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.
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')
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.
