Skip to content

Susan Holmes — "Latent variables as the best medicine for heterogeneity"

Affiliation: Stanford University, Department of Statistics Day / session: Day 1 (Wed Sep 17, 2025) — afternoon keynote Talk title: "Latent variables as the best medicine for heterogeneity" Slides (Zenodo DOI): to verify after publish Video: Bioconductor YouTube — plenaries playlist (keynotes recorded) Abstract / schedule: eurobioc2025 schedule

Thesis

Biological data — microbiome, single-cell, longitudinal multi-omics — is heterogeneous in ways that defeat both naive pooling and naive stratification: there is shared latent structure across samples that simple mixed-effects models miss and that fully separate per-stratum analyses cannot borrow strength across. The talk argues that latent-variable methods (factor models, generalized matrix factorization, latent-Dirichlet-style topic models, mixOmics-class integrative factorizations) are the right level of abstraction — they let heterogeneity be modeled as structure rather than noise, and they generalize across modalities.

Speaker context

Susan Holmes is one of the most influential statisticians for biological data of the last two decades — co-author of Modern Statistics for Modern Biology (with Wolfgang Huber), longtime collaborator on microbiome statistics (phyloseq lineage), and a recurring methodological voice on multi-table / multi-omics integration. Her group has produced foundational work on multivariate methods for compositional and high-dimensional data, longitudinal analysis of pregnancy / preterm-birth cohorts, and the statistical foundations behind several Bioconductor workflows. Inviting her places the keynote slot squarely on statistical method rather than application — a deliberate signal in a program that otherwise leans heavily on platforms and packages.

Connections to the corpus

  • omicsGMF — generalized matrix factorization is exactly the latent-variable family the keynote is about; this is the most direct package-level link in the EuroBioC2025 program
  • consICA — consensus independent component analysis; another latent-variable method shipping in the same release
  • edgeR, MOSCLIP — multi-omics methods that benefit from latent-structure framing
  • AACR axis: bioinfo-tools (the statistical-methods sub-axis); virtual-cells (latent representations are the substrate any "virtual cell" model would learn over)

Open questions

  • Which application(s) carry the talk — microbiome, single-cell, longitudinal multi-omics, or all three as a tour? Holmes's recent work has spanned all of them.
  • Does she engage directly with the deep-learning latent-variable literature (VAE / scVI-class models) or stay within classical / Bayesian factorization? The framing matters for how it lands with a younger audience.
  • Concrete software recommendations or method-pluralism? Holmes tends toward the latter, which would shape any cross-links to specific Bioconductor packages.