Bio Multi Omics Integration Design

Version Compatibility

Reference examples tested with: MultiAssayExperiment 1.36+, SummarizedExperiment 1.40+.

Before using code patterns, verify installed versions match. If versions differ:

• R: packageVersion('<pkg>') then ?function_name to verify parameters

If code throws ImportError, AttributeError, or TypeError, introspect the installed package and adapt the example to match the actual API rather than retrying.

The tool versions that matter most are MOFA2 and mixOmics, whose APIs have moved across releases; this skill routes to those tool skills rather than calling them, so the binding version here is MultiAssayExperiment (the container in which the paired-vs-mosaic decision is made).

Multi-Omics Integration Design

"How should I integrate these omics?" -> Map the biological question and the sample correspondence to a method class BEFORE running a tool - because at tens of samples and 10^5-10^6 features a spurious cross-omic signal is the default outcome, not the surprise.

• R: assemble a MultiAssayExperiment, then choose MOFA2 (shared factors) / mixOmics (signature) / SNF (subtypes) by question

Scope: the integration decision itself - method selection, correspondence (paired/horizontal/mosaic/diagonal), supervised-vs-unsupervised mapping, the n<<p discipline, and the variance-imbalance diagnostic. Running the chosen tool -> mofa-integration, mixomics-analysis, similarity-network. Cross-omic preprocessing -> data-harmonization. Single-cell multimodal -> single-cell/multimodal-integration. Horizontal same-feature meta-analysis -> differential-expression/batch-correction.

The Single Most Important Modern Insight -- Bulk Multi-Omics Is a Small-n, Huge-p Discovery Problem Where a Spurious Cross-Omic Signal Is the Default

A typical bulk cohort has n = 30-300 samples and 10^4-10^6 features per omic, so after stacking blocks n is smaller than p by three to four orders of magnitude. In that regime an integrated signature that has not been validated out-of-sample is overwhelmingly noise that fit the training samples. The deliverable is never "the integrated signature" - it is question-matched structure that survives three gates, each of which a common failure violates:

1. The correspondence gate. Vertical integration (different omics, SAME samples) and horizontal integration (same features, different cohorts) are different problems. The bulk joint-latent tools are indexed by sample; feed them unpaired same-feature data and they still run but emit factors that are pure batch. Name the correspondence first.
2. The variance gate. Total variance scales with feature count and feature scale, so a 850k-CpG block out-votes a 100-metabolite block and the shared factors become methylation PCs. Inspect the per-factor, per-view variance-explained table every time; if one view dominates every shared factor, the integration re-discovered the biggest omic.
3. The validation gate. With n = 40 and 5-fold CV each fold tests 8 samples, so in-cohort cross-validation is optimistically biased to the point of fiction. An integrated subtype or signature is not credible until it reproduces in an INDEPENDENT cohort. The held-out cohort is the finding.

Organize the analysis around defending these three gates, not around picking a favorite tool.

The Integration Taxonomy -- Three Orthogonal Axes, Not One Label

A method is a point in a 3D space, not a single name. Stating where a method sits on each axis prevents the category's two deepest errors (horizontal/vertical confusion and concatenation at n<<p).

MOFA = intermediate, vertical, unsupervised. DIABLO = intermediate, vertical, supervised. SNF = mixed/transformation, vertical, unsupervised. ComBat-across-cohorts = horizontal, unsupervised harmonization (routes OUT to differential-expression/batch-correction).

Tool Taxonomy

Decision Tree by Scenario

Default when uncertain: assemble a MultiAssayExperiment, confirm vertical paired (or mosaic) correspondence, run MOFA2 for an unsupervised map and read its per-view variance-explained table, then escalate to a supervised (DIABLO) or stratification (SNF) tool only if the question demands it.

Name the Correspondence First

Goal: Decide whether the data is a job for this category at all, and whether to model the missingness or intersect to complete cases.

Approach: Assemble the blocks into a MultiAssayExperiment (it coordinates assays, a sample map, and colData), then read off whether samples are fully paired, mosaic, or actually horizontal. Only vertical-paired and mosaic belong here.

library(MultiAssayExperiment)

mae <- MultiAssayExperiment(experiments=ExperimentList(rna=rna_mat, prot=prot_mat, methyl=methyl_mat),
                            colData=clinical)
upsetSamples(mae)                 # visualize which samples have which omics (mosaic structure)
table(complete.cases(mae))        # how many samples have EVERY omic
paired <- intersectColumns(mae)   # complete-case fallback - counts the n it would cost

If complete.cases keeps most samples, complete-case methods (mixOmics, SNF) are fine. If a large fraction is mosaic, prefer MOFA2 (it models missing-view samples in its likelihood) over intersecting, because at n<<p discarding incomplete samples is expensive and imputing a whole block fabricates data (data-harmonization owns that decision).

The Variance-Imbalance Diagnostic

Goal: Detect, before trusting any shared factor, whether one omic is set to dominate the integration purely because it has more features or larger scale.

Approach: After per-feature scaling, compare each block's total variance and feature count; a block contributing the overwhelming majority of stacked variance will hijack the shared latent space. The definitive check is the per-view variance-explained table that MOFA2 reports after fitting - if every factor loads on one view, equalize the blocks (MFA weighting, per-block keepX, or move to SNF) and refit.

block_var <- sapply(assays_list, function(x) sum(apply(x, 1, var)))   # total variance per block
share     <- block_var / sum(block_var)
share                                                                 # any block >> others = imbalance risk

A block holding most of the stacked variance is a red flag that concatenation-style integration will re-discover it. This is the single best honesty check in the category; never skip the post-fit per-view variance read-out.

The n<<p Discipline

The held-out cohort is the endpoint, not in-cohort cross-validation. Three rules follow from n<<p:

• Tune with repeated cross-validation, never a single run. At n = 40 a single CV estimate is mostly noise; mixOmics perf/tune.* take nrepeat (10-50) - use it. Generic CV/overfitting theory lives in machine-learning/model-validation.
• Report out-of-sample performance, not the in-sample fit. A supervised signature's training-set discrimination is guaranteed by construction; only an independent cohort makes it a biomarker.
• An unsupervised factor that correlates with the outcome is a hypothesis. MOFA found it without the label, which is a strength - but calling it predictive requires held-out validation, not the in-cohort correlation that found it.
• Prefer regularized/sparse methods over early concatenation plus plain CCA. Classical CCA divides out within-block variance and, when p > n, achieves correlation 1 trivially by overfitting; sparse PLS (covariance plus an L1 penalty) and sparse factor models stay identifiable, so the regularization is what makes the fit real, not a stylistic choice.

Per-Method Failure Modes

Horizontal data fed to a vertical method

Trigger: running MOFA/DIABLO/SNF on same-feature, multi-cohort data ("integrate my three RNA-seq studies"). Mechanism: the shared latent is indexed by sample and has nothing to align across feature-identical cohorts. Symptom: the tool runs and the top factors track cohort/run, not biology. Fix: recognize this as horizontal integration; use MINT, ComBat/sva, or differential-expression/batch-correction.

Unvalidated in-cohort signature reported as a result

Trigger: reporting a DIABLO panel or a MOFA-factor-vs-outcome correlation from one cohort. Mechanism: at n<<p thousands of cross-omic feature pairs clear any threshold under the null; in-cohort CV is optimistically biased. Symptom: a beautiful signature that fails to replicate. Fix: hold out an independent cohort; frame an unvalidated finding as hypothesis-generating, never as a biomarker.