SKILL.md

name	bio-crispr-screens-batch-correction
description	Batch effect correction for CRISPR screens. Covers normalization across batches, technical replicate handling, and batch-aware analysis. Use when combining screens from multiple batches or correcting systematic technical variation.
tool_type	python
primary_tool	scipy

Version Compatibility

Reference examples tested with: DESeq2 1.42+, MAGeCK 0.5+, matplotlib 3.8+, numpy 1.26+, pandas 2.2+, scikit-learn 1.4+, scipy 1.12+

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

• Python: pip show <package> then help(module.function) to check signatures

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

Batch Correction

"Correct batch effects in my CRISPR screens" → Normalize and harmonize sgRNA count data across screen batches to remove systematic technical variation while preserving biological signal.

• Python: scipy/sklearn for median normalization and batch correction
• CLI: mageck test with batch-aware design

Median Normalization

Goal: Remove systematic library-size differences between batches.

Approach: Scale each sample within a batch so that sample medians match a global median, correcting for sequencing depth variation.

import numpy as np
import pandas as pd
from scipy import stats

def median_normalize(counts_df, batch_column='batch'):
    '''Normalize counts to median within each batch.'''
    normalized = counts_df.copy()

    guide_columns = [c for c in counts_df.columns if c not in [batch_column, 'gene', 'guide']]

    for batch in counts_df[batch_column].unique():
        batch_mask = counts_df[batch_column] == batch
        batch_data = counts_df.loc[batch_mask, guide_columns]

        sample_medians = batch_data.median(axis=0)
        global_median = sample_medians.median()

        scale_factors = global_median / sample_medians
        normalized.loc[batch_mask, guide_columns] = batch_data * scale_factors

    return normalized

counts_df = pd.read_csv('screen_counts.csv')
normalized = median_normalize(counts_df, 'batch')

Size Factor Normalization

def size_factor_normalize(counts_df, reference='geometric_mean'):
    '''DESeq2-style size factor normalization.'''
    guide_cols = [c for c in counts_df.columns if c.startswith('sample_')]
    counts = counts_df[guide_cols].values

    counts_nonzero = np.where(counts == 0, np.nan, counts)

    if reference == 'geometric_mean':
        log_counts = np.log(counts_nonzero)
        geometric_mean = np.exp(np.nanmean(log_counts, axis=1))
    else:
        geometric_mean = counts_nonzero.mean(axis=1)

    ratios = counts_nonzero / geometric_mean[:, np.newaxis]
    size_factors = np.nanmedian(ratios, axis=0)

    normalized_counts = counts / size_factors
    normalized_df = counts_df.copy()
    normalized_df[guide_cols] = normalized_counts

    return normalized_df, size_factors

normalized, size_factors = size_factor_normalize(counts_df)
print('Size factors:', size_factors)

Quantile Normalization

def quantile_normalize(counts_df, guide_cols=None):
    '''Quantile normalization across samples.'''
    if guide_cols is None:
        guide_cols = [c for c in counts_df.columns if c.startswith('sample_')]

    data = counts_df[guide_cols].values.copy()

    sorted_data = np.sort(data, axis=0)
    mean_values = sorted_data.mean(axis=1)

    ranks = np.argsort(np.argsort(data, axis=0), axis=0)
    normalized = mean_values[ranks]

    result = counts_df.copy()
    result[guide_cols] = normalized

    return result

qn_counts = quantile_normalize(counts_df)

Control-Based Normalization

def normalize_to_controls(counts_df, control_genes, method='median'):
    '''Normalize using non-targeting or negative control guides.'''
    guide_cols = [c for c in counts_df.columns if c.startswith('sample_')]

    is_control = counts_df['gene'].isin(control_genes)
    control_data = counts_df.loc[is_control, guide_cols]

    if method == 'median':
        control_values = control_data.median(axis=0)
    elif method == 'mean':
        control_values = control_data.mean(axis=0)
    elif method == 'sum':
        control_values = control_data.sum(axis=0)

    reference = control_values.median()
    scale_factors = reference / control_values

    normalized = counts_df.copy()
    normalized[guide_cols] = counts_df[guide_cols] * scale_factors

    return normalized, scale_factors

nontargeting = counts_df[counts_df['gene'].str.startswith('NonTargeting')]['gene'].unique()
normalized, factors = normalize_to_controls(counts_df, nontargeting)

Batch Effect Removal with ComBat

Goal: Remove batch effects using empirical Bayes adjustment while preserving biological signal.

Approach: Log-transform counts, apply pyCombat with a batch vector, and back-transform to count space.

def combat_correction(counts_df, batch_vector, guide_cols=None):
    '''ComBat batch correction for count data.'''
    from combat.pycombat import pycombat

    if guide_cols is None:
        guide_cols = [c for c in counts_df.columns if c.startswith('sample_')]

    data = counts_df[guide_cols].values.T

    log_data = np.log2(data + 1)
    corrected = pycombat(log_data, batch_vector)
    corrected_counts = np.power(2, corrected) - 1
    corrected_counts = np.maximum(corrected_counts, 0)

    result = counts_df.copy()
    result[guide_cols] = corrected_counts.T

    return result

batches = [1, 1, 1, 2, 2, 2]
corrected = combat_correction(counts_df, batches)

Batch-Aware Log-Fold Change

def batch_aware_lfc(counts_df, treatment_cols, control_cols, batch_vector):
    '''Calculate LFC accounting for batch structure.'''
    batches = np.unique(batch_vector)

    lfc_by_batch = []
    for batch in batches:
        batch_treat = [c for c, b in zip(treatment_cols, batch_vector) if b == batch and c in treatment_cols]
        batch_ctrl = [c for c, b in zip(control_cols, batch_vector) if b == batch and c in control_cols]

        if len(batch_treat) == 0 or len(batch_ctrl) == 0:
            continue

        treat_mean = counts_df[batch_treat].mean(axis=1)
        ctrl_mean = counts_df[batch_ctrl].mean(axis=1)

        batch_lfc = np.log2((treat_mean + 1) / (ctrl_mean + 1))
        lfc_by_batch.append(batch_lfc)

    combined_lfc = pd.concat(lfc_by_batch, axis=1).mean(axis=1)
    lfc_var = pd.concat(lfc_by_batch, axis=1).var(axis=1)

    return combined_lfc, lfc_var

Replicate Correlation Check

def check_replicate_correlation(counts_df, sample_cols, replicate_groups):
    '''Check correlation between replicates.'''
    correlations = []

    for group, replicates in replicate_groups.items():
        if len(replicates) < 2:
            continue

        for i in range(len(replicates)):
            for j in range(i+1, len(replicates)):
                r1, r2 = replicates[i], replicates[j]
                if r1 in sample_cols and r2 in sample_cols:
                    log_r1 = np.log2(counts_df[r1] + 1)
                    log_r2 = np.log2(counts_df[r2] + 1)

                    corr, pval = stats.pearsonr(log_r1, log_r2)
                    correlations.append({
                        'group': group,
                        'rep1': r1,
                        'rep2': r2,
                        'pearson_r': corr,
                        'pvalue': pval
                    })

    return pd.DataFrame(correlations)

replicate_groups = {
    'treatment_batch1': ['sample_1', 'sample_2'],
    'treatment_batch2': ['sample_4', 'sample_5'],
    'control_batch1': ['sample_3'],
    'control_batch2': ['sample_6']
}

corr_df = check_replicate_correlation(counts_df, counts_df.columns[3:], replicate_groups)
print(corr_df)

Batch QC Metrics

Goal: Quantify batch effect magnitude to determine whether correction is needed.

Approach: Run PCA on log-transformed counts, compute between-batch vs within-batch variance ratio, and assess whether batch structure dominates the first principal components.

def batch_qc_metrics(counts_df, batch_vector, sample_cols):
    '''Calculate batch-related QC metrics.'''
    from sklearn.decomposition import PCA
    from scipy.spatial.distance import pdist

    log_counts = np.log2(counts_df[sample_cols].values.T + 1)

    pca = PCA(n_components=min(5, len(sample_cols)))
    pcs = pca.fit_transform(log_counts)

    batch_labels = np.array(batch_vector)
    unique_batches = np.unique(batch_labels)

    if len(unique_batches) > 1:
        batch_means = [pcs[batch_labels == b].mean(axis=0) for b in unique_batches]
        batch_separation = np.mean(pdist(batch_means))

        within_batch_var = np.mean([pcs[batch_labels == b].var() for b in unique_batches])
        between_batch_var = np.var(batch_means, axis=0).sum()

        batch_effect_ratio = between_batch_var / (within_batch_var + 1e-10)
    else:
        batch_separation = 0
        batch_effect_ratio = 0

    return {
        'batch_separation': batch_separation,
        'batch_effect_ratio': batch_effect_ratio,
        'pca_variance_explained': pca.explained_variance_ratio_,
        'n_batches': len(unique_batches)
    }

qc = batch_qc_metrics(counts_df, [1,1,1,2,2,2], sample_cols)
print(f"Batch effect ratio: {qc['batch_effect_ratio']:.2f}")

Visualization

import matplotlib.pyplot as plt

def plot_batch_effect(counts_df, batch_vector, sample_cols, output_file):
    '''Visualize batch effects with PCA.'''
    from sklearn.decomposition import PCA

    log_counts = np.log2(counts_df[sample_cols].values.T + 1)

    pca = PCA(n_components=2)
    pcs = pca.fit_transform(log_counts)

    fig, ax = plt.subplots(figsize=(8, 6))

    for batch in np.unique(batch_vector):
        mask = np.array(batch_vector) == batch
        ax.scatter(pcs[mask, 0], pcs[mask, 1], label=f'Batch {batch}', s=100)

    ax.set_xlabel(f'PC1 ({pca.explained_variance_ratio_[0]:.1%})')
    ax.set_ylabel(f'PC2 ({pca.explained_variance_ratio_[1]:.1%})')
    ax.legend()
    ax.set_title('PCA - Batch Effects')

    plt.tight_layout()
    plt.savefig(output_file, dpi=150)
    plt.close()

plot_batch_effect(counts_df, [1,1,1,2,2,2], sample_cols, 'batch_pca.png')

Related Skills

• mageck-analysis - Batch-aware MAGeCK analysis
• screen-qc - Quality control before correction
• hit-calling - Analysis after batch correction
• library-design - Control guide design