Experimental Design and Batch Effects

1. Experimental units

The experimental unit is the entity independently assigned to a condition, such as a donor, animal, or culture. Cells from one donor are not automatically independent biological replicates.

donor,condition,batch,library_id
D01,control,1,L01
D02,control,2,L02
D03,treated,1,L03
D04,treated,2,L04

2. Confounding

If every control is processed in batch 1 and every treated sample in batch 2, condition and batch are indistinguishable. No algorithm can prove which caused a difference. Balance batches when possible.

Good: each batch contains control and treated samples.
Risky: batch 1 contains only controls and batch 2 only treated samples.

3. Metadata validation

Treat metadata as data. Check unique IDs, missing values, valid factor levels, and agreement with filenames before running a pipeline.

import pandas as pd
meta = pd.read_csv("metadata.csv")
assert meta.sample_id.is_unique
assert meta.condition.notna().all()
print(pd.crosstab(meta.batch, meta.condition))

meta <- read.csv("metadata.csv", stringsAsFactors = FALSE)
stopifnot(!anyDuplicated(meta$sample_id))
stopifnot(all(!is.na(meta$condition)))
with(meta, table(batch, condition))

4. Analysis consequences

Include pre-specified covariates in the design, document exclusions, and distinguish biological correction from technical removal. Over-correction can erase real biology.

Design formula example: ~ batch + condition
Only use it when the design contains enough information to estimate both terms.

5. Experimental Design Principles

Good experimental design prevents confounding before data generation. The key principles are:

# GOOD: Balanced design – each batch processes both conditions
Batch 1: Control_1, Control_2, Treated_1, Treated_2
Batch 2: Control_3, Control_4, Treated_3, Treated_4

# BAD: Confounded design – batch = condition
Batch 1: Control_1, Control_2, Control_3, Control_4
Batch 2: Treated_1, Treated_2, Treated_3, Treated_4

Why the bad design fails: Any difference between groups could be caused by the biological condition OR by batch-specific technical variation (different reagent lots, different operators, different sequencing lanes). These effects are mathematically inseparable.

6. Identifying Batch Effects

Before correction, you must first detect whether batch effects exist. Use dimensionality reduction (PCA) to visualize sample clustering:

library(DESeq2)

# Perform variance-stabilizing transformation
vsd <- vst(dds, blind = TRUE)

# PCA plot colored by condition, shaped by batch
plotPCA(vsd, intgroup = c("condition", "batch")) +
  labs(title = "PCA: Check for Batch Effects") +
  theme_minimal()

Interpretation guide: - If samples cluster primarily by condition → minimal batch effect, proceed. - If samples cluster primarily by batch → strong batch effect, correction needed. - If batch and condition are confounded → correction is impossible; redesign the experiment.

7. Batch Correction Methods

# Correct approach: include batch in the DESeq2 model
design(dds) <- ~ batch + condition

# The condition effect is estimated AFTER accounting for batch
dds <- DESeq(dds)
results <- results(dds, contrast = c("condition", "treated", "control"))

Critical warning: Never use batch-corrected expression values as input for differential expression testing. Instead, include batch as a covariate in the statistical model. Using corrected values can deflate variance estimates and inflate false positive rates.

8. Sample Size and Power Considerations

Underpowered experiments waste resources and produce unreliable results. Use power analysis before designing your experiment:

# RNA-seq power analysis (rough estimate)
library(RNASeqPower)
rnapower(
  depth = 20,        # millions of reads per sample
  cv = 0.4,          # biological coefficient of variation
  effect = 1.5,      # minimum fold change to detect
  alpha = 0.05,      # significance level
  power = 0.8        # desired power
)
# Output: n = 5 samples per group needed

Practical Exercise

Create a balanced two-condition metadata table with at least four biological replicates across two batches. Use a crosstab to prove each batch contains both conditions.

Pass criteria: IDs are unique, no required metadata is missing, both conditions occur in every batch, and the analysis design is written before inspecting results.

Troubleshooting

If condition and batch are perfectly confounded, report the limitation and avoid claiming a batch-corrected causal result.

Next Steps

Continue with Reference Genomes and Annotation Databases and Reproducible Project Structure. Record the software versions, dataset or example inputs, and any decisions you made.