Deep Canonical Correlation Analysis (DCCA)

The classical statistical foundation upon which DCCA is built. CCA finds pairs of linear projections—one for each dataset—such that the projected variables are maximally correlated. In multi-omics, it identifies coordinated patterns between, for example, gene expression and metabolite abundance. The limitation is its inability to capture non-linear relationships, which is the primary gap DCCA addresses by replacing linear projections with deep neural networks.

An extension of classical CCA that incorporates L1 or elastic net regularization to produce sparse linear combinations. This forces many feature weights to zero, selecting only the most relevant variables from each omics modality. For biomarker discovery, Sparse CCA is critical because it yields interpretable results—identifying a small set of genes and proteins driving the cross-modal correlation—whereas standard CCA and DCCA can produce dense, hard-to-interpret loadings.

A neural architecture that learns a compressed, non-linear latent representation by encoding high-dimensional inputs into a bottleneck layer and reconstructing the original modalities. Unlike DCCA, which maximizes correlation between separate views, autoencoders focus on reconstruction fidelity. Variational autoencoders (VAEs) extend this by learning a probabilistic latent space, enabling generative tasks like synthesizing realistic multi-omics profiles for data augmentation.

A deep learning mechanism that dynamically weighs the relevance of features from one omics modality when processing another. For instance, when predicting drug response, a cross-attention layer can learn that specific gene mutations should focus attention on particular protein expression levels. This captures complex, asymmetric inter-modal relationships that symmetric correlation maximization in DCCA might miss, making it a powerful alternative for integrated prediction tasks.

A non-deep-learning method that integrates multiple omics types by constructing patient similarity networks for each data type and iteratively fusing them into a single comprehensive network. SNF excels at identifying patient subtypes by capturing both shared and complementary information across modalities. While DCCA learns a joint latent space, SNF operates on the graph level, making it particularly effective for unsupervised patient stratification when labeled outcome data is scarce.

An unsupervised statistical framework that decomposes the variation from multiple omics data types into a sparse set of latent factors. Each factor captures a principal source of biological or technical variability and can be active in one, several, or all data modalities. MOFA provides interpretability by annotating which features drive each factor, bridging the gap between the black-box nature of DCCA and the need for mechanistic insight in biomarker discovery.