Multi-Omic Factor Analysis (MOFA)

MOFA infers a low-dimensional latent space from multiple heterogeneous data matrices without requiring labeled training data. It simultaneously factorizes each omics layer—such as RNA-seq, DNA methylation, and proteomics—into a shared set of latent factors and modality-specific weight matrices. This reveals the principal axes of variation that are either shared across modalities (capturing coordinated biological programs) or private to a single modality (capturing assay-specific noise).

A core strength of MOFA is its ability to partition the total variance explained by each latent factor across data modalities and sample groups. For each factor, the model quantifies:

• Percentage of variance explained in each omics layer
• Top-weighted features (genes, CpG sites, proteins) driving the factor
• Factor values for each sample, enabling downstream association with clinical covariates This allows researchers to distinguish factors driven by biological signal (e.g., immune activation visible in both RNA and protein) from technical artifacts (e.g., batch effects isolated to one assay).

MOFA employs a probabilistic Bayesian framework based on factor analysis that naturally accommodates incomplete data matrices. Samples do not need to have measurements across all omics layers—the model can learn latent factors from partially overlapping datasets where some samples have only genomics data and others have only transcriptomics. This is critical for integrating real-world clinical cohorts where comprehensive multi-omic profiling is often unavailable for every patient.

To enhance interpretability in high-dimensional biological settings, MOFA applies automatic relevance determination (ARD) priors and spike-and-slab sparsity on the weight matrices. These priors automatically:

• Shrink irrelevant feature weights to zero
• Select only the most informative features per factor
• Determine the optimal number of latent factors from the data This built-in regularization prevents overfitting and yields sparse, interpretable factors where each factor is associated with a manageable set of molecular features.

The MOFA framework has been extended to single-cell resolution through variants like MOFA+, which scales to millions of cells and handles the unique statistical challenges of single-cell data, including zero-inflation and dropout events. It can jointly model scRNA-seq and scATAC-seq from the same cells, revealing coordinated epigenetic and transcriptional programs during cellular differentiation and disease progression.

Once latent factors are inferred, MOFA enables systematic association with sample-level metadata such as clinical outcomes, drug responses, or demographic variables. Factors that significantly correlate with phenotypes of interest can be traced back to their top-weighted molecular features across modalities, providing a hypothesis-generating engine for multi-omic biomarker discovery and mechanistic follow-up studies.

The shared, lower-dimensional mathematical representation where embeddings from distinct biological modalities are aligned. In MOFA, this space captures the principal sources of variation across all input omics layers simultaneously.

• Key property: Points in this space represent unified cellular or sample states
• Dimensionality: Typically 10-50 factors, compared to thousands of original features
• Interpretation: Each dimension corresponds to a latent factor with a specific biological or technical interpretation

This is the core output of MOFA—a compressed coordinate system where samples can be clustered, visualized, and correlated with clinical metadata.

A multi-block dimensionality reduction method that identifies linear combinations of variables across multiple omics datasets that are maximally correlated. Unlike MOFA's probabilistic framework, RGCCA uses a deterministic optimization approach with L1/L2 penalties to handle high-dimensional settings.

• Regularization: L1 penalty induces sparsity, selecting only the most relevant features per component
• Block weighting: Each omics block can be assigned different importance in the global objective
• Schema: Can operate in factorial, centroid, or Horst schemes depending on the desired correlation structure

RGCCA is often compared to MOFA as a frequentist alternative, particularly when linear assumptions are appropriate and computational efficiency is paramount.

A neural network that learns a latent representation of biological data that is invariant to technical confounders while preserving true biological variability. This directly addresses the same challenge MOFA solves through its factor model—separating technical from biological sources of variation.

• Adversarial training: A discriminator network attempts to predict batch labels from the latent space, forcing the encoder to remove batch-specific signals
• Reconstruction objective: The decoder must faithfully reconstruct the original data, ensuring biological information is retained
• Integration with MOFA: MOFA factors can be post-hoc corrected for batch effects, or batch-aware MOFA variants can model technical covariates explicitly

These autoencoders represent the deep learning counterpart to MOFA's matrix factorization approach for disentangling variation.

A generative probabilistic framework that learns a joint posterior distribution from multiple input omics layers. MVAEs extend the variational autoencoder paradigm to handle heterogeneous data types, making them the neural network analog to MOFA's factor analysis.

• Product-of-experts inference: Combines modality-specific encoders into a single joint posterior
• Missing modality imputation: Can generate plausible values for entirely absent omics layers by sampling from the learned joint distribution
• Comparison to MOFA: MVAEs offer non-linear factor representations and explicit generative capabilities, while MOFA provides more interpretable linear factors with sparsity constraints

MVAEs are preferred when non-linear relationships dominate and imputation of missing modalities is a primary objective.

Computational methods designed to align and jointly analyze multiple data types measured from the same individual cells. While MOFA was originally developed for bulk multi-omic cohorts, its principles have been adapted for single-cell resolution.

• Key challenge: Different modalities are often measured in separate cells (unpaired data), requiring cross-modal alignment
• MOFA+: An extension of MOFA specifically designed for single-cell multi-omic data with improved scalability and sparsity handling
• TotalVI: A deep generative model for paired CITE-seq data that jointly models RNA and protein abundance

Single-cell integration methods must handle extreme sparsity, zero-inflation, and the massive scale of modern single-cell datasets.

An integration approach that constrains multi-omic model architecture or training using prior biological databases such as Reactome, Gene Ontology, or STRING. This contrasts with MOFA's purely data-driven factor discovery.

• Pathway-aware embeddings: Aggregate multi-omic signals at the pathway level rather than individual gene level
• Graph-regularized MOFA: Incorporates known gene-gene interaction networks as a penalty term in the factor model
• Mechanistic plausibility: Ensures discovered factors align with established biological knowledge, increasing interpretability for translational research

Knowledge-guided approaches are particularly valuable when sample sizes are limited and prior biological knowledge can compensate for statistical power.