Sccomp

sccomp

Workflows

Standard Workflow

Estimate and test differences in cell type composition across groups from single-cell count data, with optional outlier removal and visualization.

library(sccomp)
library(dplyr)

# Estimate composition and test differences with outlier removal
sccomp_result <- counts_obj |>
  sccomp_estimate(
    formula_composition = ~ type,
    sample = "sample",
    cell_group = "cell_group",
    abundance = "count",
    cores = 1, 
    verbose = FALSE
  ) |>
  sccomp_remove_outliers(cores = 1, verbose = FALSE, max_sampling_iterations = 2000) |>
  sccomp_test()

# Generate boxplot visualization
sccomp_result |> sccomp_boxplot(factor = "type")

Input: A count data frame or SingleCellExperiment/Seurat object. Output: A tibble containing composition effects, FDR, and outlier annotations.

Differential Variability Analysis

Model and test differences in cell-group variability across groups.

library(sccomp)

# Estimate both composition and variability
res <- seurat_obj |>
  sccomp_estimate(
    formula_composition = ~ type,
    formula_variability = ~ type,
    sample = "sample",
    cell_group = "cell_group",
    cores = 1, 
    verbose = FALSE
  )

# Test and plot credible intervals
plots <- res |> sccomp_test() |> plot()
plots$credible_intervals_1D

Input: A Seurat or SingleCellExperiment object. Output: A model fit with variability effects and 1D/2D credible interval plots.

Model Comparison Loo

Compare different composition models (e.g., with vs. without a covariate) using leave-one-out (loo) cross-validation.

library(sccomp)
library(loo)

# Fit model with factor association
model_with_factor_association <- seurat_obj |>
  sccomp_estimate(
    formula_composition = ~ type,
    sample = "sample",
    cell_group = "cell_group",
    inference_method = "hmc",
    enable_loo = TRUE,
    verbose = FALSE
  )

# Fit baseline model without association
model_without_association <- seurat_obj |>
  sccomp_estimate(
    formula_composition = ~ 1,
    sample = "sample",
    cell_group = "cell_group",
    inference_method = "hmc",
    enable_loo = TRUE,
    verbose = FALSE
  )

# Compare models using leave-one-out cross-validation
loo_compare(
  attr(model_with_factor_association, "fit")$loo(),
  attr(model_without_association, "fit")$loo()
)

Input: A Seurat or SingleCellExperiment object. Output: A comparison table containing elpd_diff and se_diff.

Random Effects Modeling

Model cell-type proportions with both fixed effects and random effects (e.g., random intercepts or random slopes).

library(sccomp)

# Fit a random intercept model
res <- seurat_obj |>
  sccomp_estimate(
    formula_composition = ~ type + (1 | group__),
    sample = "sample",
    cell_group = "cell_group",
    bimodal_mean_variability_association = TRUE,
    cores = 1, 
    verbose = FALSE
  )

Input: A Seurat or SingleCellExperiment object with grouping variables in metadata. Output: A model fit containing fixed and random effect estimates.

When to Use

• To perform robust differential composition analysis on single-cell RNA-seq, CyTOF, or microbiome data using sum-constrained Beta-binomial modeling (sccomp_estimate).
• To identify and probabilistically remove outliers from composition estimation (sccomp_remove_outliers).
• To perform differential variability analysis across groups (sccomp_estimate with formula_variability).
• To model multi-level or hierarchical structures with random effects (e.g., random intercepts or random slopes) in single-cell datasets.

When NOT to Use

• For standard differential gene expression analysis, use packages like DESeq2 or edgeR instead of sccomp which is designed for cell-type composition.
• When counts are available, do not model proportions directly (e.g., using abundance = "proportion") because it fails to model uncertainty for rare cell groups; use raw counts instead.

Data Requirements

• Input can be a Seurat object, SingleCellExperiment object, or a count tibble/data frame containing columns for sample, cell group, and abundance (counts).
• For single-cell RNA sequencing, it is recommended to set bimodal_mean_variability_association = TRUE. For CyTOF and microbiome data, set bimodal_mean_variability_association = FALSE.

Key Parameters

• formula_composition: Formula for modeling cell-type composition (e.g., ~ type).
• formula_variability: Formula for modeling cell-group variability (e.g., ~ type).
• sample: Column name representing the sample identifier.
• cell_group: Column name representing the cell group/type identifier.
• abundance ("count"): Column name representing counts or proportions.
• bimodal_mean_variability_association (FALSE): Logical indicating whether to model bimodal mean-variability association (recommended TRUE for scRNA-seq).
• enable_loo (FALSE): Logical indicating whether to enable leave-one-out cross-validation.
• inference_method ("hcm"): Method used for inference (e.g., "hmc").

Best Practices

• Do not run sccomp_remove_outliers() when performing model comparison with loo_compare(), as leave-one-out cross-validation requires the models to be fit on the exact same data.
• Verify the mean-variability association using plot_2D_intervals() or plots$credible_intervals_2D to ensure the linear association is satisfied before interpreting differential variability.
• Use sccomp_proportional_fold_change() to obtain intuitive proportional fold changes for communication, but rely on sccomp_test() for statistical significance.

Common Pitfalls

• Using proportions instead of counts when counts are available: This discards critical uncertainty information for rare cell types. Fix: Pass raw counts and set abundance = "count".
• Running outlier removal before model comparison: This causes loo_compare() to fail or be invalid because different numbers of data points may be removed. Fix: Omit sccomp_remove_outliers() when comparing models.

Alternatives

• scCODA: Uses a Dirichlet-multinomial model but does not model group-specific variability or outliers.
• propeller: Uses a logit-linear model with limma but does not model data count distribution directly.
• corncob: Uses a Beta-binomial model but does not support mixed-effect modeling or outlier detection.
• ANCOM-BC: Uses a log-linear model designed primarily for microbiome data.

Citations

• Mangiola, Stefano, et al. 2023. "Sccomp: Robust Differential Composition and Variability Analysis for Single-Cell Data." Proceedings of the National Academy of Sciences 120 (33): e2203828120. https://doi.org/10.1073/pnas.2203828120

References

• Homepage: bioconductor.org/packages/sccomp
• Vignette: https://bioconductor.org/packages/release/bioc/vignettes/sccomp/inst/doc/vignettes.html