bonferroni-corrected p-value). The best specific markers have high avg_log2FC, high pct.1 (expressed in most of your cluster), and low pct.2 (rare outside it) — genes with high pct.2 may be significant but are not cluster-specific. Rows are sorted by p-value by default, not fold-change, so scroll or re-sort by avg_log2FC to find the strongest effect-size markers.

Think about it

What does min.pct = 0.25 do, and why is that a useful default?
The output has avg_log2FC, pct.1, pct.2. Which gene would you pick as a specific marker for this cluster?

Show answers

Only tests genes detected in ≥ 25% of cells in at least one of the two groups. Prunes low-expression noise and speeds the test up significantly.
A specific marker has high pct.1 and low pct.2 (expressed in almost all cluster-1 cells, almost none elsewhere) and high positive avg_log2FC. Markers that are broadly expressed (high pct.2) may be real but aren’t specific.

Step 2 — Markers for every cluster

ifnb.markers <- FindAllMarkers(
  ifnb,
  only.pos        = TRUE,
  min.pct         = 0.25,
  logfc.threshold = 0.25
)

top_markers <- ifnb.markers |>
  group_by(cluster) |>
  slice_max(avg_log2FC, n = 5)
top_markers

# --- Tables out: all positive markers + the per-cluster top 5 --------------
readr::write_csv(ifnb.markers, file.path(out_dir, "Mod3_C4_all_markers.csv"))
readr::write_csv(top_markers,  file.path(out_dir, "Mod3_C4_top5_markers_per_cluster.csv"))

Reading the output

top_markers is a tibble grouped by cluster, showing the five genes with the highest avg_log2FC in each cluster. Scan the gene column cluster-by-cluster: most clusters will be topped by recognizable PBMC markers (LYZ/S100A8 for monocytes, MS4A1/CD79A for B cells, GNLY/NKG7 for NK, IL7R/CCR7 for naive T). If a cluster’s top genes are interferon-stimulated genes (ISG15, IFI6, IFIT3, MX1) rather than cell-type markers, that cluster is likely a condition-split copy of a cell type rather than a biologically distinct population — expected on unintegrated data and resolved in Tutorial 05.

Think about it

Most of the top per-cluster markers are cell-type genes (e.g. LYZ/S100A8 for monocytes, MS4A1/CD79A for B cells, GNLY/NKG7 for NK). But for some clusters — especially the monocytes — you’ll also see interferon-stimulated genes (ISG15, IFI6, IFIT3, MX1) near the top. If those ISGs aren’t cell-type markers, what are they marking?
For condition-level DE (STIM vs CTRL), is FindAllMarkers the right tool?

Show answers

Those ISGs are condition markers — they distinguish IFN-β–stimulated cells from control cells, not cell types. They surface because, on the unintegrated data, the cell types with the strongest interferon response (monocytes most of all) split into separate CTRL and STIM clusters — so a “STIM monocyte” cluster’s top markers are ISGs. Most clusters are still driven by cell-type identity, which is why cell-type markers dominate the lists overall. Tutorial 05 integrates CTRL and STIM so clusters track biology, not condition, and the ISGs drop out of the per-cluster marker lists.
No. FindAllMarkers / FindMarkers treat each cell as an independent observation, which is anti-conservative for condition-level comparisons across biological samples. For STIM vs CTRL within a cell type, use pseudobulk + DESeq2/edgeR/limma — that’s exactly what Tutorial 06 covers, on this dataset.

Step 3 — Visualizing markers

PBMC canonical markers — useful to keep handy:

Cell type	Canonical markers
CD14+ Monocyte	`CD14`, `LYZ`, `S100A8`, `S100A9`
CD16+ Monocyte	`FCGR3A`, `MS4A7`, `CDKN1C`
Dendritic cell (DC)	`FCER1A`, `CST3`, `CLEC10A`
pDC	`LILRA4`, `IL3RA`, `CLEC4C`
CD4 Naive T	`IL7R`, `CCR7`, `LEF1`
CD4 Memory T	`IL7R`, `S100A4`
CD8 T	`CD8A`, `CD8B`, `GZMK`
NK	`GNLY`, `NKG7`, `KLRD1`
B	`MS4A1`, `CD79A`, `CD79B`
T-activated	`CD69`, `IL2RA`
Mk (megakaryocyte)	`PPBP`, `PF4`
Eryth	`HBA1`, `HBB`

3a. Violin plots — expression distribution per cluster

p_vln <- VlnPlot(ifnb,
        features = c("CD14", "LYZ", "MS4A1", "CD79A", "CD8A", "GNLY"),
        pt.size = 0, ncol = 3) &
  xlab("Cluster (RNA_snn_res.0.5)") &
  ylab("Log-normalized expression") &
  theme(plot.title = element_text(size = 11))
p_vln <- p_vln +
  patchwork::plot_annotation(
    title    = "Canonical PBMC marker expression by cluster",
    subtitle = "Monocyte (CD14/LYZ), B (MS4A1/CD79A), CD8 T (CD8A) and NK (GNLY) markers",
    caption  = "Module 3 · Markers & Annotation"
  )
p_vln

# --- Figure out ------------------------------------------------------------
save_fig(file.path(out_dir, "Mod3_C5_marker_violins.png"), p_vln,
       width = 10, height = 6, dpi = 300)

Reading the output

Each panel is one marker gene; x-axis = cluster number, y-axis = log-normalized expression. The width of a violin at a given height represents how many cells in that cluster have that expression level, so a tall, narrow spike means most cells in the cluster are consistently expressing the gene. Look for clusters with a violin that rises visibly higher than all others for a given marker — for example, one or two clusters dominating CD14 identify the monocyte clusters, and one or two dominating MS4A1/CD79A identify B cells. Clusters with flat (near-zero) violins for all six markers are likely T cells or NK cells not covered by this panel — check CD8A and GNLY for those.

3b. Feature plots — expression on UMAP

p_feature <- FeaturePlot(ifnb,
            features = c("CD14", "MS4A1", "CD8A", "GNLY"),
            ncol = 2) &
  xlab("UMAP 1") &
  ylab("UMAP 2") &
  theme(plot.title = element_text(size = 11))
p_feature <- p_feature +
  patchwork::plot_annotation(
    title    = "Marker expression overlaid on the UMAP",
    subtitle = "CD14 (monocytes), MS4A1 (B), CD8A (CD8 T), GNLY (NK)",
    caption  = "Module 3 · Markers & Annotation"
  )
p_feature

# --- Figure out ------------------------------------------------------------
save_fig(file.path(out_dir, "Mod3_C6_marker_featureplots.png"), p_feature,
       width = 9, height = 8, dpi = 300)

Reading the output

Each of the four panels places every cell at its UMAP coordinates, coloring it from light (low expression) to dark blue/purple (high expression). Healthy marker geography looks like a bright patch localized to one or a few islands — CD14 should light up the monocyte region, MS4A1 the B-cell island, CD8A the cytotoxic T region, and GNLY the NK region. Cells that appear bright in two non-adjacent islands may indicate doublets or a cross-reactive probe; diffuse high expression across the whole UMAP usually means the gene is broadly expressed and not a good discriminating marker for annotation.

3c. Heatmap — top markers across clusters

top10 <- ifnb.markers |> group_by(cluster) |> slice_max(avg_log2FC, n = 10)

# Tutorial 01 only scaled the HVGs; re-scale the top-marker genes so any
# non-HVG markers also appear in the heatmap rather than being dropped.
ifnb <- ScaleData(ifnb, features = unique(top10$gene))

p_heatmap <- DoHeatmap(ifnb, features = top10$gene) +
  NoLegend() +
  labs(
    title    = "Top-10 markers per cluster — scaled expression heatmap",
    subtitle = "Each column is a cell, grouped by cluster; rows are marker genes",
    caption  = "Module 3 · Markers & Annotation"
  )
p_heatmap

# --- Figure out ------------------------------------------------------------
save_fig(file.path(out_dir, "Mod3_C7_marker_heatmap.png"), p_heatmap,
       width = 12, height = 10, dpi = 300)

Reading the output

Columns are individual cells grouped by cluster (cluster boundaries are marked by thin colored bars on the x-axis); rows are the top-10 marker genes for each cluster. Color encodes z-scored expression: red = high relative expression, blue = low. A well-behaved marker block shows a bright red diagonal band — each set of marker rows lights up in exactly its own cluster and is blue elsewhere. Marker rows that stay red across many clusters are broadly expressed genes (less informative for annotation); rows that are consistently blue everywhere except one cluster are the most specific markers. Note that because columns represent cells rather than aggregated clusters, large clusters dominate the x-axis width, which can make small populations hard to see — the dot plot (code) normalizes for cluster size.

3d. Dot plot — % expressing × mean expression

canonical_markers <- c("CD14", "LYZ", "FCGR3A", "MS4A7", "FCER1A", "CST3",
                       "IL7R", "CCR7", "S100A4", "CD8A", "GZMK", "GNLY",
                       "NKG7", "MS4A1", "CD79A", "PPBP")
p_dot <- DotPlot(ifnb, features = canonical_markers) +
  RotatedAxis() +
  labs(
    title    = "Canonical PBMC marker panel across clusters",
    subtitle = "Dot size = % of cells expressing; colour = scaled average expression",
    x        = "Marker gene",
    y        = "Cluster (RNA_snn_res.0.5)",
    colour   = "Avg. expression",
    size     = "Percent expressed"
  )
p_dot

# --- Figure out ------------------------------------------------------------
save_fig(file.path(out_dir, "Mod3_C8_marker_dotplot.png"), p_dot,
       width = 10, height = 6, dpi = 300)

Reading the output

The grid has clusters on the y-axis and marker genes on the x-axis. Each dot encodes two things simultaneously: size = the fraction of cells in that cluster expressing the gene (larger = more cells expressing it), and color = the scaled average expression level (darker = higher). To identify a cluster, look for the column(s) where it shows large, dark dots — those are its marker genes. A cluster with a big dark dot on CD14 and LYZ is a monocyte; one showing MS4A1 and CD79A is a B cell. Clusters with no large dots across the whole panel need additional markers (check IL7R, CD3D, or CD3E for T cells, which aren’t in this panel). This is the primary visualization to use when deciding labels in code below.

Think about it

Which plot is best for showing a 16-gene marker panel across 12 clusters on one page?
Which is best for checking whether a single gene has spatial structure on the UMAP?
Which is most misleading if your cluster sizes are very unequal, and why?

Show answers

Dot plot — compact, shows both percent-expressing (dot size) and mean expression (color) in a genes × clusters grid.
Feature plot — overlays expression on the 2-D embedding so you can see geography.
Heatmap can be misleading: it shows cells, not clusters, so a dominant cluster (say 40% of cells) visually monopolizes the plot. Dot plot normalizes per cluster and avoids this.

Step 4 — Manual annotation

Use the canonical marker table above and the dot plot (code) to assign labels. Your cluster count and numbering depend on the random seed, Seurat version, and the fact that this is unintegrated data — so you’ll often have more clusters than cell types (e.g. 15–20, with two clusters of the same type that split CTRL vs STIM). Do not copy these labels blindly — run levels(ifnb) to see your clusters, then label each one from the table that matches your dot plot (it’s fine to give two clusters the same label).

# First, see how many clusters you actually have, and their ids:
levels(ifnb)
# On the UNINTEGRATED data, cell types often split by CONDITION (e.g. CTRL vs
# STIM monocytes), so you'll usually get MORE clusters than cell types — often
# ~15-20, sometimes two clusters of the same type. That's expected here;
# Tutorial 05 integrates the samples and collapses these.

# TEMPLATE — give each cluster id a label. Edit this named vector so the names on
# the LEFT (cluster ids, in quotes) cover YOUR clusters from the dot plot. The
# left side is the cluster id, the right side is the label you assign. You can
# reuse a label (e.g. two monocyte clusters both "CD14 Mono"). It's fine to label
# only some clusters while you iterate — unlisted ones keep their number.
new.cluster.ids <- c(
  "0"  = "CD14 Mono",
  "1"  = "CD4 Naive T",
  "2"  = "CD4 Memory T",
  "3"  = "B",
  "4"  = "CD8 T",
  "5"  = "NK",
  "6"  = "CD16 Mono",
  "7"  = "T activated",
  "8"  = "DC",
  "9"  = "Mk",
  "10" = "B activated",
  "11" = "pDC",
  "12" = "Eryth"
  # ... add entries for any remaining cluster ids your run produced (13, 14, ...)
)

# Apply only the labels whose cluster id actually exists. This keys on the vector
# NAMES (cluster ids), so it works for any number of clusters and never throws the
# "'names' attribute [N] must be the same length as the vector [M]" error.
new.cluster.ids <- new.cluster.ids[names(new.cluster.ids) %in% levels(ifnb)]
ifnb <- RenameIdents(ifnb, new.cluster.ids)
ifnb$celltype_manual <- Idents(ifnb)

DimPlot(ifnb, reduction = "umap", label = TRUE, repel = TRUE) + NoLegend()

Reading the output

The UMAP now shows your cell-type labels instead of cluster numbers — each island should carry a readable name. A well-annotated map has compact, non-overlapping label positions (aided by repel = TRUE) and each island dominated by a single biological category. If two neighboring islands share the same label, that is expected on unintegrated data (the same cell type split by condition); if a large island has a label that doesn’t feel right (e.g. “CD14 Mono” sitting next to the T-cell area), revisit the dot plot (code) and check whether a different marker combination fits better.

Think about it

Two clusters share CD3D / CD3E as top markers. Are they the same cell type?
What’s the risk of assigning a label from a single marker gene?

Show answers

Not necessarily — CD3D/E are pan-T markers. Both clusters are likely T cells, but could be CD4 Naive vs CD4 Memory vs CD8 vs T-activated. You need discriminating markers (IL7R+CCR7 for naive, S100A4 for memory, CD8A for cytotoxic, IL2RA for activated) on top of the shared T-cell markers.
Many markers are non-specific, and biological “marker genes” are often graded rather than binary. Label from combinations (a marker plus absence of another marker) and always cross-check.

Step 5 — Validate against the ground-truth labels

The ifnb dataset ships with author-curated labels in seurat_annotations. This is rare in real life — but it makes a beautiful teaching moment to compare your manual labels to a published ground truth.

# Side-by-side UMAPs
p_manual <- DimPlot(ifnb, reduction = "umap", label = TRUE, repel = TRUE) +
  NoLegend() +
  labs(
    title  = "Manual annotation (yours)",
    x      = "UMAP 1",
    y      = "UMAP 2"
  )
p_truth  <- DimPlot(ifnb, reduction = "umap", group.by = "seurat_annotations",
                    label = TRUE, repel = TRUE) +
  NoLegend() +
  labs(
    title  = "Author labels (ground truth)",
    x      = "UMAP 1",
    y      = "UMAP 2"
  )

p_manual
p_truth

# Combined two-panel manual-vs-truth comparison, shared annotation
p_validate <- (p_manual | p_truth) +
  patchwork::plot_annotation(
    title    = "Manual annotation vs author ground truth on the same embedding",
    subtitle = "Cells should land in matching positions if your manual labels agree with the authors'",
    caption  = "Module 3 · Markers & Annotation"
  )

# Cross-tabulate. Each row should ideally concentrate on a single column.
manual_vs_truth <- table(manual = Idents(ifnb), truth = ifnb$seurat_annotations)
manual_vs_truth

# --- Figures out -----------------------------------------------------------
save_fig(file.path(out_dir, "Mod3_C10_umap_manual.png"),         p_manual,
       width = 7,  height = 6, dpi = 300)
save_fig(file.path(out_dir, "Mod3_C10_umap_ground_truth.png"),   p_truth,
       width = 7,  height = 6, dpi = 300)
save_fig(file.path(out_dir, "Mod3_C10_umap_manual_vs_truth.png"), p_validate,
       width = 13, height = 6, dpi = 300)

# --- Table out: cross-tab of manual labels vs ground-truth labels ----------
as.data.frame(manual_vs_truth) |>
  readr::write_csv(file.path(out_dir, "Mod3_C10_manual_vs_truth.csv"))

Reading the output

The two UMAP panels side-by-side show your labels (left) and the author ground-truth (right) on the same embedding: identical islands in matching positions mean agreement; an island that carries a different name in each panel is a disagreement. The manual_vs_truth cross-tab printed below has your labels as rows and the author labels as columns — each cell is a count of cells assigned that label pair. A perfect annotator would see counts concentrated on a diagonal (each of your labels maps cleanly to one author label); off-diagonal counts identify the specific cell types where your calls and the authors’ disagree.

Think about it

Where do you and the authors agree? Where do you disagree?
Which cell type was easiest to call from markers alone? Hardest?
In Tutorial 05, after integration, the same cluster numbers will mean different things. Why is seurat_annotations (a per-cell label) more durable than your RNA_snn_res.0.5 manual labels (a per-cluster mapping)?

Show answers

Big myeloid (CD14 / CD16 / DC) and lymphoid (B / NK) clusters usually agree perfectly. T-cell subdivisions (CD4 Naive vs CD4 Memory vs T activated) and rare populations (pDC, Eryth, Mk) are the common disagreements — they require multi-gene logic and have soft boundaries.
Easiest: B cells (MS4A1+, CD79A+, otherwise quiet) and NK (GNLY+, NKG7+, no CD3). Hardest: the various CD4 T subsets, because they share most markers and differ in subtle, graded ways.
Per-cell labels travel with the cell — if a cell is CD14 Mono in seurat_annotations here, it stays CD14 Mono after integration too. Cluster numbers are recomputed every time you re-cluster (and integration recomputes), so cluster-level labels are tied to a specific clustering and have to be re-applied.

Step 6 — Save the annotated object

saveRDS(ifnb, file = "../data/ifnb_annotated.rds")

Continue to Tutorial 04 — Reference Annotation

You’ve named the clusters by hand. The immediate next step, Tutorial 04 — Reference-based Annotation, cross-checks those manual labels against an automated reference (Azimuth) on this same annotated object — confirming the confident calls and flagging the cells the reference can’t place.

After that, Tutorial 05 — Two-Sample Integration re-does the PCA + neighbor-graph + clustering steps on an integrated representation that aligns CTRL and STIM, and you’ll watch what happens to:

the UMAP (CTRL/STIM islands collapse onto each other)
the clusters (cell types stop splitting in two)
the markers (ISGs stop being top markers; cell-type markers reclaim the top)

That integrated .rds is then the input to pseudobulk DE in Tutorial 06.

Credits

Dataset: ifnb via Bioconductor’s muscData::Kang18_8vs8() — Kang et al. (2017) Nat Biotechnol
PBMC canonical markers: Seurat PBMC 3k tutorial, Azimuth PBMC reference, PanglaoDB
Inspiration: scNotebooks