Multi-Omics Factor Analysis (MOFA)

MOFA is designed to jointly analyze multiple omics data types (e.g., genomics, transcriptomics, proteomics) measured on the same samples without requiring a predefined outcome variable. It decomposes the variation into a set of latent factors that capture the global sources of variability across all data modalities. This allows researchers to identify coordinated molecular changes that would be missed by analyzing each omics layer in isolation.

The model outputs a set of latent factors ordered by the amount of variance they explain. Each factor is associated with a weight matrix that quantifies the contribution of every feature from every omics layer. This structure enables direct biological interpretation:

• Factor 1 might capture technical batch effects.
• Factor 2 might represent a specific immune signaling axis.
• Factor 3 might delineate a metabolic reprogramming event. By inspecting the top-weighted features, researchers can assign mechanistic labels to each axis of variation.

A critical advantage of MOFA is its ability to handle incomplete experimental designs where not every omics assay is performed on every sample. The model uses a probabilistic matrix factorization framework that naturally accommodates missing views. This is essential for integrating public cohort data where different subsets of patients have been profiled with different technologies, maximizing statistical power without discarding partially characterized samples.

MOFA provides a rigorous variance decomposition report that quantifies how much of the variation in each omics layer is explained by each latent factor. This allows researchers to:

• Identify which omics layers are driving specific factors.
• Distinguish factors that capture shared variation across multiple assays from those capturing data-specific noise.
• Prioritize factors that explain a high percentage of variance in clinically relevant data types, such as drug response or survival-associated features.

The latent factor matrix serves as a low-dimensional embedding of the patient cohort. This embedding can be used directly for unsupervised clustering to discover novel disease subtypes. Because the factors are denoised and integrated across omics, the resulting patient groups are more robust and biologically coherent than clusters derived from a single data type. Common downstream analyses include:

• K-means or hierarchical clustering on the factor values.
• Survival analysis to test if the identified subgroups have distinct clinical outcomes.
• Differential expression between the discovered clusters to characterize their molecular profiles.

MOFA employs sparsity-inducing priors (such as spike-and-slab or automatic relevance determination) on the feature weights. This automatically performs feature selection by driving the weights of irrelevant features to zero. The result is that each factor is associated with a concise, interpretable set of top-weighted genes, proteins, or metabolites. This built-in regularization prevents overfitting in high-dimensional settings where the number of features vastly exceeds the number of samples.

MOFA integrates genomics, transcriptomics, and proteomics to identify latent factors that define clinically relevant cancer subtypes. By decomposing variation across data modalities, it reveals driver mutations and dysregulated pathways that would be missed by single-omics analysis.

• Identifies multi-omics signatures for glioblastoma and breast cancer subtypes
• Separates technical noise from biological signal across batches
• Reveals patient-specific pathway dysregulation for personalized treatment selection

MOFA models trained on pharmacogenomics datasets can uncover latent factors that explain differential drug sensitivity. By correlating factors with IC50 values across cell lines, researchers pinpoint the multi-omics basis of drug resistance.

• Integrates somatic mutations, gene expression, and DNA methylation
• Identifies predictive biomarkers for targeted therapy response
• Enables mechanism-of-action deconvolution for novel compounds

For diseases with limited cohort sizes, MOFA leverages multi-modal data to extract maximum information from each patient. It identifies molecular endotypes within phenotypically similar rare disease groups.

• Combines whole-exome sequencing with RNA-seq and metabolomics
• Detects compensatory pathway activation unique to patient subgroups
• Prioritizes candidate genes by their contribution to latent factors

By applying MOFA to time-series multi-omics data, researchers can track how latent factors evolve during disease progression or treatment. This reveals dynamic molecular shifts that precede clinical deterioration.

• Models temporal variation in chronic diseases like diabetes and COPD
• Identifies early warning signatures of treatment failure
• Maps disease stage transitions at the molecular level

MOFA+ extends the framework to single-cell resolution, integrating scRNA-seq with scATAC-seq to resolve cellular heterogeneity. It identifies factors that define discrete cell states and developmental trajectories.

• Aligns chromatin accessibility with transcriptional output per cell
• Reveals regulatory logic governing cell fate decisions
• Discovers rare cell populations with distinct multi-omics profiles

MOFA enables integrative analysis across independent cohorts by learning a shared latent space. This overcomes batch effects and small sample sizes to identify robust, reproducible biomarkers.

• Harmonizes disparate datasets from public repositories like TCGA and GEO
• Increases statistical power for detecting weak but consistent signals
• Validates biomarker generalizability across populations and platforms