Bio Multi Omics Mixomics Analysis

Version Compatibility

Reference examples tested with: mixOmics 6.26+.

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

• R: packageVersion('mixOmics') 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 DIABLO model is block.splsda - there is no function literally named diablo. mixOmics input is samples-by-features (the opposite of MOFA2's features-by-samples), and scale=TRUE is the default, so feed appropriately transformed but not pre-standardized matrices.

mixOmics Multi-Omics Analysis

"Find a cross-omic signature that discriminates my groups" -> Project the blocks onto sparse latent components that maximize an association criterion - because these methods maximize covariance, not truth, so a signature that discriminates the training cohort is guaranteed and only out-of-sample replication makes it real.

• R: block.splsda() (DIABLO, supervised), spls() (pairwise), mint.splsda() (multi-study)

Scope: supervised multi-block discriminant integration (DIABLO), sparse pairwise correlation (sPLS), regularized CCA (rCCA), and multi-study integration (MINT). Unsupervised factor models -> mofa-integration. The method-selection decision -> integration-design. Generic cross-validation / leakage theory -> machine-learning/model-validation. Per-omic scaling/batch -> data-harmonization. Enrichment of selected features -> pathway-analysis/go-enrichment.

The Single Most Important Modern Insight -- DIABLO Maximizes Covariance, Not Truth, So a Signature Scored on the Data That Selected It Is an Artifact

These are projection methods: they find linear combinations of features that optimize an association criterion (covariance for PLS/DIABLO, correlation for CCA), and DIABLO additionally optimizes whatever the design matrix tells it to. With n much smaller than p the methods can fit almost anything, so the output is never "the cross-omic drivers of the disease" - it is the features that best satisfied the chosen criterion on the available samples. Three rules follow:

1. Cross-validation must wrap feature selection, not follow it. Tuning keepX on the full data and then reporting perf() cross-validation error on the same data is leakage (up to ~0.15 AUC inflation). mixOmics selects inside folds when tuning, which is correct for choosing keepX - but the minimized CV error it returns is still optimistic. The honest number comes from an external test set never touched during tuning, or a fully nested CV.
2. The design matrix is the central decision, not a default. The off-diagonal weights in [0,1] trade discrimination against cross-block correlation: near 1 gives biologically coherent, inter-correlated signatures at the cost of classification; near 0 gives the best classification with a disconnected network. The tutorials' 0.1 leans toward prediction and is not a recommendation from the DIABLO paper - choose it from the goal and report it.
3. The selected features are candidates, not biomarkers. selectVar discriminates the training cohort by construction; that is the null result, not the finding. Replication in an independent cohort (or MINT across studies) is the finding.

Tool Taxonomy

Decision Tree by Scenario

Set Up Matched Blocks

Goal: Guarantee the row-by-row sample matching DIABLO/sPLS/rCCA require, so the model correlates feature vectors of the same individuals.

Approach: Intersect to common samples and verify identical rowname order across every block; for combining one omic across cohorts, that is horizontal integration and needs MINT, not DIABLO.

library(mixOmics)

common <- Reduce(intersect, list(rownames(X_rna), rownames(X_prot)))   # samples x features
X_blocks <- list(RNA=X_rna[common, ], Protein=X_prot[common, ])
Y <- factor(pheno[common, 'Condition'])
stopifnot(identical(rownames(X_blocks$RNA), rownames(X_blocks$Protein)))   # matched rownames or the result is garbage

sPLS: Sparse Pairwise Correlation

Goal: Select a sparse set of features that covary between two omics.

Approach: Choose mode deliberately - 'canonical' for two omics on equal footing (the CCA-like symmetric framing), 'regression' (the default) only when one block is a designated response. Tune component count, then fit with keepX/keepY feature selection.

spls_res <- spls(X_blocks$RNA, X_blocks$Protein, ncomp=3, mode='canonical',   # symmetric; default 'regression' treats Protein as a response of RNA
                 keepX=c(50, 50, 50), keepY=c(30, 30, 30))
plotVar(spls_res, comp=c(1, 2))

DIABLO: Supervised Multi-Block Signature

Goal: Find a cross-omic feature signature that discriminates a known outcome, with the correlation-versus-discrimination trade-off chosen and reported.

Approach: Set the design matrix from the goal (high off-diagonal for coherent networks, low for prediction), tune the component count on a non-sparse model, tune keepX inside cross-validation folds using balanced error rate, fit, then report performance on an external set.

design <- matrix(0.5, nrow=length(X_blocks), ncol=length(X_blocks),
                 dimnames=list(names(X_blocks), names(X_blocks)))   # 0.5-1 favors cross-block correlation; <0.5 favors prediction - choose from the goal
diag(design) <- 0

ncomp_fit <- perf(block.plsda(X_blocks, Y, ncomp=5, design=design),               # tune ncomp on a NON-sparse model first
                  validation='Mfold', folds=10, nrepeat=10)                       # nrepeat>=10; a single split is noise

tune <- tune.block.splsda(X_blocks, Y, ncomp=2, design=design,
                          test.keepX=list(RNA=c(10, 25, 50), Protein=c(10, 25, 50)),
                          validation='Mfold', folds=10, nrepeat=10,
                          measure='BER', BPPARAM=BiocParallel::MulticoreParam(workers=4))   # BER, not overall error, for imbalanced classes; cpus= is defunct, use BPPARAM
diablo <- block.splsda(X_blocks, Y, ncomp=2, keepX=tune$choice.keepX, design=design)

The features chosen here discriminate the training cohort by construction. For an honest accuracy, hold out an external test set never used in tuning and report predict() / auroc() on it; the perf() error on the tuning data is optimistic because keepX was chosen to minimize it.

Interpret and Validate

Goal: Extract the signature as candidates and visualize cross-block structure without overclaiming.

Approach: Pull selected variables per block and component, inspect inter-block correlations, and frame the list as cohort-specific candidates requiring replication.

sel_rna  <- selectVar(diablo, block='RNA', comp=1)$RNA$name        # candidate features, not validated biomarkers
circosPlot(diablo, cutoff=0.7)                                     # inter-block correlations of the selected features
auc <- auroc(diablo, roc.block='RNA', roc.comp=1)

MINT: Multi-Study (Horizontal) Integration

Goal: Build a signature for ONE omic that replicates across cohorts by modeling study as a known effect.

Approach: Pass a study factor so the model accounts for study-specific variation; this is horizontal integration (same features, different cohorts), the opposite of DIABLO's vertical matched-sample design.

mint_res <- mint.splsda(X=X_rna, Y=Y, study=study, ncomp=3, keepX=c(50, 50, 50))   # study = fixed effect; one omic, many cohorts
plotIndiv(mint_res, study='global', legend=TRUE)

Per-Method Failure Modes

CV that follows selection instead of wrapping it

Trigger: tuning keepX on all samples, then reporting perf() CV error on all samples. Mechanism: the features were chosen with knowledge of every sample, so no fold is truly held out. Symptom: an excellent CV error that collapses in a new cohort. Fix: external test set or nested CV; the number used to pick keepX is not an estimate of performance.