Multi-Omics Factor Analysis (MOFA)

MOFA discovers a low-dimensional latent space that explains the joint variation across multiple omics data types (e.g., genomics, transcriptomics, proteomics) measured on the same samples. Unlike supervised methods, it requires no labeled outcomes.

• Learns a shared set of latent factors from heterogeneous data
• Each factor captures a distinct biological or technical source of variation
• Handles missing data naturally—not all modalities need to be measured for every sample
• Outputs a unified sample representation for downstream analysis

Each latent factor is associated with feature weights that quantify the contribution of every molecular feature from every data modality. This enables biological interpretation of the captured variation.

• Gene set enrichment analysis on top-weighted features reveals pathway associations
• Factors can correlate with clinical covariates (e.g., survival, drug response)
• Identifies which data modalities drive each factor
• Enables discovery of cross-omics molecular signatures

MOFA explicitly decomposes the total variance explained by each factor into modality-specific contributions, revealing how much each omics layer drives a given axis of variation.

• Quantifies the percentage of variance explained per modality per factor
• Distinguishes factors driven by a single modality from those capturing cross-modal coordination
• Identifies technical artifacts (e.g., batch effects) that dominate a specific assay
• Guides experimental design by highlighting informative data types

MOFA uses a probabilistic generative model based on Bayesian matrix factorization, which naturally accommodates incomplete experimental designs where not all omics assays are performed on every sample.

• Supports sparse data integration—samples can have measurements for only a subset of modalities
• Models each modality with an appropriate noise distribution (e.g., Bernoulli for binary, Poisson for counts)
• Robust to technical noise and varying feature dimensionality across assays
• Enables integration of public datasets with heterogeneous coverage

The inferred latent factors serve as a low-dimensional embedding for diverse downstream tasks, including clustering, trajectory inference, and association testing.

• Sample clustering in factor space reveals molecular subtypes
• UMAP or t-SNE visualizations colored by factor values highlight structure
• Factors can be used as predictors in survival models or regression analyses
• Integration with MOFA+ (the Python and R implementation) provides seamless workflows with Scanpy and Seurat ecosystems

MOFA differs from other multi-omics integration approaches in its probabilistic, factor-based framework and explicit variance decomposition.

• vs. Canonical Correlation Analysis (CCA): MOFA handles more than two modalities and non-linear relationships through its Bayesian formulation
• vs. Similarity Network Fusion (SNF): MOFA provides interpretable factors with feature weights rather than only a fused sample network
• vs. Variational Autoencoders (VAEs): MOFA offers more direct interpretability of latent dimensions through linear factor loadings
• vs. iCluster: MOFA scales to thousands of features and samples with efficient variational inference

A core mathematical technique for transforming high-dimensional omics data into a lower-dimensional space. MOFA relies on this principle to distill thousands of noisy molecular features into a small set of interpretable latent factors. Key methods include:

• PCA: Finds orthogonal linear combinations maximizing variance.
• t-SNE & UMAP: Non-linear techniques primarily used for visualization, preserving local neighborhood structures.
• Factor Analysis: The direct statistical ancestor of MOFA, modeling observed variables as linear combinations of latent factors plus noise.

The probabilistic framework powering MOFA's model fitting. Unlike methods that produce a single point estimate, Bayesian inference quantifies uncertainty in the learned latent factors and weights. MOFA uses variational inference, an efficient approximation technique, to estimate the posterior distribution of model parameters. This provides a principled way to regularize the model, automatically determining the relevance of each factor and handling missing data without imputation.

A family of statistical models that explain observed correlations among many variables through a smaller number of unobserved, or 'latent', constructs. In MOFA, each latent factor captures a principal axis of variation across all data modalities. A single factor might represent:

• An immune activation axis visible in both transcriptomic and proteomic data.
• A metabolic shift captured by metabolomics and flux data.
• A technical artifact such as batch effects, which MOFA can isolate and interpret.

The linear algebra engine underlying MOFA. The algorithm decomposes each omics data matrix into a product of two smaller matrices: a factor matrix (shared across all data types) and a weight matrix (modality-specific). This decomposition reveals the modular structure of biological variation. MOFA extends standard matrix factorization with:

• Sparsity priors: Encouraging factors to be active in only a subset of features, aiding interpretability.
• Multi-view learning: Simultaneously factorizing multiple matrices with a shared factor matrix.

The overarching computational discipline that MOFA serves. The goal is to create a holistic biological model by combining layers like genomics, transcriptomics, proteomics, and metabolomics from the same samples. Integration strategies fall into three categories:

• Early integration: Concatenating all features before modeling.
• Late integration: Building separate models and fusing predictions.
• Intermediate integration: Finding a joint latent representation, which is MOFA's approach, preserving modality-specific noise models while capturing cross-modal covariance.

A rapidly advancing experimental frontier where multiple omics measurements are taken from the same individual cell. Technologies like CITE-seq (RNA + surface proteins) and 10x Multiome (RNA + ATAC-seq) produce data perfectly suited for MOFA. The model can disentangle the coordinated epigenetic and transcriptional programs defining cell states, identify cell-type-specific latent factors, and impute missing modalities in cells where only one assay was performed.