Fusing genomics, transcriptomics, proteomics, and metabolomics into one disease model — MOFA, similarity network fusion, and pathway-level integration for precision medicine
Multi-omics integration begins with generating each molecular layer on the same set of patients or samples — a design choice that is far harder logistically than it sounds, since DNA, RNA, protein, and metabolite extraction protocols compete for finite tissue.
Genomics layer: • Whole-genome or whole-exome sequencing at 30-100x coverage • Output: variant calls (SNPs, indels, CNVs) per patient
Transcriptomics layer: • Bulk or single-cell RNA-seq; ~20-40M reads per sample • Output: gene-level counts, normalized to TPM/CPM
Proteomics layer: • Tandem mass spectrometry (TMT-multiplexed or label-free) • Output: relative protein abundance, ~5,000-10,000 proteins quantified, 20-40% missing values (stochastic MS sampling)
Metabolomics layer: • LC-MS or GC-MS untargeted profiling • Output: relative intensities of several hundred annotated metabolites plus thousands of unannotated features
Sample-matching constraint: • All four layers must derive from the same biopsy/blood draw and same timepoint • Tissue-limited studies often sacrifice one layer (commonly metabolomics) due to insufficient material • Batch design must randomize patients across sequencing/MS runs to avoid confounding disease signal with technical batch
Before any layer can be compared to another, each must be normalized on its own terms — RNA-seq counts, protein intensities, and metabolite peaks live on completely different numerical and noise scales, and naive concatenation would let the noisiest layer dominate.
Genomics: variant quality filtering (GATK VQSR), removal of low-confidence calls, no cross-sample normalization needed (binary presence/absence)
RNA-seq: library-size normalization (TPM/CPM/DESeq2 median-of-ratios), log2(x+1) transform to stabilize variance, filtering of low-expression genes
Proteomics: median-centering per sample, missing-value imputation — MNAR (missing not at random, low-abundance proteins below detection) handled via left-censored imputation (e.g. QRILC), MCAR handled via KNN
Metabolomics: quantile normalization across samples, log transform, Pareto scaling before multivariate analysis
Batch correction (all layers): ComBat or limma removeBatchEffect models batch as a covariate and removes its additive/multiplicative contribution while preserving biological variance of interest — critical since sequencing runs and MS batches are rarely fully randomized in real clinical cohorts
MOFA generalizes PCA to multiple simultaneous data views: instead of one factorization per omics layer, it finds a small set of latent factors that jointly explain variance across all layers, distinguishing factors that are layer-specific from those that reflect a truly shared biological axis of variation.
MOFA model: • Each omics view Y_m modeled as: Y_m = W_m Z^T + ε_m • Z = shared latent factor matrix (samples × factors), common across all views • W_m = view-specific factor loadings (features × factors) — sparse via automatic relevance determination priors • Variational Bayesian inference estimates posterior over Z and W_m jointly • Output: for each factor, a "variance explained per view" score reveals whether it is genomics-driven, RNA-driven, or genuinely shared
Similarity Network Fusion (alternative): • Build one k-nearest-neighbor similarity network per omics layer (patient × patient) • Iteratively update each network using information diffused from the others (nonlinear message passing) • Converges to one fused similarity network capturing consensus patient relationships across layers • Strength: nonlinear, robust to layer-specific noise; weakness: less interpretable factor loadings than MOFA
Choosing number of factors: MOFA uses an elbow/ARD-shrinkage criterion — factors whose variance explained drops below ~2% in every view are pruned automatically
A latent factor is only useful once it is mapped back onto biology: gene set enrichment applied to each layer's loadings on a given factor reveals whether that axis of patient variation corresponds to a recognizable pathway, and whether the direction of effect is concordant across DNA, RNA, and protein.
For each retained MOFA factor, rank features within each view by loading magnitude, then run gene set enrichment analysis (GSEA) or over-representation analysis against curated pathway databases (KEGG, Reactome, MSigDB hallmark sets).
Concordance check: a factor is biologically credible when the same pathway is independently enriched among top-loading genomics variants, top-loading transcripts, AND top-loading proteins — triangulation across independent measurement technologies dramatically reduces the false-positive rate compared to single-omics enrichment.
Example disease module: a factor loading heavily on mitochondrial electron-transport-chain genes, with concordant DNA variant burden in mtDNA-interacting nuclear genes, decreased RNA expression of Complex I subunits, and accumulation of the metabolite lactate — together describing a coherent oxidative-phosphorylation dysfunction module without any single layer being sufficient to detect it alone.
The final translational step clusters patients directly on their multi-omics latent factor scores, producing molecular subtypes that in published pan-cancer studies stratify survival and treatment response more sharply than histology or any single-omics classifier alone.
Cluster patients on their Z matrix (factor scores) using consensus clustering (repeated k-means or hierarchical clustering with resampling) to find a robust number of subgroups k.
Evaluate cluster stability via cophenetic correlation and silhouette score; validate clinical relevance via log-rank test on survival curves stratified by subtype, and via association with treatment response in a held-out cohort.
Deployment: an integrated subtype classifier trained on the discovery cohort's factor loadings can assign a NEW patient to a subtype using only their omics profile, run through the frozen MOFA loading matrix — enabling prospective clinical use without re-fitting the entire model.