← Về thư mục
Batch Correction - Usage Guide
Overview
Batch effects are technical variations between experimental batches that can confound biological signals. Different correction methods suit different analysis goals.
Prerequisites
BiocManager::install(c('sva', 'limma', 'DESeq2'))
install.packages('harmony') # for single-cell
Quick Start
Tell your AI agent what you want to do:
- "Remove batch effects from my RNA-seq data using ComBat-Seq"
- "Add batch as a covariate in DESeq2 analysis"
- "Estimate surrogate variables for unknown batch effects"
Example Prompts
Design-Based Correction
"Include batch in my DESeq2 design formula"
"Model batch effects as a covariate in edgeR"
Count Correction
"Apply ComBat-Seq to my raw count matrix"
"Correct counts for batch before clustering"
Unknown Batches
"Use SVA to find hidden batch effects in my data"
"Estimate surrogate variables and include in DE analysis"
Visualization
"Make a PCA plot colored by batch before and after correction"
"Assess batch effect severity in my samples"
What the Agent Will Do
- Visualize batch effects with PCA colored by batch
- Choose appropriate correction method based on downstream analysis
- Apply correction (design formula for DE, ComBat-Seq for visualization)
- Validate correction with post-correction PCA
Method Selection
| Method |
Input |
Use For |
| DESeq2 design formula |
Raw counts |
DE analysis (preferred) |
| ComBat-Seq |
Raw counts |
Visualization, clustering |
| ComBat |
Normalized |
Visualization, ML |
| limma removeBatchEffect |
Normalized |
Visualization only |
| SVA |
Normalized |
Unknown batch sources |
| Harmony |
Embeddings |
Single-cell integration |
Tips
- For differential expression: always include batch in the design formula, never use corrected counts
- For visualization/clustering: Use corrected values (ComBat-Seq for counts, removeBatchEffect for normalized)
- Known vs unknown batches: Use SVA when batch sources are unknown
- Confounding: Batch perfectly correlated with condition is unfixable -- no statistical method can separate them
- Over-correction: Can remove biological signal; compare results with and without correction
- Balanced design: Best results when conditions are evenly spread across batches
- differential-expression/deseq2-basics - DE with batch in design formula
- differential-expression/de-visualization - PCA plots for batch assessment
- single-cell/clustering - Harmony and other integration methods
- expression-matrix/normalization - Data normalization and transformation