Batch design

Version Compatibility

Reference examples tested with: limma 3.58+

Before using code patterns, verify installed versions match. If versions differ:

• R: packageVersion('<pkg>') then ?function_name to verify parameters

If code throws ImportError, AttributeError, or TypeError, introspect the installed package and adapt the example to match the actual API rather than retrying.

Batch Design and Mitigation

"Design experiment to avoid batch effects" → Plan sample-to-batch assignments that confound biology with technical variation, and apply correction methods post-hoc.

• R: sva::ComBat(), limma::removeBatchEffect()
• Python: scanpy.pp.combat() for single-cell data

Core Principle

Batch effects are unavoidable. Good design makes them correctable.

Design Rules

1. Never confound batch with condition - Each batch must contain all conditions
2. Balance samples across batches - Equal numbers per condition per batch
3. Randomize within constraints - Avoid systematic patterns
4. Include controls - Same samples across batches if possible

Balanced Design Example

# BAD: Confounded design
# Batch 1: All treated samples
# Batch 2: All control samples
# -> Cannot separate batch from treatment

# GOOD: Balanced design
# Batch 1: 3 treated, 3 control
# Batch 2: 3 treated, 3 control
# -> Batch effect can be estimated and removed

Sample Assignment

library(designit)

# Create balanced assignment
samples <- data.frame(
  sample_id = paste0('S', 1:24),
  condition = rep(c('ctrl', 'treat'), each = 12),
  sex = rep(c('M', 'F'), 12)
)

# Optimize batch assignment
batch_design <- osat(samples, batch_size = 8,
                     balance_cols = c('condition', 'sex'))

Detecting Batch Effects

Goal: Identify hidden batch effects in expression data by estimating surrogate variables that capture unmodeled technical variation.

Approach: Fit a model matrix for the biological variable, estimate the number of surrogate variables using num.sv, then compute surrogate variables with sva for inclusion in downstream differential analysis.

library(sva)

# From count matrix
mod <- model.matrix(~condition, colData)
mod0 <- model.matrix(~1, colData)

# Estimate number of surrogate variables (hidden batches)
n_sv <- num.sv(counts_normalized, mod)

# Estimate surrogate variables
svobj <- sva(counts_normalized, mod, mod0, n.sv = n_sv)

Correction Methods

Method	When to Use
ComBat	Known batches, moderate effects
SVA	Unknown batches, exploratory
RUVseq	Using control genes
limma::removeBatchEffect	Visualization only

Documenting Design

Always record:

• Date of sample processing
• Reagent lot numbers
• Operator
• Equipment/lane assignments
• Any deviations from protocol

Related Skills

• experimental-design/power-analysis - Account for batch in power calculations
• differential-expression/batch-correction - Correcting batch effects in analysis
• single-cell/batch-integration - scRNA-seq batch correction