henhaobio
Sign in

Molecular Systems BiologyBioinformaticsPeer-reviewed journal article

Multi-Omics factor analysis - a framework for unsupervised integration of multi-omic data sets

Ricard ArgelaguetiD, Britta VelteniD, Damien ArnoliD, Sascha DietrichiD, Thorsten ZenziD, John C. MarioniiD, Wolfgang HuberiD, Florian BuettneriD, Oliver StegleiD

PDFDOIPreprintCodeData
In-depth full-text digest available · Jump to digest ↓

Abstract

Multi-omic studies promise the improved characterization of biological processes across molecular layers. However, methods for the unsupervised integration of the resulting heterogeneous datasets are lacking. We present Multi-Omics Factor Analysis (MOFA), a computational method for discovering the principal sources of variation in multi-omic datasets. MOFA infers a set of (hidden) factors that capture biological and technical sources of variability. It disentangles axes of heterogeneity that are shared across multiple modalities and those specific to individual data modalities. The learnt factors enable a variety of downstream analyses, including identification of sample subgroups, data imputation, and the detection of outlier samples. We applied MOFA to a cohort of 200 patient samples of chronic lymphocytic leukaemia, profiled for somatic mutations, RNA expression, DNA methylation and ex-vivo drug responses. MOFA identified major dimensions of disease heterogeneity, including immunoglobulin heavy chain variable region status, trisomy of chromosome 12 and previously underappreciated drivers, such as response to oxidative stress. In a second application, we used MOFA to analyse single-cell multiomics data, identifying coordinated transcriptional and epigenetic changes along cell differentiation.

In-depth summary

Claude Opus 5.5 · xhigh effort · 8 Nov 2026 · from the full text of version 2 (11 Apr 2018) · CC BY 4.0

MOFA, a sparse Bayesian factor model with mixed likelihoods, separates shared from modality-specific multi-omic variation, recovering IGHV, trisomy 12 and oxidative-stress axes in 200 leukaemia samples.

Earlier integration strategies (marginal association tests such as QTL mapping, kernel or graph similarity networks, multi-view clustering, prior factor models) are local, hard to interpret, non-sparse, tuning-heavy, or weak on missing and non-Gaussian data. MOFA learns sparse, interpretable factors that show which omic layers each axis of variation affects, with variational inference, missing-value handling and per-view likelihoods for continuous, binary and count data.

Key results

  1. ShownIn 200 chronic lymphocytic leukaemia (CLL) samples with mutation, mRNA, methylation and ex-vivo drug response data, MOFA recovered immunoglobulin heavy-chain variable region (IGHV) status (Factor 1) and trisomy 12 (Factor 2) as the two dominant multi-layer axes.

    Nearly 40% of the 200 samples lacked at least one assay; the 10 factors retained (2% variance threshold) explained 41% of drug response, 38% of mRNA, 24% of methylation and 24% of mutation variance; Factors 1 and 2 were active in most assays and matched IGHV and trisomy 12 through top mutation weights; Factor 1 was consistent with three subgroups, though hinting at a continuum. · Application to Chronic Lymphocytic Leukaemia; MOFA identifies important clinical markers in…; Data processing

  2. SuggestedBeyond the established markers, MOFA exposed Factor 5, shared by mRNA and drug response, which points to oxidative stress and heat shock protein expression as an underexplored source of CLL heterogeneity.

    IGHV and trisomy 12 factors together accounted for less than 20% of the variance MOFA explained; Factor 5 mRNA weights were enriched for Reactome oxidative stress and senescence gene sets (FDR<1%), with heat shock proteins as top genes; its top drugs related to oxidative stress responses (reactive oxygen species, DNA damage response, apoptosis). · MOFA identifies important clinical markers in…; Gene set enrichment analysis

  3. ShownOn parallel single-cell methylation and transcriptome profiles of mouse embryonic stem cells (mESCs), MOFA reconstructed a continuous naive-to-primed-to-differentiated trajectory coupling expression with genome-wide methylation shifts, which SNF and iCluster clusters missed.

    87 mESCs (16 in 2i, 71 in serum); 3 factors kept at a 2% threshold; Factor 1 explained 7% of RNA and between 53% and 72% of promoter, CpG island and enhancer methylation variance, loading on naive markers such as Tbx3 and Fbxo15; Factor 2 loaded on keratins and annexins; Factor 3 tracked cellular detection rate; SNF and iCluster only separated discrete subpopulations. · MOFA reveals coordinated changes in single cells…; Details on the scMT analysis

  4. ShownMOFA was more accurate and faster than GFA (group factor analysis) and iCluster, and imputed scattered missing values and entirely missing drug-response assays better than common imputation methods.

    In simulations GFA and iCluster inferred redundant factors and recovered cross-view activity less accurately; CLL training took 45 minutes (MOFA) vs 34 hours (GFA) and 5-6 days (iCluster); on N=121 complete CLL samples with masked drug-response data, MOFA had lower mean squared error than feature-wise mean, SoftImpute and k-nearest neighbours, and was more robust than GFA. · Model validation and comparison on simulated…; MOFA identifies outlier samples and accurately…; Imputation; Model comparison

  5. ShownThe factors also worked as clinical tools: continuous Factor 1 refined binary IGHV labels and flagged outlier patients, and the 10 factors predicted time to next treatment better than principal components.

    Factor 1 agreed with clinical IGHV status for 176 out of 200 patients, classified 12 unlabelled ones and reassigned 12 (nine borderline; three discordant, confirmed as outliers by independent drug and exome data); Factors 1, 7 and 8 associated with time to next treatment (Cox, FDR<1%, N=174); 10-factor Cox model beat PCA in 5-fold cross-validated C-index. · MOFA identifies outlier samples and accurately…; Latent factors inferred by MOFA are predictive…; Survival Analysis

Methods, data and reproducibility

Bayesian group factor analysis factorising each omic view into shared sample factors and view-specific weights, with automatic relevance determination plus spike-and-slab sparsity priors, Gaussian, Poisson or Bernoulli likelihoods per view, and mean-field variational inference; low-variance factors are pruned.

  • CLL cohort: somatic mutations (targeted plus whole-exome sequencing), RNA-seq, methylation arrays and ex-vivo drug screens; D=310 drug, 69 mutation, 5000 mRNA, 4248 CpG features · human (CLL patient samples) · n = 200 patients (121 with all assays) · EGAS00001001746 (European Genome-phenome Archive)
  • CLL patients with time-to-next-treatment records used for Cox survival modelling · human · n = 174 (96 uncensored)
  • Validation data for discordant IGHV cases not used in training: ONO-4509 drug response and whole-exome IGHV mutation counts · human · n = three outlier samples, compared with other M-CLL and U-CLL samples
  • Parallel single-cell DNA methylation and transcriptome data of mouse embryonic stem cells grown in 2i or serum · mouse · n = 87 cells (16 in 2i, 71 in serum) · GSE74535 (GEO)
  • Data simulated from the MOFA generative model, varying views, features, factors, missingness and likelihoods, plus mixed-likelihood comparison sets · n = 10 realizations per configuration
Methods & tools
MOFA with MOFAtools R package, GFA (group factor analysis), iCluster, SNF, PCA, DESeq2 (size factors, variance-stabilizing transformation), SoftImpute, k-nearest neighbour imputation, glmnet (ridge-penalised Cox regression), survminer (maximally selected rank statistics), Reactome gene sets, CellTiter Glo ATP-based drug response assay
Experimental conditions
CLL model training: 45 minutes for MOFA vs 34 hours for GFA and 5-6 days for iCluster; in scalability benchmarks iCluster needed on average more than 200 minutes in every setting except the lowest number of views; hardware not stated.
Reproducibility gaps
No software versions, containers or hardware details are given; full simulation parameters are in Appendix Table S1 rather than the main text.
Limitations (6)
  • MOFA is linear, so strongly non-linear dependencies among features, within an assay or between assays, can be missed; non-linear extensions would trade off complexity, speed and interpretability. · stated by the authors
  • Only Gaussian, Poisson and Bernoulli likelihoods are implemented; zero-inflated data (e.g. scRNA-seq) and binomial data (e.g. splicing events) would require new noise models. · stated by the authors
  • Inference is approximate and downstream analyses use posterior point estimates, so estimation uncertainty is not carried through to diagnostics or downstream results. · stated by the authors
  • The objective is non-convex, so a global optimum is not assured; reported results rely on 10-25 random restarts, keeping the model with the highest evidence lower bound (ELBO). · stated by the authors
  • Method comparisons simulate from MOFA's own generative model, including non-Gaussian views that GFA cannot model, which may favour MOFA; benchmark accuracy is reported only in appendix figures, without numeric values in the main text. · noted in this summary
  • The oxidative-stress reading of Factor 5 rests on gene set enrichment and drug loadings without functional validation, and the single-cell application uses one dataset of 87 cells. · noted in this summary

Read this if you integrate incomplete multi-omic or single-cell multi-modal datasets and want a sparse, interpretable factor model linking latent axes to features, clinical markers and outcomes.

Citations
1,508
FWCI
29.07
Field percentile
top 1%

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISCapraz '24Brown '23Argelaguet '20Iperi '24Qoku '25Hirst '25
74% semantic matchBioinformatics
Tümay Capraz, Harald S. Vöhringer, Klaus Sebastian Augusto Kruger Serrano+3 more·bioRxiv·Oct 10, 2024· 5 citations