Regularized Generalized Canonical Correlation Analysis (RGCCA)

A statistical framework that decomposes the variance of multiple omics datasets into a low-rank matrix of latent factors, revealing the principal sources of biological and technical variation. Unlike RGCCA's focus on maximizing cross-block correlation, MOFA explicitly partitions variance into shared factors (capturing variation across all data blocks) and dataset-specific factors. This makes MOFA particularly suited for exploratory analysis where distinguishing between biological signal and technical noise is critical. The model uses Bayesian inference with sparsity-inducing priors to automatically determine the number of relevant latent dimensions.

The computational process of mapping feature vectors from different biological assays into a common coordinate system so that semantically similar biological states occupy proximal positions. While RGCCA finds linear combinations that maximize correlation, modern alignment techniques use contrastive learning or manifold alignment to handle non-linear relationships. Key approaches include:

• Canonical correlation analysis (CCA) for linear alignment
• Dynamic time warping for sequential data
• Optimal transport for distribution-level matching
• Contrastive losses that pull paired samples together and push unpaired samples apart

A shared, lower-dimensional mathematical representation where embeddings from distinct biological modalities (e.g., RNA-seq and ATAC-seq) are aligned to enable cross-modal comparison and integration. RGCCA constructs this space by finding canonical variates—linear combinations of original variables—that maximize pairwise correlations. The dimensionality of this space is determined by the number of retained canonical components. In practice, the joint latent space serves as input for downstream tasks including clustering, trajectory inference, and phenotype prediction, effectively creating a unified molecular portrait from fragmented omic measurements.

An integration approach that constrains multi-omic model architecture or training using prior biological databases (e.g., Reactome, Gene Ontology) to ensure mechanistic plausibility. RGCCA can incorporate this through structural regularization—penalizing the model to favor solutions where known interacting genes receive similar weights. This transforms the purely data-driven correlation maximization into a mechanistically-informed optimization. Key knowledge sources include:

• Protein-protein interaction networks
• Gene regulatory networks
• Pathway databases (KEGG, Reactome)
• Gene co-expression modules

An architecture that explicitly models multi-modal interactions by computing the outer product between modality-specific embeddings, capturing high-order multiplicative correlations across omics layers. While RGCCA finds linear correlations between blocks, tensor fusion captures tri-linear and higher-order interactions that linear methods miss. The resulting tensor is flattened and passed through a neural network for prediction. This approach is computationally expensive—the tensor dimension grows exponentially with the number of modalities—but captures synergistic effects where the combination of two omics signals is more predictive than either alone.

A regularization technique where entire data modalities (e.g., DNA methylation) are randomly zeroed out during training to force the model to learn robust representations that handle missing clinical assays. In the RGCCA framework, this translates to training on subsets of blocks and penalizing solutions that rely too heavily on any single modality. Benefits include:

• Missing data resilience at inference time
• Reduced overfitting to dominant modalities
• Fair contribution across all omics layers
• Clinical applicability where not all assays are always available