omicverse/omicverse-skills/bulk-combat-correction

Bulk RNA-seq batch correction with ComBat

Overview

Apply this skill when a user has multiple bulk expression matrices measured across different batches and needs to harmonise them before downstream analysis. It follows `t_bulk_combat.ipynb`, w hich demonstrates the pyComBat workflow on ovarian cancer microarray cohorts.

Instructions

1. Import core libraries

• Load omicverse as ov, anndata, pandas as pd, and matplotlib.pyplot as plt.
• Call ov.ov_plot_set() (aliased ov.plot_set() in some releases) to align figures with omicverse styling.

1. Load each batch separately

• Read the prepared pickled matrices (or user-provided expression tables) with pd.read_pickle(...)/pd.read_csv(...).
• Transpose to gene × sample before wrapping them in anndata.AnnData objects so adata.obs stores sample metadata.
• Assign a batch column for every cohort (adata.obs['batch'] = '1', '2', ...). Encourage descriptive labels when availa

ble.

1. Concatenate on shared genes

• Use anndata.concat([adata1, adata2, adata3], merge='same') to retain the intersection of genes across batches.
• Confirm the combined adata reports balanced sample counts per batch; if not, prompt users to re-check inputs.

1. Run ComBat batch correction

• Execute ov.bulk.batch_correction(adata, batch_key='batch').
• Explain that corrected values are stored in adata.layers['batch_correction'] while the original counts remain in adata.X.

1. Export corrected and raw matrices

• Obtain DataFrames via adata.to_df().T (raw) and adata.to_df(layer='batch_correction').T (corrected).
• Encourage saving both tables (.to_csv(...)) plus the harmonised AnnData (`adata.write_h5ad('adata_batch.h5ad', compressio

n='gzip')`).

1. Benchmark the correction

• For per-sample variance checks, draw before/after boxplots and recolour boxes using ov.pl.red_color, blue_color, `gree

n_color` palettes to match batches.

• Copy raw counts to a named layer with adata.layers['raw'] = adata.X.copy() before PCA.
• Run ov.pp.pca(adata, layer='raw', n_pcs=50) and ov.pp.pca(adata, layer='batch_correction', n_pcs=50).
• Visualise embeddings with ov.pl.embedding(..., basis='raw|original|X_pca', color='batch', frameon='small') and repeat fo

r the corrected layer to verify mixing.

1. Defensive validation

``python # Before ComBat: verify batch column exists and has >1 batch assert 'batch' in adata.obs.columns, "adata.obs must contain a 'batch' column" n_batches = adata.obs['batch'].nunique() assert n_batches > 1, f"Only {n_batches} batch — need >1 for batch correction" # Verify gene overlap after concatenation if adata.n_vars < 100: print(f"WARNING: Only {adata.n_vars} shared genes after concat — check gene ID harmonization") ``

1. Troubleshooting tips

• Mismatched gene identifiers cause dropped features—remind users to harmonise feature names (e.g., gene symbols) before conca

tenation.

• pyComBat expects log-scale intensities or similarly distributed counts; recommend log-transforming strongly skewed matrices.
• If batch_correction layer is missing, ensure the batch_key matches the column name in adata.obs.

Examples

• "Combine three GEO ovarian cohorts, run ComBat, and export both the raw and corrected CSV matrices."
• "Plot PCA embeddings before and after batch correction to confirm that batches 1–3 overlap."
• "Save the harmonised AnnData file so I can reload it later for downstream DEG analysis."

References

• Tutorial notebook: `t_bulk_combat.ipynb`
• Example inputs: see the tutorial page for example batch-correction inputs: t_bulk_combat
• Quick copy/paste commands: `reference.md`