Differential Analysis

Differential Abundance (DA)

library(CATALYST)
library(diffcyt)

# Load clustered data
sce <- readRDS('sce_clustered.rds')

# Create design matrix
design <- createDesignMatrix(ei(sce), cols_design = 'condition')

# Create contrast
contrast <- createContrast(c(0, 1))  # Treatment vs Control

# Differential abundance test
res_DA <- testDA_edgeR(sce, design, contrast, cluster_id = 'meta20')

# View results
rowData(res_DA)$cluster_id
rowData(res_DA)$p_adj

# Significant clusters
sig_DA <- rowData(res_DA)$p_adj < 0.05
table(sig_DA)

Differential State (DS)

# Test for marker expression differences within clusters
res_DS <- testDS_limma(sce, design, contrast,
                        cluster_id = 'meta20',
                        markers_include = rownames(sce)[rowData(sce)$marker_class == 'state'])

# Results per marker per cluster
ds_results <- rowData(res_DS)

Visualization

# DA results heatmap
plotDiffHeatmap(sce, res_DA, all = TRUE, fdr = 0.05)

# DS results heatmap
plotDiffHeatmap(sce, res_DS, all = TRUE, fdr = 0.05)

# Abundance by condition
plotAbundances(sce, k = 'meta20', by = 'cluster_id', group_by = 'condition')

Manual Statistical Testing

library(tidyverse)

# Get cluster frequencies per sample
freqs <- colData(sce) %>%
    as.data.frame() %>%
    group_by(sample_id, condition, cluster_id = cluster_ids(sce, 'meta20')) %>%
    summarise(n = n(), .groups = 'drop') %>%
    group_by(sample_id) %>%
    mutate(freq = n / sum(n) * 100)

# Test each cluster
test_abundance <- function(df, cluster) {
    cluster_data <- filter(df, cluster_id == cluster)
    ctrl <- filter(cluster_data, condition == 'Control')$freq
    treat <- filter(cluster_data, condition == 'Treatment')$freq

    if (length(ctrl) >= 2 && length(treat) >= 2) {
        test <- t.test(treat, ctrl)
        return(data.frame(
            cluster = cluster,
            fc = mean(treat) / mean(ctrl),
            pvalue = test$p.value
        ))
    }
    return(NULL)
}

results <- map_dfr(unique(freqs$cluster_id), ~test_abundance(freqs, .x))
results$padj <- p.adjust(results$pvalue, method = 'BH')

Mixed Effects Models

library(lme4)
library(lmerTest)

# For paired/repeated measures designs
# Random effect for patient/donor

fit_mixed <- function(df, cluster) {
    cluster_data <- filter(df, cluster_id == cluster)

    model <- lmer(freq ~ condition + (1|patient_id), data = cluster_data)

    coef <- summary(model)$coefficients
    return(data.frame(
        cluster = cluster,
        estimate = coef[2, 'Estimate'],
        pvalue = coef[2, 'Pr(>|t|)']
    ))
}

CITRUS (Automated Discovery)

library(citrus)

# Prepare data
fcs_files <- list.files('data', pattern = '\\.fcs$', full.names = TRUE)
labels <- c(rep('Control', 2), rep('Treatment', 2))

# Run CITRUS
citrus_result <- citrus(
    fcs_files,
    labels,
    fileSampleSize = 1000,
    featureType = 'abundances',
    modelType = 'glmnet',
    family = 'classification'
)

# Get significant clusters
citrus_plot(citrus_result)

Volcano Plot

library(ggplot2)

# From DA results
da_df <- as.data.frame(rowData(res_DA))
da_df$significant <- da_df$p_adj < 0.05

ggplot(da_df, aes(x = logFC, y = -log10(p_adj), color = significant)) +
    geom_point() +
    geom_hline(yintercept = -log10(0.05), linetype = 'dashed') +
    geom_vline(xintercept = c(-1, 1), linetype = 'dashed') +
    scale_color_manual(values = c('gray', 'red')) +
    theme_bw() +
    labs(title = 'Differential Abundance')

Export Results

# Combine DA and DS results
da_results <- as.data.frame(rowData(res_DA))
da_results$analysis <- 'DA'

ds_results <- as.data.frame(rowData(res_DS))
ds_results$analysis <- 'DS'

# Save
write.csv(da_results, 'da_results.csv', row.names = FALSE)
write.csv(ds_results, 'ds_results.csv', row.names = FALSE)

Multiple Comparisons

# For multiple conditions
design_full <- model.matrix(~ 0 + condition, data = ei(sce))
colnames(design_full) <- levels(factor(ei(sce)$condition))

# Multiple contrasts
contrasts <- makeContrasts(
    TreatA_vs_Ctrl = TreatmentA - Control,
    TreatB_vs_Ctrl = TreatmentB - Control,
    TreatA_vs_B = TreatmentA - TreatmentB,
    levels = design_full
)

# Test each contrast
res_list <- lapply(1:ncol(contrasts), function(i) {
    testDA_edgeR(sce, design_full, contrasts[, i], cluster_id = 'meta20')
})

Related Skills

• clustering-phenotyping - Cluster data first
• gating-analysis - Compare gated populations
• differential-expression/de-results - Similar statistical concepts