Abstract
Multi-omic studies offer a holistic view of biological processes across molecular layers, yet the unsupervised integration of these heterogeneous datasets remains a significant computational challenge. We present Multi-Omics Factor Analysis (MOFA), a statistical framework designed to discover the principal sources of variation in multi-omic datasets. MOFA infers latent factors that capture both shared and modality-specific biological and technical variability. By handling missing data and diverse data distributions, MOFA enables downstream tasks such as sample subgrouping, data imputation, and outlier detection. We demonstrate MOFA's utility on a cohort of 200 chronic lymphocytic leukaemia (CLL) patients and single-cell multi-omics data, identifying key drivers of disease heterogeneity and cellular differentiation.
Introduction
Biological systems are inherently multi-layered, with signals propagating across the genome, transcriptome, and proteome. While modern experiments often collect data across these modalities, integrating them into a cohesive biological narrative is difficult. Traditional methods often struggle with the "missing data" problem—where not all samples are profiled across all assays—and fail to distinguish whether a source of variation is shared across molecular layers or unique to one.
Multi-Omics Factor Analysis (MOFA) addresses these challenges by providing a Bayesian framework for unsupervised integration. It moves beyond simple clustering to identify continuous axes of variation (latent factors) that explain the heterogeneity within a population.
Methodology
MOFA generalizes Principal Component Analysis (PCA) to handle multiple data matrices simultaneously. It decomposes these matrices into a shared set of latent factors and view-specific weight matrices.
Technical Innovations
- Variational Bayesian Inference: To ensure scalability, MOFA utilizes variational inference rather than traditional Markov Chain Monte Carlo (MCMC) sampling. This allows the model to process large clinical and single-cell datasets in minutes or hours rather than days.
- Two-Level Sparsity: The model employs "spike-and-slab" priors to enforce sparsity. This ensures that factors are driven by a small, interpretable subset of features and views, preventing the model from over-fitting to noise.
- Flexible Likelihood Models: MOFA is not restricted to continuous data. It can integrate Gaussian (continuous), Poisson (counts), and Bernoulli (binary) distributions within a single model, making it ideal for combining RNA-seq, mutation status, and methylation data.
- Robustness to Missing Values: MOFA is designed to handle incomplete data matrices. It ignores missing entries during the learning process and can subsequently use the learned factors to impute missing values with high accuracy.
Applications and Insights
The authors validated MOFA across simulated and real-world biological datasets, highlighting its versatility.
Chronic Lymphocytic Leukaemia (CLL)
In a study of 200 CLL patients, MOFA integrated somatic mutations, RNA expression, DNA methylation, and ex-vivo drug responses. Despite 40% of samples having at least one missing assay, MOFA identified 10 major factors of heterogeneity.
- Shared vs. Private Variation: MOFA can pinpoint exactly which data type informs a factor. For example, Factor 1 (IGHV status) was visible across all four data types, while Factor 4 was specific only to mRNA, likely representing cell-type contamination.
- Novel Drivers: MOFA identified a previously underappreciated source of variation related to oxidative stress and heat shock proteins (HSPs), which correlated with specific drug response patterns.
- Clinical Prediction: The factors identified by MOFA were more predictive of "time to next treatment" than standard PCA or individual molecular features alone.
Single-Cell Multi-Omics
Applied to mouse embryonic stem cells (mESCs) with parallel methylation and transcriptome sequencing, MOFA recovered a smooth differentiation trajectory. Unlike discrete clustering methods, MOFA’s continuous factors accurately mapped the transition from "naive" to "differentiated" states, capturing coordinated transcriptional and epigenetic changes along the differentiation path.
Performance and Utility
MOFA demonstrates significant computational efficiency. In the CLL cohort, MOFA reached convergence in 45 minutes, whereas Group Factor Analysis (GFA) required 34 hours and iCluster took nearly a week. Furthermore, MOFA’s imputation capabilities outperformed standard methods like k-Nearest Neighbors (kNN) and SoftImpute, particularly when entire data modalities were missing for a subset of samples. This robustness allows researchers to include all available samples in an analysis rather than discarding those with incomplete records.
Limitations and Future Directions
While powerful, MOFA is currently a linear model, which may limit its ability to capture complex non-linear biological interactions. Future iterations may incorporate non-linear kernels or deep learning components to address this. Additionally, while MOFA is unsupervised, incorporating biological priors (such as known pathway information) could further enhance the interpretability of the discovered factors, guiding the model toward known biological processes while still allowing for the discovery of novel signals.
Summary updated