Multi-Omics Data Integration

Introduction

Modern biology rarely measures one thing at a time. A single patient sample can now yield mRNA expression, DNA methylation, copy-number alterations, protein and phospho-protein levels, metabolites, and — at single-cell and spatial resolution — paired transcriptome + surface protein (CITE-seq) or transcriptome + chromatin accessibility (Multiome). Multi-omics data integration is the family of statistical and machine-learning methods that jointly analyze these heterogeneous "views" of the same samples to find shared structure: disease subtypes, patient strata, co-regulated modules, and multi-modal biomarkers that no single assay reveals. The central challenge is that the views differ in scale, noise model, dimensionality (often p ≫ n), and biological meaning, so naïvely concatenating them lets the highest-variance modality dominate. The field is organized below by approach family; each section explains the method, applications, strengths and limitations, with links to the literature. Two large independent benchmarks (Rappoport & Shamir 2018; Cantini et al. 2021) converge on a consistent finding: there is no universal winner — the best method is task- and data-dependent.

Joint latent-variable / factor models

Joint latent-variable models represent omics views as noisy, view-specific projections of a small set of shared latent factors. Integration estimates these shared factors.

How the method works

A generative model links latent factors Z to each view X_m through view-specific loadings W_m (X_m ≈ W_m Z + noise_m), with a per-view noise model:

• iCluster / iClusterPlus / iClusterBayes — joint latent Gaussian (later GLM / fully Bayesian) model; samples clustered in the shared space; Bayesian variants use sparsity priors + MCMC, each view carrying its own likelihood.
• MOFA / MOFA+ — Bayesian group factor analysis with ARD priors + spike-and-slab sparsity; ARD prunes unneeded factors and reveals, per factor, which views it is active in. MOFA+ adds stochastic variational inference scaling to millions of cells + multiple groups.
• JIVE — decomposes total variation into joint structure common to all views + individual private structures + noise.
• MCIA — multivariate ordination maximizing co-inertia (squared covariance) between each view and a synthetic reference; unsupervised, fast.

Tools. iClusterPlus/iClusterBayes (R); MOFA2 (R+Python); r.jive; omicade4/mogsa (MCIA).

Use. Cancer subtyping (breast, lung, ovarian, glioma), patient stratification, factor–survival/covariate association (MOFA's canonical CLL analysis: a factor ↔︎ IGHV status + drug response).

Strengths. Interpretable factors with per-view weights; principled heterogeneous noise; MOFA/ARD auto-downweights junk views; JIVE separates shared vs private.

Limitations. Linear/Gaussian misses nonlinearity; iCluster MCMC is slow/tuning-sensitive (λ, K); factor count K must be chosen; MCIA needs prior standardization or a high-magnitude view swamps others.

Citations. Shen et al. 2009, Bioinformatics 25:2906 (10.1093/bioinformatics/btp543, PMID 19759197); Mo et al. 2013, PNAS 110:4245 (10.1073/pnas.1208949110, PMID 23431203); Mo et al. 2018, Biostatistics 19:71 (10.1093/biostatistics/kxx017, PMID 28541380); Argelaguet et al. 2018, Mol Syst Biol 14:e8124 (10.15252/msb.20178124, PMID 29925568); Argelaguet et al. 2020, Genome Biol 21:111 (10.1186/s13059-020-02015-1, PMID 32393329); Lock et al. 2013, Ann Appl Stat 7:523 (10.1214/12-AOAS597, PMID 23745156); Meng et al. 2014, BMC Bioinformatics 15:162 (10.1186/1471-2105-15-162, PMID 24884486).

Network fusion

Network-fusion methods combine patient-similarity networks, one per data view, into a consensus network for clustering.

How the method works

SNF builds a sample×sample similarity graph per view, then iteratively updates each via cross-diffusion / message-passing — weak view-specific edges wash out, cross-view-supported edges reinforce; spectral-cluster the fused graph. NEMO simplifies to a closed-form average of per-view relative-similarity matrices (no iteration) and natively handles partial data (samples missing some views).

Tools. SNFtool (R/MATLAB/Python); NEMO (R).

Use. Subtype discovery from mRNA + methylation + miRNA (SNF's original GBM/kidney/lung); incomplete-assay cohorts (NEMO).

Strengths. Nonparametric; combines any views with a definable similarity; SNF robust to view-specific noise; NEMO handles missing modalities and is fast.

Limitations. SNF's equal-weight fusion is fragile to a wholly-uninformative view; output is a graph, not per-feature loadings (indirect biomarker attribution); hyperparameters (K, iterations) need tuning.

Citations. Wang et al. 2014, Nat Methods 11:333 (10.1038/nmeth.2810, PMID 24464287); Rappoport & Shamir 2019, Bioinformatics 35:3348 (10.1093/bioinformatics/btz058, PMID 30698637) (NEMO).

Matrix factorization (non-negative)

Non-negative matrix factorization decomposes each view's non-negative matrix into a shared sample-coefficient matrix + view-specific non-negative bases → parts-based, additive "metafeature" modules.

How the method works

Joint NMF minimizes Σ_m ‖X_m − W_m H‖² with non-negativity and H shared across views. intNMF is the canonical integrative-clustering implementation.

Tools. IntNMF (R); joint-NMF in scikit-fusion (Python).

Use. Cancer subtyping; extracting co-varying multi-omics modules.

Strengths. Highly interpretable additive factors; best-in-class for recovering ground-truth clusters in the Cantini benchmark; minimal distributional assumptions.

Limitations. Requires non-negative inputs (shift/transform signed data); non-convex (local optima, init-dependent); rank K must be chosen.

Citations. Chalise & Fridley 2017, PLoS ONE 12:e0176278 (10.1371/journal.pone.0176278, PMID 28459819).

Supervised integration

Supervised integration uses a known outcome/label to guide integration toward jointly-discriminative features across views.

How the method works

DIABLO (in mixOmics) extends sparse PLS / generalized CCA to find sparse latent components across all views that are (a) mutually correlated and (b) predictive of the outcome; L1 sparsity → a compact multi-omics signature.

Tools. mixOmics (R), block.splsda (DIABLO).

Use. Multi-omics biomarker panels/classifiers (e.g., breast-cancer subtype from mRNA+miRNA+methylation+protein).

Strengths. Directly optimizes for the outcome; compact interpretable cross-modal signatures; rich diagnostics/CV.

Limitations. Needs labels (no pure discovery); overfitting risk at small n/high p (nested CV mandatory); linear/correlational; design-matrix is a modeling choice.

Citations. Singh et al. 2019, Bioinformatics 35:3055 (10.1093/bioinformatics/bty1054, PMID 30657866); Rohart et al. 2017, PLoS Comput Biol 13:e1005752 (10.1371/journal.pcbi.1005752, PMID 29099853).

Deep-learning integration

Autoencoders, VAEs, and GNNs learn a shared nonlinear embedding; dominant for single-cell & spatial multi-omics.

Mechanism (by tool).

• MOGONET — one GCN per view (nodes=patients) + a View Correlation Discovery Network fusing cross-omics label correlations; classification + biomarker ID (bulk).
• totalVI — VAE jointly modeling CITE-seq RNA (NB) + surface protein (with background) → denoised, batch-corrected joint latent space.
• MultiVI — VAE integrating paired and unpaired RNA/ATAC/protein, imputing missing modalities (mosaic regime).
• WNN / Seurat v4 — learns per-cell relative modality weights → a single weighted neighbor graph.
• GLUE — VAE integration of unpaired modalities linked by a prior-knowledge guidance graph (ATAC peaks→genes), enabling regulatory inference with no shared cells.
• SpatialGlue — GNN with dual-attention integrating spatial location with each omics (intra) then across omics (cross) to resolve spatial domains.

Tools. MOGONET (PyTorch); scvi-tools (totalVI, MultiVI); Seurat v4 (WNN); scglue; SpatialGlue.

Use. Atlas building; cell-type/state discovery across RNA+protein+chromatin; cross-modality imputation; spatial domains; patient classification (MOGONET).

Strengths. Nonlinear; scale to millions of cells; handle batch, unpaired/mosaic designs, prior graphs; SOTA on single-cell/spatial.

Limitations. Data/compute-hungry (GPU); low interpretability; many hyperparameters + stochastic training; preprocessing-sensitive; over-integration risk. Deep does not automatically beat tuned classical baselines — must be shown per task.

Citations. Wang et al. 2021, Nat Commun 12:3445 (10.1038/s41467-021-23774-w, PMID 34103512); Gayoso et al. 2021, Nat Methods 18:272 (10.1038/s41592-020-01050-x, PMID 33589839); Ashuach et al. 2023, Nat Methods 20:1222 (10.1038/s41592-023-01909-9, PMID 37386189); Hao et al. 2021, Cell 184:3573 (10.1016/j.cell.2021.04.048, PMID 34062119); Cao & Gao 2022, Nat Biotechnol 40:1458 (10.1038/s41587-022-01284-4, PMID 35501393); Long et al. 2024, Nat Methods 21:1658 (10.1038/s41592-024-02316-4, PMID 38907114).

Ensemble / consensus meta-frameworks

Consensus methods combine the results of multiple integration or clustering algorithms, followed by standardized downstream evaluation.

How the method works

Each method produces a clustering; a consensus step (co-association / majority voting) merges into a stable partition, then computes standardized metrics (silhouette, survival, clinical enrichment) — operationalizing "no universal winner." MOVICS wraps ~10 algorithms (iClusterBayes, SNF, MoCluster, CIMLR, IntNMF, ConsensusClustering, LRAcluster, PINSPlus, NEMO, COCA). CancerSubtypes unifies consensus clustering, SNF, iCluster.

Tools. MOVICS (R); CancerSubtypes (R).

Strengths. Robustness through aggregation; standardized reproducible evaluation; lowers barrier to running many methods.

Limitations. Consensus can wash out a genuinely-best single method; still needs K + method choice; cancer-oriented; heavy dependency stacks.

Citations. Lu et al. 2020, Bioinformatics 36:5539 (10.1093/bioinformatics/btaa1018, PMID 33315104); Xu et al. 2017, Bioinformatics 33:3131 (10.1093/bioinformatics/btx378, PMID 28605519).

Benchmarks — how the field is actually measured

• Rappoport & Shamir 2018 (NAR 46:10546, 10.1093/nar/gky889, PMID 30295871): 9 algorithms × 10 TCGA cancers (expression+methylation+miRNA). No algorithm universally wins; in aggregate MCCA gave the best prognostic value, rMKL-LPP the most enriched clinical parameters.
• Cantini et al. 2021 (Nat Commun 12:124, 10.1038/s41467-020-20430-7, PMID 33402734): 9 joint-dimensionality-reduction methods across simulated + TCGA + single-cell. No universal winner; intNMF best at cluster recovery, MCIA most consistent all-rounder.

Choosing an approach

• No universal winner — the single most reproducible message; performance is task- and data-dependent.
• Reference performers by objective: cluster recovery → intNMF; consistent bulk all-rounder → MCIA; interpretable view-attributable factors → MOFA/MOFA+; survival prognosis → MCCA/rMKL-LPP; supervised signatures → DIABLO; robustness → SNF/NEMO + consensus (MOVICS).
• Single-cell & spatial are deep-learning-led (totalVI/MultiVI, Seurat WNN, GLUE, SpatialGlue) — but deep ≠ automatic win over tuned baselines.
• Practical implication: run several methods and compare with a fair quantitative harness on your data.

Applying the methods with BioMate

• Integration methods in one workbench. Seven integration methods run end-to-end on AWS Batch (Nextflow): MOFA2, SNF, intNMF, MCIA, MFA, iClusterPlus, JIVE (StabMap in progress), via a universal CSV interface (features×samples per modality, sample-intersected).
• Head-to-head on real data, scored reproducibly. A worked pilot compared MOFA2 vs SNF (and others) on real CLL (Dietrich 2018) and TCGA cohorts, matched-k, scored by a reproducible metric harness against gold-standard labels (IGHV for CLL, PAM50 for BRCA). It reproduces the literature's message on the user's own data: MOFA2 led on CLL, MCIA led on TCGA — leader changes by cohort, validating Cantini's "MCIA most robust."
• Modality-specific normalization applied automatically (log2-CPM bulk RNA/miRNA, log1p-CP10k scRNA, M-value methylation, TF-IDF scATAC, median-centering proteomics) — precisely the step that fixes the high-magnitude-view-dominance failure mode.
• Validation scope: not a new algorithm, not a new benchmark result; several modalities (MS proteomics, metabolomics, spatial, CNV, microbiome) have the universal interface in principle but are not yet runtime-validated end-to-end. Ongoing development focuses on unifying bulk + single-cell integration under one maintained, your-own-data cloud platform, with experimental feedback informing subsequent analyses.