Epigenomic Batch Correction

The foundational statistical approach that assumes batch effects manifest as systematic shifts in the mean and variance of data distributions. ComBat is the canonical implementation, using an empirical Bayes framework to estimate batch-specific additive and multiplicative parameters. The method borrows strength across genes, making it robust for small batch sizes. It adjusts the data matrix to remove these location-scale differences while preserving the biological covariates of interest.

• Additive effects: Shifts the mean expression or signal level
• Multiplicative effects: Scales the variance of the distribution
• Empirical Bayes: Pools information across features for stable parameter estimation

A matching-based strategy that identifies pairs of cells or samples from different batches that are mutual nearest neighbors in a shared latent space. MNN Correct uses these matched pairs to compute a correction vector, which is then applied to the query batch. This approach assumes that the biological composition between batches is partially overlapping and that batch effects are orthogonal to the biological subspace.

• Shared latent space: Typically derived from PCA or autoencoders
• Correction vector: The difference between paired cell profiles
• Gaussian smoothing: Propagates correction to non-overlapping populations

A linear alignment technique that identifies a common low-dimensional subspace where the correlation between batches is maximized. Seurat's integration method uses canonical correlation analysis (CCA) to find shared sources of variation, then applies dynamic time warping to align the datasets. This method is particularly effective when the biological signal is conserved across batches but obscured by technical noise.

• CCA: Finds linear combinations of features with maximum cross-batch correlation
• Anchor identification: Selects cells that are mutual nearest neighbors in CCA space
• Transformation matrix: Projects query data onto the reference batch's coordinate system

A probabilistic method that iteratively corrects batch effects by softly assigning cells to clusters and then moving cluster centroids to maximize cluster diversity across batches. Harmony uses a maximum diversity clustering objective, penalizing clusters that are dominated by a single batch. The correction is applied in a low-dimensional embedding space, making it computationally efficient for large-scale single-cell epigenomic atlases.

• Fuzzy clustering: Each cell has a probability of belonging to multiple clusters
• Diversity penalty: Encourages each cluster to contain cells from all batches
• Linear mixture model: Assumes batch effects are additive in the embedding space

Neural network-based approaches that learn a latent representation where batch identity and biological state are explicitly factorized. scVI and trVAE use variational autoencoders with conditional architectures to model the data-generating process. The decoder is conditioned on batch labels, forcing the encoder to learn a batch-invariant latent space. This approach can model complex, non-linear batch effects that linear methods cannot capture.

• Conditional VAE: Decoder receives both latent code and batch label
• Adversarial training: A discriminator network penalizes batch-specific information in the latent space
• Negative binomial likelihood: Models the discrete count nature of epigenomic data

A non-parametric global adjustment technique that forces the empirical distribution of values in each batch to be identical. The method computes a reference distribution, typically the average of all batch distributions, and then projects each batch's data onto this reference. While effective for removing gross distributional differences, it assumes that global differences are purely technical and that biological differences manifest only in the relative ordering of features.

• Rank-invariant: Preserves the rank order of features within each sample
• Reference distribution: Usually the mean of all batch empirical distributions
• Assumption: Biological variation is orthogonal to batch-level distributional shifts

The gold-standard statistical framework for removing known batch effects from epigenomic count data while preserving biological variability. It uses an empirical Bayes framework to shrink batch effect estimates toward a pooled mean, making it robust for small sample sizes. Unlike standard ComBat, ComBat-Seq directly models negative binomial distributions inherent to sequencing data, avoiding the need for log-transformation which can distort zero-inflated signals in assays like ATAC-seq or ChIP-seq.

A correction method that identifies pairs of cells or samples from different batches that are mutually nearest neighbors in a shared feature space. It estimates batch vectors based on the differences between these paired observations and applies a smoothing correction to the remaining population. MNN excels at correcting non-linear batch distortions and is a cornerstone of single-cell epigenomic integration workflows, preserving heterogeneous cell-type structures that linear models might erase.

An iterative algorithm that projects cells into a shared embedding using PCA, then applies a soft-clustering approach to group cells by both cell type and batch. It corrects batch effects by shifting cluster centroids toward a global consensus location using fuzzy clustering. Harmony is computationally efficient and scales to massive single-cell epigenomic atlases, making it a preferred tool for integrating chromatin accessibility data across dozens of donors and experimental conditions.

A straightforward batch correction approach where principal components correlated with known batch variables are identified and regressed out from the original data matrix. While simple and interpretable, it assumes batch effects are linear and orthogonal to biological signal. This method is often used as a baseline or quick diagnostic tool to visualize the magnitude of batch effects in histone modification ChIP-seq data before applying more sophisticated non-linear corrections.

A subspace alignment technique that identifies shared correlation structures across batches by finding linear combinations of features that maximize cross-batch correlation. In the Seurat integration pipeline, CCA anchors are used to identify conserved cell populations across datasets, enabling robust batch correction for single-cell ATAC-seq and multi-omic epigenomic data. It preserves fine-grained biological distinctions that global normalization methods often collapse.

A calibration strategy that uses spike-in controls or genomic 'dark matter' regions expected to have no biological signal to estimate and remove technical variation. By modeling the observed variation in these negative control features, a correction factor is derived and applied genome-wide. This approach is particularly valuable for DNA methylation arrays where control probes on the Illumina Infinium platform provide direct measurements of technical noise independent of biological methylation state.