Canonical Correlation Analysis

CCA identifies pairs of canonical variates—linear combinations of variables from two datasets—that maximize the Pearson correlation between them. Unlike simple pairwise correlation, CCA finds the directions in feature space where the two views co-vary most strongly. The first pair captures the dominant mode of shared variation, with subsequent pairs constrained to be uncorrelated with previous ones, ensuring each component reveals a distinct axis of coordinated biology.

When integrating ATAC-seq (chromatin accessibility) with RNA-seq (gene expression), CCA compresses thousands of peaks and transcripts into a low-dimensional latent space. This shared embedding aligns cells or samples based on their combined molecular profiles, enabling downstream tasks like clustering or trajectory inference. The reduced dimensions filter out technical noise while preserving the biological signal common to both assays.

Standard CCA fails when the number of features exceeds the number of samples—a common scenario in genomics. Regularized CCA introduces L2 (ridge) or L1 (lasso) penalties on the canonical weights to stabilize estimation and prevent overfitting. This shrinkage ensures the model generalizes beyond the training data and produces interpretable weight vectors that highlight the most influential genomic features driving the correlation.

Sparse CCA applies L1 regularization to force many canonical weights to exactly zero, performing simultaneous feature selection and integration. In practice, this identifies a small set of enhancers or promoters whose accessibility directly correlates with the expression of specific target genes. The resulting sparse loadings are biologically interpretable, pinpointing the regulatory elements that mediate the observed association.

Biological relationships are often non-linear. Kernel CCA maps both datasets into high-dimensional reproducing kernel Hilbert spaces before performing CCA, capturing complex dependencies without explicitly computing the transformation. This allows the method to detect non-linear regulatory logic—such as cooperative transcription factor binding—that linear CCA would miss, at the cost of reduced interpretability of the canonical directions.

The popular single-cell analysis toolkit Seurat uses a variant of CCA to align scRNA-seq and scATAC-seq data from the same biological system. After identifying anchors—mutual nearest neighbors in the CCA space—the method transfers labels and imputes missing modalities. This CCA-based integration enables joint analysis of transcriptomic and epigenomic landscapes across thousands of single cells.

A training paradigm where a single neural network is simultaneously trained on multiple related prediction tasks, such as forecasting expression across different tissues. By sharing hidden representations, the model learns a joint latent space that naturally aligns modalities, effectively performing a non-linear analog of CCA. Hard parameter sharing reduces overfitting on small genomic datasets, while soft parameter sharing uses distance constraints between task-specific weights.

Genomic loci, known as eQTLs, where genetic variants are statistically associated with variation in mRNA expression levels. While CCA finds global correlation structures between omics layers, eQTL analysis pinpoints specific cis-regulatory variants driving expression changes. Deep learning models like Enformer now predict eQTL effects in silico by comparing wild-type and mutated sequence predictions.

Systematic non-biological variations introduced by differences in sample processing, reagent lots, or sequencing platforms. These technical artifacts can dominate the correlation structure found by CCA, producing spurious associations between modalities. Correction methods like ComBat-Seq use negative binomial regression to adjust count data while preserving biological variability, ensuring CCA captures genuine cross-modal relationships.

A model interpretability technique that attributes predictions to input features by accumulating gradients along a path from a baseline to the actual input. When applied to multi-modal genomic models, it reveals which epigenomic features most influence expression predictions. This provides a mechanistic complement to CCA: while CCA identifies correlated latent variables, Integrated Gradients exposes the specific sequence motifs driving the correlation.