Canonical Correlation Analysis (CCA)

CCA finds pairs of canonical variates—linear combinations of features from each dataset—that maximize the Pearson correlation between them. This is achieved by solving a generalized eigenvalue problem on the cross-covariance matrix. The first pair of canonical variates captures the strongest shared signal, with subsequent pairs constrained to be uncorrelated with previous ones, ensuring each dimension captures orthogonal modes of co-variation.

CCA projects both datasets into a common low-dimensional embedding where cells or samples from different batches are aligned by their correlated biological signal rather than technical artifacts. The canonical correlation vectors define this shared subspace, effectively filtering out batch-specific noise while preserving the underlying biological heterogeneity. This makes CCA a foundational step in Seurat's integration workflow for single-cell data.

Standard CCA fails when the number of features exceeds the number of samples, as the cross-covariance matrix becomes singular. Regularized CCA (rCCA) introduces L2 penalties on the canonical vectors to stabilize the estimation. This shrinkage parameter controls the bias-variance tradeoff, preventing overfitting and enabling robust integration of single-cell RNA-seq datasets where thousands of genes are measured across relatively few cells.

In the Seurat v3 integration framework, CCA is used to identify cross-dataset anchors—pairs of cells from different batches that are mutual nearest neighbors in the CCA subspace. These anchors represent cells in a similar biological state across batches. The differences between anchored cells are used to compute a correction vector, which is then applied to harmonize the entire dataset without requiring all cells to have a direct counterpart.

CCA generalizes several common statistical techniques. When one dataset contains a single variable, CCA reduces to multiple regression. When both datasets are identical, it becomes principal component analysis (PCA). Unlike PCA, which maximizes variance within a single dataset, CCA maximizes the correlation between two datasets, making it uniquely suited for multi-modal data integration tasks such as aligning scRNA-seq and scATAC-seq measurements.

Sparse CCA (sCCA) adds L1 penalties to the canonical vectors, forcing many coefficients to zero. This yields linear combinations that depend on only a small subset of features, enhancing biological interpretability. In batch correction contexts, sparse CCA can identify the specific genes driving cross-batch alignment, distinguishing them from genes that exhibit strong batch-specific variation and should be downweighted during integration.

A critical experimental design flaw where the batch variable is perfectly correlated with the biological condition of interest. When this occurs, no computational method—including CCA—can reliably separate technical artifacts from true biological signal. For example, if all control samples are processed on Day 1 and all treatment samples on Day 2, the batch effect and the treatment effect are statistically indistinguishable, rendering any downstream analysis uninterpretable.

A diversity score used to quantitatively evaluate the success of CCA-based integration. For each cell, iLISI computes the effective number of different batches in its local neighborhood. A well-integrated dataset will have an iLISI score approaching the total number of batches, indicating that a cell's nearest neighbors are drawn uniformly from all batches rather than clustering by technical origin. This metric provides an objective measure of batch mixing quality.