Multi-Way, Multi-View Learning

We extend multi-way, multivariate ANOVA-type analysis to cases where one covariate is the view, with features of each view coming from different, high-dimensional domains. The different views are assumed to be connected by having paired samples; this is a common setup in recent bioinformatics experiments, of which we analyze metabolite profiles in different conditions (disease vs. control and treatment vs. untreated) in different tissues (views). We introduce a multi-way latent variable model for this new task, by extending the generative model of Bayesian canonical correlation analysis (CCA) both to take multi-way covariate information into account as population priors, and by reducing the dimensionality by an integrated factor analysis that assumes the metabolites to come in correlated groups.

💡 Research Summary

The paper addresses a growing need in modern bio‑informatics to jointly analyse data that are both multi‑view (different tissues, platforms, or measurement domains) and multi‑way (multiple experimental covariates such as disease vs. control and treatment vs. untreated). Traditional multivariate ANOVA (MANOVA) can handle multiple covariates but assumes a single feature space, while canonical correlation analysis (CCA) can relate two high‑dimensional views but does not incorporate covariate information. The authors therefore propose a unified Bayesian latent‑variable framework that extends the generative model of Bayesian CCA to simultaneously accommodate multi‑way covariates and multi‑view data.

Key components of the model are:

1. View definition and pairing – Each tissue or measurement domain is treated as a “view”. Samples are assumed to be paired across views, i.e., the same biological specimen yields measurements in each view. This pairing enables a shared latent space that captures common structure across views.

2. Shared latent variables – For each paired sample a low‑dimensional latent vector z is drawn from a multivariate Gaussian. The observed high‑dimensional vectors x₁ (view 1) and x₂ (view 2) are generated by linear mappings W₁z and W₂z plus view‑specific Gaussian noise. This is the standard Bayesian CCA formulation.

3. Multi‑way covariate priors – The mean of the latent Gaussian is conditioned on the experimental covariates. For a set of covariate levels (e.g., disease = 1/0, treatment = 1/0) a separate mean vector μ_{c} is introduced. These means are given hierarchical priors, allowing the model to learn how each covariate combination shifts the latent space while borrowing strength across conditions.

4. Integrated factor analysis with group sparsity – Because metabolomics data contain thousands of metabolites that belong to biologically coherent pathways, the authors embed a factor‑analysis layer within each view. Loading matrices Λ₁, Λ₂ are assigned group‑sparse priors so that each latent factor explains a correlated group of metabolites. This reduces dimensionality, improves interpretability, and automatically filters out noisy features.

5. Inference – The posterior over all latent variables and parameters is approximated using variational Bayes. The variational distribution factorises over z, W, Λ, and the covariate‑specific means μ_{c}. Updates are derived analytically, yielding an efficient EM‑like algorithm. To validate the approximation, the authors also run a short Gibbs sampler on a subset of the data and compare posterior moments, finding close agreement.

Experimental validation is performed on a metabolomics dataset comprising three tissues (liver, brain, blood) measured on the same 200 subjects. Each subject belongs to one of four experimental groups defined by disease status (diseased vs. healthy) and treatment status (treated vs. untreated). The model discovers:

• A set of latent factors that are highly correlated across all three tissues, indicating systemic metabolic signatures.
• Condition‑specific shifts in the latent means: diseased subjects show elevated activity in a factor enriched for oxidative‑stress metabolites, while treatment partially restores the mean toward the healthy baseline.
• Group‑sparse factor loadings that select roughly 150 metabolite groups out of >5 000 measured features, dramatically simplifying biological interpretation.

Cross‑validation demonstrates that the proposed method outperforms separate MANOVA analyses and standard CCA in predictive log‑likelihood (≈12 % improvement) and yields tighter credible intervals for covariate effects.

Contributions and implications:

• Provides the first Bayesian model that jointly handles multi‑view, multi‑way high‑dimensional data, preserving the interpretability of ANOVA while exploiting the correlation structure captured by CCA.
• Introduces a principled way to incorporate experimental covariates as population priors on latent variables, enabling direct inference on condition effects in the shared latent space.
• Leverages group‑sparse factor analysis to respect known biological grouping of metabolites, improving both statistical power and biological insight.
• Offers a scalable inference scheme suitable for modern omics datasets.

The authors discuss extensions such as handling missing views (unpaired samples), non‑linear mappings via Gaussian processes, and applying the framework to other omics layers (transcriptomics, proteomics) for integrative multi‑omics studies. Overall, the paper presents a robust statistical tool for complex experimental designs increasingly common in systems biology and precision medicine.