Data-Driven Selection of Single-Cell Normalization Methods

A quantitative framework for selecting optimal normalization strategies (e.g., SCTransform vs. LogNormalization) based on diagnostic metrics rather than heuristics. Evaluates three critical dimensions of preprocessing quality:

1. Variance stabilization: Decoupling of mean-variance relationship in normalized expression (lower correlation = better).

2. Biological signal retention: Preservation of known marker genes within highly variable genes (higher retention = better).

3. Dropout robustness: Removal of technical dropout bias from normalized values (lower correlation with dropout rate = better).

Methods are ranked using a weighted composite score. Designed for head-to-head comparison of preprocessed Seurat objects.

Usage

ChooseNormalization(
  ...,
  subset_size = integer(),
  known_hvgs = list(),
  n_hvgs = 2000L,
  low_expressed_thresh = 0.2,
  weight = c(variance_stability = 0.4, marker_signal = 0.35, dropout_robustness = 0.25)
)

Arguments

...

Named arguments where each value is a Seurat object representing a distinct preprocessing strategy (e.g., SCT = sct_obj, Log = log_obj). Requirements:

• Must be named (names become method identifiers)

• Must contain normalized data in the data slot of the specified assay/layer

• Must have identical cell counts (for fair comparison)

subset_size

Integer. Number of cells to subsample for diagnostics. If missing or empty, defaults to min(10000, total_cells). Smaller subsets accelerate computation with minimal accuracy loss for large datasets (>50k cells).

known_hvgs

Named list of canonical marker genes for cell types. Format: list(cell_type1 = c("geneA", "geneB"), cell_type2 = c("geneC", ...)). Used to compute marker_retention metric. If NULL (default), this metric is omitted from scoring.

n_hvgs

Integer. Number of top highly variable genes to evaluate for marker retention. Default: 2000L. Only relevant when known_hvgs is provided.

low_expressed_thresh

Numeric (0–1). Quantile threshold for filtering lowly expressed genes during variance-mean correlation calculation. Genes below this quantile of mean expression are excluded. Default: 0.2 (bottom 20% excluded).

weight

Named numeric vector specifying weights for composite scoring. Must contain exactly these components summing to 1:

variance_stability

Weight for variance-mean decoupling (default: 0.4)

marker_signal

Weight for marker gene retention (default: 0.35)

dropout_robustness

Weight for dropout bias removal (default: 0.25)

Value

A list containing:

metrics

A data.table with diagnostic metrics per method:

• variance_mean_cor: Pearson correlation between log10(mean) and log10(variance) of normalized expression (lower = better)

• marker_retention: Proportion of known markers in top HVGs (higher = better; NA if known_hvgs not provided)

• mean_dropout_residual: Absolute Spearman correlation between dropout rate (from counts) and normalized means (lower = better)

• composite_score: Weighted combination of normalized metrics (0–1 scale)

• rank: Method ranking (1 = best)

recommendation

Character string naming the top-ranked method

plots

(Optional) Diagnostic visualizations if ggplot2 available

Workflow

1. Validates input objects (naming, cell count consistency, data slot presence)

2. Subsamples cells (if needed) for computational efficiency

3. Computes three core metrics per method:

   Variance stabilization

   Correlation between log-transformed mean and variance of normalized expression

   Marker retention

   Overlap between user-provided markers and top HVGs

   Dropout robustness

   Correlation between gene dropout rate (from counts) and normalized expression means

4. Normalizes metrics to 0-1 scale (inverting where lower=better)

5. Computes weighted composite score and ranks methods

6. Returns top recommendation with full diagnostic report

Notes

• Requires pre-normalized Seurat objects—this function does not perform normalization itself. Users must first run SCTransform(), NormalizeData(), etc., and store results in the data slot.

• Marker retention metric is only computed when known_hvgs is provided. Without it, scoring relies solely on variance stabilization and dropout robustness.

• Weights should sum to 1 (not enforced but recommended for interpretable scores).

Examples

if (FALSE) { # \dontrun{
# Compare two preprocessed Seurat objects
sct_obj <- SCTransform(seurat_raw, verbose = FALSE)
log_obj <- NormalizeData(seurat_raw, normalization.method = "LogNormalize")

result <- ChooseNormalization(
  SCT = sct_obj,
  Log = log_obj,
  subset_size = 5000,
  verbose = TRUE
)

# Top recommendation
result$recommendation

# Full metrics table
result$metrics[, .(method, composite_score, rank)]

# With marker genes for refined scoring
markers <- list(
  T_cell = c("CD3D", "CD3E", "CD8A"),
  B_cell = c("CD79A", "MS4A1", "CD19"),
  Myeloid = c("CD14", "LYZ", "FCGR3A")
)

result <- ChooseNormalization(
  SCT = sct_obj,
  Log = log_obj,
  known_hvgs = markers,
  weight = c(variance_stability = 0.3, marker_signal = 0.5, dropout_robustness = 0.2)
)
} # }