Multi-Omics Integration Design - Usage Guide

Overview

This skill is the decision layer that runs BEFORE any integration tool. It maps a biological question to a method class, names the sample correspondence (vertical/paired, mosaic, horizontal, or diagonal), enforces the small-n / large-p discipline that makes a held-out cohort the real endpoint, and runs the per-view variance-imbalance diagnostic that exposes a "shared" factor that is really one omic in disguise. It exists because the most consequential multi-omics mistakes are made in the choice of strategy, not in the syntax of a tool: feeding horizontal (same-feature, multi-cohort) data to a vertical method, reporting an unvalidated in-cohort signature as a biomarker, or mistaking the dominance of the biggest omic for integration.

It is method-agnostic and routes to the tool skills: mofa-integration (unsupervised shared factors), mixomics-analysis (supervised DIABLO signatures, sPLS pairs, MINT multi-study), and similarity-network (patient stratification by network fusion). Single-cell multimodal integration is a different paradigm owned by single-cell/multimodal-integration.

Prerequisites

BiocManager::install(c('MultiAssayExperiment', 'SummarizedExperiment'))

Conceptual prerequisites and notes: - The omics blocks must be measured on a SHARED sample axis (the same patients/specimens). If instead the same features were measured in different cohorts, that is horizontal integration (meta-analysis / batch correction), not this category. - Per-omic normalization happens FIRST, in each omic's own category (RNA-seq -> VST/logCPM; methylation -> M-values; proteomics -> log2 + MNAR-aware imputation; metabolomics -> log + Pareto). This skill assumes each block is already on a sane scale; the cross-omic harmonization decision is owned by data-harmonization. - Bulk multi-omics is almost always small-n / large-p: tens of samples and 10^4-10^6 features. That regime, not the choice of tool, dominates what is trustworthy. - A MultiAssayExperiment is the natural container for the paired-vs-mosaic decision; matrix orientation differs by downstream tool (MOFA2 wants features x samples, mixOmics wants samples x features).

Quick Start

Tell your AI agent what you want to do: - "I have RNA, protein, and methylation on the same patients - which integration method fits?" - "Is this vertical or horizontal integration?" - "My cohort is mosaic - should I intersect to complete cases or model the missingness?" - "Check whether one omic dominates my integration before I trust the factors" - "How do I validate an integrated subtype - is in-cohort cross-validation enough?"

Example Prompts

Method selection from the question

"I have transcriptomics, proteomics, and methylation on 70 tumors and want to discover molecular subtypes with no predefined labels. Which method class should I use, and what is the honest way to claim the subtypes are real rather than a clustering artifact?"

Correspondence and mosaic decision

"Three RNA-seq cohorts of the same disease, same genes - my collaborator wants to 'integrate' them with MOFA. Is that the right tool, or is this a different kind of integration? If different, where should it go?"

The n<<p validation discipline

"I have a DIABLO signature that classifies responders with 0.95 AUC in cross-validation on 44 samples. How much should I trust it, and what would make it credible?"

Variance-imbalance diagnostic

"Before I interpret my MOFA factors, check whether my 850k-CpG methylation block is dominating the shared latent space relative to my 200-metabolite block, and tell me what to do if it is."

What the Agent Will Do

1. Establish the correspondence: confirm the blocks share a sample axis (vertical) versus share features across cohorts (horizontal), and assemble a MultiAssayExperiment to quantify how mosaic the cohort is.
2. Map the question to a method class: subtype discovery (SNF/iCluster), shared axes of variation (MOFA2/JIVE), predictive signature (DIABLO), or pairwise correlation (sPLS), routing single-cell and horizontal cases out.
3. Surface the n<<p regime: count samples versus features, and set the expectation that tuning needs repeated cross-validation and that a held-out cohort is the endpoint.
4. Run the variance-imbalance diagnostic: compare per-block variance and feature counts, and plan to read the post-fit per-view variance-explained table; recommend equalization or SNF if one omic dominates.
5. Hand off to the chosen tool skill with the correspondence, scaling, and validation constraints made explicit.

Tips

• Name the correspondence before naming a tool. If the answer is "same features, different cohorts," stop - that is meta-analysis, not vertical integration.
• Treat an unvalidated integrated signature as hypothesis-generating, never as a biomarker. At small n the in-cohort fit is guaranteed and uninformative.
• Always read the per-view variance-explained table. If every shared factor loads on one omic, the integration re-discovered the biggest data type; equalize the blocks or move to SNF.
• Prefer modeling the missingness (MOFA2) over intersecting a mosaic cohort to complete cases, because discarding incomplete samples is expensive when n is already small.
• There is no universally best integration method; benchmarks disagree by design. Cross-check a headline result with a second method class - a real subtype should appear in both a MOFA map and an SNF clustering.
• Choose the DIABLO design matrix from the question a priori (discrimination-first low, mechanistic-correlation-first high) and report the classification cost; tuning it until the result looks "integrated" is p-hacking by hyperparameter.

Related Skills

• mofa-integration - Unsupervised shared-factor discovery (the default tool once correspondence is vertical)
• mixomics-analysis - Supervised DIABLO signatures, sPLS pairs, and MINT multi-study integration
• similarity-network - Patient stratification via similarity network fusion
• data-harmonization - Per-block normalization, scaling, batch, and the mosaic missing-omic decision
• single-cell/multimodal-integration - Single-cell CITE-seq/Multiome integration (different paradigm)
• differential-expression/batch-correction - Horizontal same-feature meta-analysis and batch correction
• machine-learning/model-validation - Cross-validation and overfitting theory for supervised integration
• clinical-biostatistics/survival-analysis - Survival validation of discovered subtypes
• workflows/multi-omics-pipeline - End-to-end multi-omics integration pipeline