ComBat

ComBat employs a three-level hierarchical model to pool information across genes, providing robust batch parameter estimates even with small batch sizes.

• Level 1: Models the expression data as a function of biological covariates and batch-specific additive/multiplicative effects.
• Level 2: Assumes the batch effect parameters (location and scale shifts) for each gene are drawn from a common parametric prior distribution.
• Level 3: Estimates the hyperparameters of these prior distributions empirically from the data using method of moments estimators.

This hierarchical structure shrinks extreme batch effect estimates toward the overall mean, preventing overfitting when a batch contains few samples.

ComBat corrects for two distinct types of batch-induced distortion simultaneously, making it more comprehensive than simple mean-centering approaches.

• Location Adjustment (Additive): Estimates and removes the systematic shift in mean expression levels for each gene within each batch. This addresses scenarios where an entire batch of samples is uniformly up- or down-shifted.
• Scale Adjustment (Multiplicative): Estimates and corrects for batch-specific differences in variance or dynamic range. This addresses scenarios where a batch exhibits compressed or inflated signal variability.

The model expresses the adjusted value as: Y_ijg = (Y_ijg - γ_ig) / δ*_ig, where γ* and δ* are the empirical Bayes estimates of the additive and multiplicative batch effects for gene g in batch i.

A critical design feature of ComBat is its ability to adjust for batch effects while preserving the biological signal of interest, preventing the removal of true experimental variation.

• The model includes a design matrix that explicitly specifies the biological conditions or phenotypes of interest as fixed effects.
• Batch effect parameters are estimated conditioned on these biological covariates, ensuring that variation attributable to the outcome of interest is not mistakenly identified and removed as a batch effect.
• This is essential for studies where batch and biology are partially confounded, allowing the model to disentangle the two sources of variation rather than simply removing all batch-associated variance.

ComBat offers flexibility in how the prior distribution of batch effects is specified, accommodating different assumptions about the underlying data structure.

• Parametric ComBat: Assumes the gene-wise batch effect parameters (γ_g, δ_g) follow a normal and inverse gamma distribution, respectively. This is the standard, computationally efficient formulation.
• Non-Parametric ComBat: Relaxes the distributional assumption by using a kernel density estimator to flexibly model the shape of the prior distribution directly from the observed data. This is particularly useful when the batch effects do not follow standard parametric forms.

This flexibility allows practitioners to choose a more robust prior when the standard parametric assumptions are violated, improving adjustment quality in complex datasets.

When harmonizing more than two batches, ComBat can be applied using a reference batch approach to avoid a combinatorial explosion of pairwise corrections and ensure a consistent target distribution.

• A single, high-quality batch (e.g., the largest batch or one processed under optimal conditions) is selected as the reference.
• All other batches are then adjusted one-by-one to match the mean and variance characteristics of this reference batch.
• This strategy preserves the scale of the reference batch and provides a stable, interpretable target for harmonization, preventing the drift that can occur when sequentially adjusting batches against each other.
• The choice of reference batch is a critical experimental decision that should be justified based on data quality metrics.

The effectiveness of a ComBat correction is typically validated using Principal Component Analysis (PCA) to visually and quantitatively assess the removal of batch-driven structure.

• Pre-Correction: A PCA plot colored by batch will typically show strong clustering of samples by batch, indicating a dominant technical signal that obscures biology.
• Post-Correction: A successful ComBat adjustment results in PCA plots where samples from different batches are intermixed, while clustering by biological condition becomes the primary source of variance.
• This diagnostic step is essential to confirm that the correction has removed the unwanted technical variation without overcorrecting and eliminating the biological signal of interest.

A systematic non-biological source of variation introduced across different experimental batches—such as different processing dates, reagents, or technicians—that can confound downstream analysis. In high-throughput genomics, batch effects often manifest as global shifts in expression distributions that obscure true biological signals. Understanding the nature and source of batch effects is the critical first step before applying any correction method like ComBat.

A statistical method that estimates and removes the effects of unmodeled, latent sources of variation directly from high-dimensional data without requiring knowledge of the batch variable. Unlike ComBat, which requires explicit batch labels, SVA constructs surrogate variables from the data itself to capture unwanted heterogeneity. This makes it particularly valuable when batch information is unknown, incomplete, or when the sources of technical variation are complex and multidimensional.

An iterative clustering algorithm designed for single-cell RNA sequencing data that projects cells into a shared embedding. Harmony softly clusters cells by both cell type and dataset of origin, then uses these cluster assignments to calculate dataset-specific correction factors. Unlike ComBat's parametric empirical Bayes framework, Harmony uses a fuzzy clustering approach that scales efficiently to hundreds of thousands of cells while preserving rare cell populations.

A critical experimental design flaw where the batch variable is perfectly correlated with the biological condition of interest. For example, if all control samples are processed on Day 1 and all treatment samples on Day 2, it becomes statistically impossible to separate technical artifacts from true biological signal. ComBat and other correction methods cannot rescue confounded designs—this must be addressed at the experimental planning stage through balanced block designs.

A quantitative metric for evaluating batch correction quality. kBET compares the local batch label distribution in a k-nearest neighbor graph to the global batch distribution using a chi-squared test. A perfectly mixed dataset yields an acceptance rate near 1.0, while residual batch effects produce lower scores. This metric is commonly used to benchmark ComBat against newer methods like Harmony and scVI.

A deep generative model based on a variational autoencoder that models single-cell RNA-seq data with a zero-inflated negative binomial distribution. scVI explicitly accounts for batch effects as a latent variable in its probabilistic framework, learning a batch-corrected low-dimensional representation. Unlike ComBat's linear adjustments, scVI captures non-linear batch effects and provides uncertainty estimates, making it a modern deep learning alternative for complex integration tasks.