Last updated: 2026-07-02

Checks: 5 1

Knit directory: immgenT-GP-analysis/analysis/

This reproducible R Markdown analysis was created with workflowr (version 1.7.2). The Checks tab describes the reproducibility checks that were applied when the results were created. The Past versions tab lists the development history.

Environment: empty

Great job! The global environment was empty. Objects defined in the global environment can affect the analysis in your R Markdown file in unknown ways. For reproduciblity it’s best to always run the code in an empty environment.

Seed: set.seed(1)

The command set.seed(1) was run prior to running the code in the R Markdown file. Setting a seed ensures that any results that rely on randomness, e.g. subsampling or permutations, are reproducible.

Session information: recorded

Great job! Recording the operating system, R version, and package versions is critical for reproducibility.

Cache: none

Nice! There were no cached chunks for this analysis, so you can be confident that you successfully produced the results during this run.

File paths: relative

Great job! Using relative paths to the files within your workflowr project makes it easier to run your code on other machines.

Repository version: no version control

Tracking code development and connecting the code version to the results is critical for reproducibility. To start using Git, open the Terminal and type git init in your project directory.

This project is not being versioned with Git. To obtain the full reproducibility benefits of using workflowr, please see ?wflow_start.

All panels except (a) are produced by script-refactor/Figure6.R, which shares its CITE-seq setup with Figure S6 via code-refactor/R/citeseq_shared_setup.R. The code below is shown for reference (not re-executed on this page, since this script takes about a minute to run); the images are its pre-rendered output.

Setup

library(ggplot2)
library(ggrepel)
library(dplyr)
library(patchwork)
library(pheatmap)
library(tidyr)
library(Matrix) # protein matrices are dgCMatrix; must be attached for `[` to dispatch

data_path <- "data/"
figure_path <- "figure-refactor/Figure 6/"
source("code-refactor/R/gated_protein_helpers.R")

Panels (b) onward additionally load the shared CITE-seq cell/protein setup:

# ============================================================
# Load data (shared with FigureS6.R)
# ============================================================
source("code-refactor/R/citeseq_shared_setup.R")

# Re-derive the normalized protein matrix used only by this script's panel
# 6b (Protein_F_pm_raw filtered/scaled -- FigureS6.R doesn't need it).
Protein_F_pm <- Protein_F_pm_raw[!rownames(Protein_F_pm_raw) %in% isotype_proteins, ]
Protein_F_pm <- Protein_F_pm[rownames(Protein_F_pm) %in% good_proteins, ]
Protein_F_pm <- Protein_F_pm[!rownames(Protein_F_pm) %in% exclude_proteins, ]
Protein_F_pm <- Protein_F_pm[!rownames(Protein_F_pm) %in% thy11_proteins, ]
D_lognorm <- diag(1 / apply(Protein_F_pm, 2, function(x) max(abs(x))))
Protein_F_pm <- Protein_F_pm %*% D_lognorm
colnames(Protein_F_pm) <- paste0("GP", 1:ncol(Protein_F_pm))
Protein_F_pm[is.na(Protein_F_pm)] <- 0

(a) Projection schematic

This panel is a hand-drawn schematic, not generated from R – there is no code or pre-rendered image to show here. See figures/final-selected/bits/Figure 6/6a.pdf for the published panel.

(b) Protein-program heatmap

# ============================================================
# 6b: sparse protein-program heatmap, contamination GPs removed
# ============================================================
threshold_simplified <- 0
keep_rows_simplified <- apply(Protein_F_pm, 1, function(v) any(abs(v) > threshold_simplified, na.rm = TRUE))
Protein_F_pm_simplified <- Protein_F_pm[keep_rows_simplified, , drop = FALSE]
keep_cols_simplified <- apply(Protein_F_pm_simplified, 2, function(v) any(abs(v) > threshold_simplified, na.rm = TRUE))
Protein_F_pm_simplified <- Protein_F_pm_simplified[, keep_cols_simplified, drop = FALSE]

GP_contamination <- c("GP40", "GP50", "GP55", "GP188")
Protein_F_pm_simplified_no_contamination <- Protein_F_pm_simplified[, !colnames(Protein_F_pm_simplified) %in% GP_contamination, drop = FALSE]

sparse_cutoff <- 0.5
bk_sparse <- unique(c(seq(-1, -sparse_cutoff, length.out = 26), seq(-sparse_cutoff, sparse_cutoff, length.out = 51), seq(sparse_cutoff, 1, length.out = 26)))
cols_sparse <- c(colorRampPalette(c("#4575B4", "white"))(25), rep("white", 50), colorRampPalette(c("white", "#D73027"))(25))

pdf(paste0(figure_path, "6b.pdf"), width = 20, height = 40)
pheatmap::pheatmap(
  t(Protein_F_pm_simplified_no_contamination),
  main = sprintf("Protein programs (sparse, no contamination GPs): %d proteins x %d GPs", nrow(Protein_F_pm_simplified_no_contamination), ncol(Protein_F_pm_simplified_no_contamination)),
  color = cols_sparse,
  breaks = bk_sparse,
  border_color = "black"
)
dev.off()

Fig. 6b. Heatmap of the re-estimated protein matrix (U). The protein scores in each GP are scaled so its maximum |score| = 1, after removing isotype and low-quality proteins and four putative contamination GPs (GP40, GP50, GP55, GP188); 47 proteins (columns) x 179 GPs (rows). Entries with |score| < 0.5 are shown white, with color running from blue (-1) through white to red (+1).

(c-f) Protein-gated vs. GP-loading populations

# ============================================================
# 6c-6f: protein-gate vs. GP-loading comparison for the 4 curated main-figure GPs
# (df_markers2, thymocyte/proliferating/miniverse_cells, L_pm_for_gating,
# select_proteins, threshold_results_subset_manual all come from
# citeseq_shared_setup.R above)
# ============================================================
GPs_fig6 <- c("GP171", "GP23", "GP12", "GP80")
fig6_letter <- c("GP171" = "6c", "GP23" = "6d", "GP12" = "6e", "GP80" = "6f")
enlarge_gps <- c("GP8", "GP30", "GP170", "GP107")
for (gp in GPs_fig6) {
  k_name <- paste0("K", sub("^GP", "", gp))
  plot_gated_gp_vs_protein(
    gp_name = k_name,
    df_markers = df_markers2,
    protein_mat = protein_mat_normalized_lognorm,
    loading_mat = L_pm_for_gating,
    mde_emb = mde_result,
    missing_threshold_action = "skip",
    threshold_df = threshold_results_subset_manual,
    exclude_cells = c(thymocyte_cells, proliferating_cells, miniverse_cells),
    selected_proteins = select_proteins,
    loading_q = NULL,
    min_pointsize = if (gp %in% enlarge_gps) 3L else 0L,
    save_path = paste0(figure_path, fig6_letter[gp], ".pdf")
  )
}

Fig. 6c-f. Transcriptional GPs recover protein-gated populations. For each GP, cells are shown twice on the same MDE embedding. Left, cells passing a protein gate built from the GP’s curated marker signature (positive markers above, and negative markers below); right, an equally sized set of cells with the highest GP loading (the loading cutoff is chosen to match the protein-gate count). Highlighted cells are colored by two-dimensional density and all other cells are grey; thymocytes, proliferating, and “miniverse” cells are excluded.

(g, h) KLRG1 modulation across lineages

# ============================================================
# 6g/6h: KLRG1 modulation (CD8 vs CD4, CD8 vs Treg)
# ============================================================
FlashierDGE_corrected <- function(F1, L1, group1, group2, title_plot = "") {
  loadings_group1 <- colMeans(L1[group1, ])
  loadings_group2 <- colMeans(L1[group2, ])
  mean_change_loadings <- loadings_group1 - loadings_group2
  vplot <- data.frame(SYMBOL = names(mean_change_loadings), mean_change_loadings = mean_change_loadings, AveExpr = colMeans(L1[c(group1, group2), ]))
  list(diff_factors = vplot)
}
get_klrg1_split <- function(cell_type_label, meta, protein_data, threshold) {
  cells <- meta$cellID[meta$annotation_level1 == cell_type_label]
  cells <- intersect(cells, rownames(protein_data))
  list(pos = cells[protein_data[cells, "KLRG1"] >= threshold], neg = cells[protein_data[cells, "KLRG1"] < threshold])
}
run_checked_dge <- function(group_list, F_mat, L_mat, label) {
  if (length(group_list$pos) < 3 || length(group_list$neg) < 3) stop(paste("Insufficient data:", label))
  df <- FlashierDGE_corrected(F1 = F_mat, L1 = L_mat, group1 = group_list$pos, group2 = group_list$neg)$diff_factors
  if (!"SYMBOL" %in% colnames(df)) df$SYMBOL <- rownames(df)
  df
}
plot_target_gps <- function(df, x_var, y_var, label_var, target_gps, highlight_color = "darkorange", background_color = "black",
                             x_limits = c(-0.5, 0.5), y_limits = c(-0.5, 0.5), background_alpha = 0.5,
                             xlab = "Difference in Mean Loading", ylab = "Difference in Mean Loading", title = "Comparison of Specific GP Loadings") {
  highlight_df <- df %>% filter({{ label_var }} %in% target_gps) %>% mutate(.label_display = as.character({{ label_var }}))
  ggplot(df, aes(x = {{ x_var }}, y = {{ y_var }})) +
    geom_point(color = background_color, alpha = background_alpha) +
    geom_abline(slope = 1, intercept = 0, linetype = "dashed", color = "red") +
    geom_hline(yintercept = 0, linetype = "dashed", color = "blue") +
    geom_vline(xintercept = 0, linetype = "dashed", color = "blue") +
    geom_point(data = highlight_df, aes(color = {{ label_var }}), size = 2) +
    ggrepel::geom_text_repel(data = highlight_df, aes(label = .label_display, color = {{ label_var }}), max.overlaps = Inf, size = 3.5, box.padding = 0.35, point.padding = 0.5, segment.color = "grey50", show.legend = FALSE) +
    scale_color_manual(values = highlight_color, guide = "none") +
    coord_cartesian(xlim = x_limits, ylim = y_limits) +
    labs(x = xlab, y = ylab, title = title) +
    theme_minimal()
}

klrg1_threshold <- threshold_results_subset_manual$Threshold[threshold_results_subset_manual$Protein == "KLRG1"]
cd8_split <- get_klrg1_split("CD8", seurat_meta_filtered, protein_mat_normalized_lognorm, klrg1_threshold)
CD4_split <- get_klrg1_split("CD4", seurat_meta_filtered, protein_mat_normalized_lognorm, klrg1_threshold)
treg_split <- get_klrg1_split("Treg", seurat_meta_filtered, protein_mat_normalized_lognorm, klrg1_threshold)

diff_CD8 <- run_checked_dge(cd8_split, F_pm_filtered, L_pm_filtered, "CD8") %>% rename(mean_change_CD8 = mean_change_loadings, AveExpr_CD8 = AveExpr)
diff_CD4 <- run_checked_dge(CD4_split, F_pm_filtered, L_pm_filtered, "CD4") %>% rename(mean_change_CD4 = mean_change_loadings, AveExpr_CD4 = AveExpr)
diff_Treg <- run_checked_dge(treg_split, F_pm_filtered, L_pm_filtered, "Treg") %>% rename(mean_change_Treg = mean_change_loadings, AveExpr_Treg = AveExpr)

# 6g: CD8 vs CD4
merged_cd4 <- inner_join(diff_CD4, diff_CD8, by = "SYMBOL")
p_6g <- plot_target_gps(
  df = merged_cd4, x_var = mean_change_CD8, y_var = mean_change_CD4, label_var = SYMBOL,
  target_gps = c("GP10", "GP58", "GP25", "GP26", "GP43"), background_alpha = 0.8, x_limits = c(-0.2, 0.4), y_limits = c(-0.2, 0.4),
  highlight_color = c("GP10" = "darkorange2", "GP25" = "blue", "GP43" = "blue", "GP26" = "blue", "GP58" = "darkorange2"),
  title = "KLRG1 Modulation: CD8 vs CD4", xlab = "Effect Size in CD8 (KLRG1+ - KLRG1-)", ylab = "Effect Size in CD4 (KLRG1+ - KLRG1-)"
) + theme_bw()
ggsave(paste0(figure_path, "6g.pdf"), p_6g, width = 7, height = 6)

# 6h: CD8 vs Treg
merged_treg <- inner_join(diff_Treg, diff_CD8, by = "SYMBOL")
p_6h <- plot_target_gps(
  df = merged_treg, x_var = mean_change_CD8, y_var = mean_change_Treg, label_var = SYMBOL,
  target_gps = c("GP6", "GP10", "GP12", "GP27", "GP68", "GP58"), background_alpha = 0.8, x_limits = c(-0.2, 0.4), y_limits = c(-0.2, 0.4),
  highlight_color = c("GP10" = "darkorange2", "GP27" = "deeppink", "GP6" = "deeppink", "GP68" = "deeppink", "GP12" = "deeppink", "GP58" = "darkorange2"),
  title = "KLRG1 Modulation: CD8 vs Treg", xlab = "Effect Size in CD8 (KLRG1+ - KLRG1-)", ylab = "Effect Size in Treg (KLRG1+ - KLRG1-)"
) + theme_bw()
ggsave(paste0(figure_path, "6h.pdf"), p_6h, width = 7, height = 6)

Fig. 6g, h. KLRG1 modulation of GPs across lineages. Within each lineage, cells are split KLRG1+ versus KLRG1- based on the threshold on the KLRG1 protein, and for every GP the effect size is the difference in mean GP loading (KLRG1+ minus KLRG1-). The scatter compares CD8 (x-axis) against CD4 (g) or Treg (h) (y-axis); each point is a GP, the dashed red line marks equal effect (y = x), and the dashed blue lines mark zero. Selected GPs are labeled – orange for large, concordant effects in both lineages, and blue/pink for lineage-biased GPs.

(i-k) GPs associated with CD69

# ============================================================
# 6i/6j/6k: the 10 GPs most associated with CD69
# ============================================================
D_scale6 <- diag(1 / apply(F_pm_filtered, 2, function(x) max(abs(x), na.rm = TRUE)))
F_pm_filtered_scaled <- F_pm_filtered %*% D_scale6
colnames(F_pm_filtered_scaled) <- paste0("GP", 1:ncol(F_pm_filtered_scaled))

cd69_top_gps_subset <- c("GP35", "GP6", "GP170", "GP26", "GP58", "GP171", "GP63", "GP62", "GP3", "GP29")
shared_cells_cd69 <- intersect(rownames(L_pm_filtered), rownames(protein_mat_normalized_lognorm))
cd69_expr_vec <- protein_mat_normalized_lognorm[shared_cells_cd69, "CD69"]
cd69_corr <- sapply(cd69_top_gps_subset, function(gp) cor(L_pm_filtered[shared_cells_cd69, gp], cd69_expr_vec, method = "spearman"))
cd69_top_gps_sorted <- names(sort(cd69_corr, decreasing = FALSE)) # most-correlated GP ends up at top of y-axis

plot_factor_heatmap <- function(F_matrix, gp_vector, n_top = 5, min_abs_loading = 0.5, transpose = FALSE,
                                 title = "Factor loadings", low_color = "steelblue", mid_color = "white", high_color = "firebrick", font_size = 9) {
  F_sub <- F_matrix[, gp_vector, drop = FALSE]
  selected_genes <- lapply(gp_vector, function(gp) {
    vals <- F_sub[, gp]
    top_pos <- names(sort(vals, decreasing = TRUE))[seq_len(min(n_top, sum(vals > 0)))]
    top_neg <- names(sort(vals, decreasing = FALSE))[seq_len(min(n_top, sum(vals < 0)))]
    c(top_pos, top_neg)
  })
  selected_genes <- unique(unlist(selected_genes))
  if (min_abs_loading > 0) {
    max_abs <- apply(F_sub[selected_genes, , drop = FALSE], 1, function(x) max(abs(x), na.rm = TRUE))
    selected_genes <- names(max_abs[max_abs >= min_abs_loading])
  }
  hc_genes <- hclust(dist(F_sub[selected_genes, , drop = FALSE]))
  gene_order <- rownames(F_sub[selected_genes, , drop = FALSE])[hc_genes$order]
  plot_df <- F_sub[selected_genes, , drop = FALSE] %>%
    as.data.frame() %>%
    tibble::rownames_to_column("Gene") %>%
    tidyr::pivot_longer(cols = -Gene, names_to = "GP", values_to = "Loading") %>%
    mutate(GP = factor(GP, levels = gp_vector), Gene = factor(Gene, levels = gene_order))
  limit <- max(abs(plot_df$Loading), na.rm = TRUE)
  x_aes <- if (transpose) "Gene" else "GP"
  y_aes <- if (transpose) "GP" else "Gene"
  ggplot(plot_df, aes(x = .data[[x_aes]], y = .data[[y_aes]], fill = Loading)) +
    geom_tile() +
    scale_fill_gradient2(low = low_color, mid = mid_color, high = high_color, midpoint = 0, limits = c(-limit, limit), name = "Loading") +
    labs(title = title, x = NULL, y = NULL) +
    theme_minimal(base_size = font_size) +
    theme(axis.text.x = element_text(angle = if (transpose) 90 else 45, hjust = 1, size = font_size), axis.text.y = element_text(size = font_size), panel.grid = element_blank())
}

p_heatmap <- plot_factor_heatmap(F_matrix = F_pm_filtered_scaled, gp_vector = cd69_top_gps_sorted, n_top = 5, font_size = 9, transpose = TRUE, low_color = "#4DAF4A", mid_color = "white", high_color = "#984EA3")

corr_strip_df <- data.frame(GP = factor(cd69_top_gps_sorted, levels = cd69_top_gps_sorted), Correlation = cd69_corr[cd69_top_gps_sorted], x = "Corr")
corr_limit <- max(abs(corr_strip_df$Correlation))
p_corr_strip <- ggplot(corr_strip_df, aes(x = x, y = GP, fill = Correlation)) +
  geom_tile() +
  scale_fill_gradient2(low = "royalblue", mid = "white", high = "tomato", midpoint = 0, limits = c(-corr_limit, corr_limit), name = "Corr\n(CD69)") +
  labs(x = NULL, y = NULL) +
  theme_minimal(base_size = 9) +
  theme(axis.text.x = element_text(size = 9, angle = 45, hjust = 1), axis.text.y = element_blank(), axis.ticks.y = element_blank(), panel.grid = element_blank())

p_6i <- p_corr_strip + p_heatmap + patchwork::plot_layout(widths = c(0.06, 1), guides = "collect")
ggsave(paste0(figure_path, "6i.pdf"), p_6i, width = 11, height = 5)

# 6j/6k: mean loading of these GPs per tissue (j) and per lineage (k)
cells_for_heatmap <- intersect(rownames(L_pm_filtered), rownames(seurat_meta_filtered))
L_cd69_sub <- L_pm_filtered[cells_for_heatmap, cd69_top_gps_sorted, drop = FALSE]
meta_hm <- seurat_meta_filtered[cells_for_heatmap, c("annotation_level1", "organ_simplified")]

mean_loading_long <- function(L_mat, group_vec, gp_levels) {
  as.data.frame(L_mat) %>%
    mutate(group = group_vec) %>%
    tidyr::pivot_longer(cols = -group, names_to = "GP", values_to = "Loading") %>%
    group_by(group, GP) %>%
    summarise(mean_loading = mean(Loading, na.rm = TRUE), .groups = "drop") %>%
    mutate(GP = factor(GP, levels = gp_levels))
}
make_mean_loading_heatmap <- function(df, title) {
  fill_max <- max(df$mean_loading, na.rm = TRUE)
  ggplot(df, aes(x = group, y = GP, fill = mean_loading)) +
    geom_tile() +
    scale_fill_gradient(low = "white", high = "firebrick", limits = c(0, fill_max), name = "Mean\nloading") +
    labs(title = title, x = NULL, y = NULL) +
    theme_minimal(base_size = 9) +
    theme(axis.text.x = element_text(angle = 45, hjust = 1, size = 9), axis.text.y = element_text(size = 9), panel.grid = element_blank())
}

df_organ <- mean_loading_long(L_cd69_sub, meta_hm$organ_simplified, cd69_top_gps_sorted)
p_6j <- make_mean_loading_heatmap(df_organ, "Mean GP loading by tissue (organ_simplified)")
ggsave(paste0(figure_path, "6j.pdf"), p_6j, width = 9, height = 5)

df_level1 <- mean_loading_long(L_cd69_sub, meta_hm$annotation_level1, cd69_top_gps_sorted)
p_6k <- make_mean_loading_heatmap(df_level1, "Mean GP loading by cell type (level1)")
ggsave(paste0(figure_path, "6k.pdf"), p_6k, width = 7, height = 5)

Fig. 6i-k. GPs associated with CD69. The ten GPs most associated with CD69 are ordered by the Spearman correlation between CD69 protein expression and GP loading. (i) Heatmap of the top five gene scores per GP (green-white-purple), with a left strip giving each GP’s CD69 correlation (blue-white-red). (j) Mean loading of these GPs per tissue and (k) per lineage.


sessionInfo()
R version 4.5.1 (2025-06-13)
Platform: aarch64-apple-darwin20
Running under: macOS Sequoia 15.6.1

Matrix products: default
BLAS:   /System/Library/Frameworks/Accelerate.framework/Versions/A/Frameworks/vecLib.framework/Versions/A/libBLAS.dylib 
LAPACK: /Library/Frameworks/R.framework/Versions/4.5-arm64/Resources/lib/libRlapack.dylib;  LAPACK version 3.12.1

locale:
[1] en_CA/en_CA/en_CA/C/en_CA/en_CA

time zone: Asia/Tokyo
tzcode source: internal

attached base packages:
[1] stats     graphics  grDevices utils     datasets  methods   base     

loaded via a namespace (and not attached):
 [1] vctrs_0.7.3     cli_3.6.6       knitr_1.50      rlang_1.2.0    
 [5] xfun_0.55       stringi_1.8.7   otel_0.2.0      promises_1.5.0 
 [9] jsonlite_2.0.0  workflowr_1.7.2 glue_1.8.1      rprojroot_2.1.1
[13] git2r_0.36.2    htmltools_0.5.9 httpuv_1.6.16   sass_0.4.10    
[17] rmarkdown_2.30  evaluate_1.0.5  jquerylib_0.1.4 tibble_3.3.0   
[21] fastmap_1.2.0   yaml_2.3.12     lifecycle_1.0.5 stringr_1.6.0  
[25] compiler_4.5.1  fs_1.6.6        Rcpp_1.1.1-1.1  pkgconfig_2.0.3
[29] later_1.4.4     digest_0.6.39   R6_2.6.1        pillar_1.11.1  
[33] magrittr_2.0.5  bslib_0.9.0     tools_4.5.1     cachem_1.1.0