What is Dr. Zubair Khalid's research focus?

Dr. Zubair Khalid specializes in molecular virology, mRNA vaccine development, and computational biology, with a focus on avian pathogens like IBDV and Avian Reovirus.

Where is Dr. Zubair Khalid currently working?

Dr. Zubair Khalid is a Postdoctoral Research Associate at the University of Maryland (UMD), specifically within the Department of Animal and Avian Sciences.

Combat Batch Effect Removal in Omics Data Analysis

By Dr. Zubair Khalid, DVM, MS, PhD · 2026-08-11

Introduction to Batch Effects and Combat

Batch effects are systematic, non-biological technical variations that arise when samples are processed in groups, or batches, rather than under a single uniform condition. In omics experiments, these effects are pervasive and can arise from differences in reagent lots, instrument calibration drift, operator handling, temperature fluctuations during RNA extraction, or the day on which a sequencing run was performed. For example, two otherwise identical RNA-seq libraries prepared on different days may show hundreds of differentially expressed genes purely because of subtle differences in reverse transcriptase efficiency or adapter ligation yields. Similarly, in mass spectrometry-based proteomics, a change in liquid chromatography column performance between batches can shift peptide retention times and ion intensities, producing apparent abundance differences that have no biological basis.

The magnitude of batch effects should not be underestimated. In multi-center studies or longitudinal cohorts where samples are collected over months or years, batch effects can account for more variance in the data than the biological signal of interest. A classic illustration is the observation that clustering of samples by processing date often produces cleaner separation than clustering by disease status. This is why batch effect correction is not an optional refinement but a mandatory step in any integrated omics analysis.

What Are Batch Effects?

Batch effects are defined as technical sources of variation shared by samples processed together, which are independent of the biological variables under study. They are distinct from confounding variables in that they are not inherently linked to the experimental design—a batch effect exists because of how the experiment was physically executed, not because of what the samples represent. Common sources include:

Reagent lot changes: Different lots of TRIzol, antibodies, or sequencing kits can introduce systematic shifts.
Instrument drift: Mass spectrometers, microarrays scanners, and sequencers all degrade or shift in performance over time.
Environmental factors: Temperature, humidity, and oxygen levels in a lab can vary seasonally.
Operator effects: Different technicians may perform extractions or library preparations with slightly different efficiencies.
Time-dependent degradation: Samples stored for different durations before processing may show systematic changes in RNA integrity or protein stability.

The key feature of a batch effect is that it is additive and multiplicative in a predictable way across many features (genes, proteins, metabolites) simultaneously. This property is what makes statistical correction possible: if the effect were random per feature, no correction method could distinguish it from biological noise.

Why ComBat Is Widely Used

ComBat (Combating Batch Effects) was introduced by Johnson, Li, and Rabinovic in 2007 as an empirical Bayes method for adjusting batch effects in microarray data. Its popularity has persisted and expanded to RNA-seq, proteomics, and metabolomics because it addresses a fundamental limitation of simpler correction methods: when batch sizes are small, estimating batch-specific means and variances directly from the data is noisy and can introduce more variance than it removes.

ComBat's innovation is to borrow information across features (genes, proteins) to stabilize the estimates of batch parameters. Instead of estimating each gene's batch effect independently, ComBat assumes that the batch effects across all genes follow a common prior distribution. This empirical Bayes framework allows the method to shrink extreme estimates toward the overall mean, producing more robust corrections, particularly when the number of samples per batch is modest (e.g., 3–10 samples per batch).

ComBat is also widely used because it is implemented in accessible software packages (the sva package in R and pycombat in Python), requires only a data matrix and a batch label vector, and preserves the original data scale, making downstream analyses straightforward.

The Statistical Model Behind ComBat

ComBat operates on a location-scale model. For each feature \( g \) (gene, protein, metabolite) and each sample \( j \) in batch \( i \), the observed expression value \( Y_{ijg} \) is modeled as:

\[ Y_{ijg} = \alpha_g + X_{ij}\beta_g + \gamma_{ig} + \delta_{ig}\varepsilon_{ijg} \]

where:

\( \alpha_g \) is the overall mean expression for feature \( g \),
\( X_{ij} \) is the design matrix of biological covariates (e.g., disease status, age, sex),
\( \beta_g \) is the regression coefficient for the biological covariates,
\( \gamma_{ig} \) is the additive batch effect (location shift) for feature \( g \) in batch \( i \),
\( \delta_{ig} \) is the multiplicative batch effect (scale factor) for feature \( g \) in batch \( i \),
\( \varepsilon_{ijg} \) is the residual error, assumed to follow a normal distribution with mean zero and feature-specific variance \( \sigma_g^2 \).

The goal is to estimate \( \gamma_{ig} \) and \( \delta_{ig} \) and then adjust the data:

\[ Y^{*}_{ijg} = \frac{Y_{ijg} - \hat{\alpha}_g - X_{ij}\hat{\beta}_g - \hat{\gamma}_{ig}}{\hat{\delta}_{ig}} + \hat{\alpha}_g + X_{ij}\hat{\beta}_g \]

This adjustment removes the batch-specific shift and scaling while preserving the biological signal captured by \( X_{ij}\beta_g \).

Location and Scale Adjustments

The location parameter \( \gamma_{ig} \) accounts for the fact that a gene might be systematically higher or lower in one batch relative to another. For example, if batch 1 was processed with a less efficient reverse transcription step, all genes might show slightly lower raw counts. The scale parameter \( \delta_{ig} \) accounts for differences in variability—batch 2 might have higher technical noise, compressing or expanding the dynamic range of measured intensities.

The distinction between location and scale is critical. A method that only corrects means (such as simple centering per batch) will fail when batches have different variances. Conversely, a method that only standardizes variances but ignores mean shifts will leave systematic offsets in place. ComBat handles both simultaneously, which is why it outperforms simpler normalization approaches.

Empirical Bayes Shrinkage

The core statistical innovation in ComBat is the empirical Bayes step. Rather than estimating \( \gamma_{ig} \) and \( \delta_{ig} \) independently for each gene (which would be equivalent to simple per-gene, per-batch standardization), ComBat assumes that the batch effect parameters themselves are drawn from probability distributions:

\[ \gamma_{ig} \sim N(\gamma_i, \tau_i^2) \] \[ \delta_{ig}^2 \sim \text{Inverse Gamma}(\lambda_i, \theta_i) \]

The hyperparameters (\( \gamma_i, \tau_i^2, \lambda_i, \theta_i \)) are estimated from the data across all genes in each batch. Then, the posterior means of \( \gamma_{ig} \) and \( \delta_{ig} \) are computed given the observed data for each gene. This produces shrinkage: genes with extreme batch effect estimates (often due to noise) are pulled back toward the batch-level mean, while genes with consistent batch effects across many samples retain their estimated values.

This shrinkage is particularly valuable when batch sizes are small. With only 3–5 samples per batch, the direct estimate of a gene's batch mean has high variance. The empirical Bayes approach reduces this variance substantially, at the cost of introducing slight bias. In practice, the bias-variance tradeoff strongly favors shrinkage, as demonstrated by improved downstream classification and differential expression results.

Assumptions and Prerequisites for ComBat

ComBat is not a universal solution. It operates under specific assumptions that must be met for the correction to be valid.

Known Batch Labels

ComBat requires that batch labels are known and provided as input. This seems obvious, but in practice, batch information is often incompletely recorded. If samples were processed across multiple plates, runs, or days, the investigator must have accurate records of which sample belongs to which batch. In some cases, batch structure can be inferred from the data (e.g., using principal component analysis to identify clusters of samples processed together), but this is a separate step and introduces additional uncertainty.

The Batch Record is an essential laboratory document that tracks reagent lots, instrument IDs, dates, and operator names. Without a meticulous batch record, applying ComBat is impossible because the batch variable is unknown.

Sample Size Considerations

ComBat requires a sufficient number of samples per batch to estimate the hyperparameters reliably. As a rule of thumb, each batch should contain at least 3–5 samples, and ideally more. With fewer than 3 samples per batch, the empirical Bayes shrinkage becomes less effective because the batch-level hyperparameter estimates themselves are noisy.

Additionally, the total number of batches matters. With only 2 batches, ComBat can still be applied, but the correction is essentially equivalent to a paired analysis. With 5 or more batches, the empirical Bayes framework has enough data to estimate the prior distributions well.

A related consideration is the balance of biological groups across batches. If all cases are in batch 1 and all controls are in batch 2, ComBat cannot distinguish batch effects from biological effects—the design is confounded. ComBat will remove the difference between batches, which in this case is indistinguishable from the case-control difference, and the biological signal will be lost. This is not a failure of ComBat but a fundamental limitation of the experimental design. The solution is to randomize samples across batches and include biological replicates of each condition in every batch.

ComBat Variants and Extensions

The original ComBat was designed for microarray data with approximately normally distributed intensities. Omics data types have different distributions, and several variants have been developed to address these differences.

ComBat-seq for Count Data

RNA-seq data are counts, not continuous intensities. The negative binomial distribution is a more appropriate model for count data than the normal distribution assumed by the original ComBat. ComBat-seq, developed by Zhou et al. in 2022, adapts the empirical Bayes framework to negative binomial models. It estimates batch effects on the count scale and produces adjusted counts that preserve the integer nature of the data.

ComBat-seq is particularly important because applying the original ComBat to RNA-seq counts (even after log transformation) can produce negative adjusted values or distort the mean-variance relationship. ComBat-seq avoids these issues by modeling the dispersion directly. However, ComBat-seq is computationally more intensive and requires the raw count matrix as input, not normalized or log-transformed data.

For proteomics data, which often involves peptide intensity values that are approximately log-normal after preprocessing, the original ComBat or a variant that accounts for missing values (such as ComBat with a missingness model) is typically more appropriate. See Proteomics Batch Effect Correction for a detailed treatment of batch effects specific to mass spectrometry workflows.

ComBat with Biological Covariates

The original ComBat formulation includes a design matrix \( X \) for biological covariates. This is a crucial feature: ComBat can preserve the effects of known biological variables while removing batch effects. For example, if you are studying gene expression differences between tumor and normal tissue, you can include the tissue type as a covariate. ComBat will then estimate batch effects after accounting for tissue type, ensuring that the correction does not remove the tumor-normal differences.

This is implemented by fitting the model \( Y_{ijg} = \alpha_g + X_{ij}\beta_g + \gamma_{ig} + \delta_{ig}\varepsilon_{ijg} \) and only adjusting the residual after removing the covariate effects. In practice, this means running ComBat with a mod argument that specifies the biological covariates. Failing to include known biological covariates can lead to overcorrection, where genuine biological variation is removed along with batch effects.

Evaluating the Effectiveness of Batch Effect Removal

After applying ComBat, it is essential to verify that batch effects have been reduced and that biological signal is preserved. Several diagnostic approaches are standard.

Visual Diagnostics

Principal component analysis (PCA) is the most common visual tool. Before correction, a PCA plot colored by batch should show samples clustering by batch. After correction, the batch clusters should disperse, and samples should intermingle across batches. Simultaneously, if the data contain a strong biological grouping (e.g., cases vs. controls), the PCA plot should still show separation along the principal components that capture biological variance.

A Heat Map of Genes can also be informative. Before correction, the heat map may show block-like patterns corresponding to batches. After correction, these blocks should disappear, and the heat map should reflect biological structure.

Quantitative Metrics

Visual inspection is subjective. Quantitative metrics provide more rigorous assessment:

Batch entropy: This metric measures how uniformly samples from different batches are mixed in a low-dimensional embedding. Higher entropy indicates better batch mixing.
Silhouette width: Compute the silhouette width of samples labeled by batch. After correction, the mean silhouette width should decrease, indicating that batches are no longer well-separated.
Biological signal preservation: After correction, test for differential expression between biological groups. The number of significant genes should be similar to (or greater than) the number found before correction, provided batch effects were obscuring true signal.
Variance decomposition: Use a mixed-effects model to estimate the proportion of variance explained by batch before and after correction. This should decrease substantially.

A useful approach is to compute the ratio of between-batch variance to total variance for each gene before and after correction. The median of this ratio should decrease after ComBat.

Common Pitfalls and Misconceptions

ComBat is powerful but frequently misapplied. The following failure modes are common in practice.

Overcorrection Risks

The most serious pitfall is overcorrection, where ComBat removes biological variation along with batch effects. This occurs when biological covariates are not included in the model. For example, if you are comparing two cell types and all cell type A samples happen to be in batch 1 while cell type B samples are in batch 2, ComBat without covariates will interpret the cell type difference as a batch effect and remove it. The corrected data will show no difference between cell types—a catastrophic loss of signal.

Even with covariates included, overcorrection can occur if the biological covariate is correlated with batch. In such cases, ComBat will attribute some of the biological variation to batch and remove it. This is an inherent limitation of any batch correction method when the experimental design is confounded. The only robust solution is to design the experiment with balanced batches from the start.

Another overcorrection scenario involves the inclusion of too many covariates. If you include a covariate that is itself affected by batch (e.g., a quality metric that differs between batches), ComBat may remove variation that is actually biological.

Incorrect Application to Test Data

A common error in machine learning pipelines is to apply ComBat to the entire dataset, including the test set, before splitting into training and test sets. This introduces information leakage: the batch effect parameters estimated from the full dataset incorporate information from the test samples, which can inflate performance metrics.

The correct procedure is to estimate ComBat parameters (the batch effect estimates and hyperparameters) on the training set only, then apply the same transformation to the test set. In R, this is done by fitting ComBat on the training data and then using the predict function (or manually applying the stored batch parameters) to adjust the test data. This ensures that the test set remains independent.

Small Batch Sizes

Applying ComBat to batches with only 1–2 samples is problematic. The empirical Bayes shrinkage becomes unreliable, and the correction may introduce more noise than it removes. In such cases, alternative approaches such as merging small batches into larger ones (if biologically justifiable) or using simpler methods like mean-centering may be preferable.

Applying ComBat to Non-Normalized Data

ComBat assumes that the data are approximately normally distributed within each feature after accounting for batch and covariates. Raw count data from RNA-seq or raw peak intensities from mass spectrometry are not normally distributed. Applying the original ComBat to such data can produce poor results. For RNA-seq, use ComBat-seq. For proteomics, apply ComBat after appropriate transformation (e.g., log2 transformation of intensity values).

Practical Workflow for Applying ComBat

The following workflow outlines the steps for applying ComBat to a typical omics dataset.

Step-by-Step Implementation

Preprocess the data: For RNA-seq, generate a count matrix and filter low-expression genes. For microarray or proteomics, apply appropriate normalization (e.g., quantile normalization for microarrays, log2 transformation for proteomics intensities). The data matrix should have features in rows and samples in columns.

Define batch and covariates: Create a data frame with sample IDs, batch labels, and biological covariates (e.g., disease status, age, sex). Verify that batch labels are complete and accurate.

Check the design: Confirm that each biological group is represented in each batch. If not, consider whether ComBat is appropriate or whether the design is too confounded.

Run ComBat: In R, use the sva package:

library(sva)
# data_matrix: genes x samples
# batch: factor vector of batch labels
# mod: model matrix of biological covariates
mod <- model.matrix(~ disease_status + age, data = sample_info)
corrected_data <- ComBat(dat = data_matrix, batch = batch, mod = mod)

In Python, use pycombat:

from combat.pycombat import pycombat
corrected_data = pycombat(data_matrix, batch)

Validate the correction: Run PCA before and after correction. Compute batch entropy or silhouette width. Check that biological differential expression is preserved.

Document the parameters: Record the batch effect estimates, the number of samples per batch, and the covariates used. This documentation is essential for reproducibility.

Validation and Reporting

When reporting results, state explicitly that ComBat was applied, which version was used, what covariates were included, and how the correction was validated. Include before-and-after PCA plots in supplementary materials. If ComBat-seq was used for RNA-seq data, note that the adjusted counts are no longer integers and should not be used with tools that expect raw counts (e.g., DESeq2's internal normalization).

Alternatives to ComBat and Comparative Insights

ComBat is not the only batch correction method, and in some situations, alternatives may be more appropriate.

Limma and RUV

Limma's removeBatchEffect function is a simpler approach that fits a linear model with batch as a factor and subtracts the batch coefficients. It does not perform empirical Bayes shrinkage and does not adjust for scale differences. For large batches (e.g., 20+ samples per batch), limma's approach can perform nearly as well as ComBat and is computationally simpler.

RUV (Removing Unwanted Variation) takes a different approach: it uses control genes (known to be unaffected by biological condition) or replicate samples to estimate the unwanted variation directly from the data. RUV can be more powerful than ComBat when good control genes are available, but it requires additional information that is not always accessible.

Harmony and Other Tools

Harmony, originally developed for single-cell RNA-seq data integration, uses an iterative clustering approach to identify and correct batch effects in a low-dimensional embedding. It is particularly effective for large datasets with many batches and complex population structures. However, Harmony operates on the embedding space rather than the original feature space, which means the corrected values are not directly interpretable as gene expression levels.

Other tools include:

scGen: A generative model for single-cell data that learns batch-invariant representations.
BERNN: A neural network approach for batch effect removal.
MMUPHin: Designed for meta-analysis of microbiome data.

The choice of method depends on the data type, the number of batches, the batch sizes, and whether the downstream analysis requires corrected feature-level values or just corrected embeddings. For most bulk omics applications with known batch labels and moderate batch sizes, ComBat remains the default choice due to its balance of performance, interpretability, and ease of use.

Summary and Best Practices

Key Takeaways

Batch effects are systematic technical variations that can obscure biological signal in omics data.
ComBat uses an empirical Bayes framework to estimate and remove batch-specific location and scale effects while preserving biological covariates.
ComBat requires known batch labels, adequate sample sizes per batch, and a balanced experimental design.
ComBat-seq is the appropriate variant for RNA-seq count data; the original ComBat is suitable for continuous data like microarrays and log-transformed proteomics intensities.
Always include biological covariates in the ComBat model to prevent overcorrection.
Validate batch correction with PCA, batch entropy, and differential expression analyses.
Never apply ComBat to the full dataset before splitting into training and test sets in machine learning pipelines.

Best Practices Checklist

[ ] Record batch information meticulously in a Batch Record during sample processing.
[ ] Randomize samples across batches and include all biological groups in each batch.
[ ] Use at least 3–5 samples per batch; more is better.
[ ] Include known biological covariates in the ComBat model.
[ ] Use ComBat-seq for RNA-seq counts; use original ComBat for continuous data.
[ ] Validate the correction with PCA and quantitative metrics.
[ ] Document all parameters and versions for reproducibility.
[ ] For multi-omics integration, apply ComBat separately to each data type, then integrate using a Multi-omics Approach.
[ ] When performing Biomarker Discovery, ensure that batch correction does not remove the biomarker signal by validating on independent cohorts.

Frequently Asked Questions

What is combat batch effect removal?

ComBat batch effect removal is an empirical Bayes method that adjusts omics data for systematic technical variations introduced by processing samples in different batches. It estimates batch-specific location (mean) and scale (variance) parameters for each feature and shrinks these estimates toward a common prior distribution, producing corrected data that retain biological signal.

When should I use ComBat?

Use ComBat when you have omics data (microarray, RNA-seq, proteomics, metabolomics) with known batch labels and you need to remove technical variation before downstream analysis. It is appropriate when batch sizes are moderate (3–10 samples per batch) and when the experimental design is not fully confounded with batch.

Does ComBat work for RNA-seq count data?

The original ComBat is not ideal for raw RNA-seq counts because it assumes normality. Use ComBat-seq, which models count data with a negative binomial distribution and produces adjusted counts that preserve the data's statistical properties.

Can ComBat remove biological variation?

Yes, if biological covariates are not included in the model, ComBat can remove genuine biological variation, especially when biological groups are unevenly distributed across batches. Always include known biological covariates in the model to minimize this risk.

How many samples per batch are needed for ComBat?

A minimum of 3 samples per batch is recommended, with 5 or more being preferable. With fewer samples, the empirical Bayes shrinkage becomes less reliable, and the correction may introduce noise.

Should I include test samples when applying ComBat?

No. In machine learning workflows, ComBat parameters must be estimated on the training set only and then applied to the test set. Including test samples in the correction causes information leakage and inflates performance estimates.

How do I know if batch effects were successfully removed?

Use PCA plots to check that samples no longer cluster by batch. Compute quantitative metrics such as batch entropy or silhouette width. Verify that biological differential expression is preserved or improved after correction.