Skip to content

omicsGMF

Generalized matrix factorization for omics dimensionality reduction — performs matrix factorization for dimensionality reduction, visualization, and imputation of RNA-seq, proteomics, and single-cell omics data without requiring log-transformation, with built-in batch-effect correction via covariate adjustment and missing-value handling.

Talk at EuroBioC2025

  • Speaker: Lieven Clement (Ghent University) — senior author
  • Day / session: Day 1 (Wed Sep 17, 2025) — morning short-talk session, after the Crowell keynote
  • Talk title: "omicsGMF for proteomics dimensionality reduction"
  • Slides (Zenodo DOI): to verify after publish
  • Video: not recorded
  • Abstract / schedule: eurobioc2025 schedule

What it does

omicsGMF takes counts (or intensities) plus a sample-covariate matrix and returns a low-dimensional factorization that simultaneously denoises, imputes, and corrects for known batch covariates. It accepts SummarizedExperiment, SingleCellExperiment, and QFeatures inputs, so it slots cleanly into existing Bioconductor pipelines — RNA-seq, bulk proteomics, single-cell proteomics — without an upstream log step.

How it works

Core idea. Generalized matrix factorization (GMF) on the sgdGMF engine: the assay matrix is modeled as the sum of an offset from sample/feature covariates and a low-rank factor product, fit by stochastic gradient descent with adaptive learning rates. The distribution family is user-selectable (e.g. gaussian() for log-intensity proteomics, count families for RNA-seq), and missing values are handled natively inside the likelihood rather than imputed upstream.

Inputs / outputs. Input: a SummarizedExperiment / SingleCellExperiment / QFeatures with an assay matrix, plus optional sample-level (X) and feature-level (Z) covariate matrices for batch adjustment. Output: factor scores and loadings stored back on the object, plus per-cell imputed values via imputeGMF and a 2-D plot via plotGMF.

Key innovation. Operates directly on raw counts / non-log intensities (no upstream log transform required), handles missingness inside the model (no prior imputation), and absorbs batch covariates into the factorization itself — three things conventional PCA cannot do simultaneously.

Parameters worth knowing. family — likelihood (e.g. gaussian() for proteomics intensities). ncomponents — number of latent factors; pick via runCVGMF cross-validation or calculateRankGMF (a scree-plot analogue on deviance residuals). X, Z — sample- and feature-level covariate matrices for batch / known-effect adjustment. assay.type — which assay to factorize.

Canonical example. Proteomics vignette: simulated Gaussian-distributed proteomics intensities (20 samples × 50 features) with 30% artificial missingness and an injected batch effect. The workflow runs runCVGMF over ncomponents = 1:5 to pick the rank, then runGMF to fit the factorization with the batch covariate in X — the resulting plotGMF shows samples grouped by biology with the batch axis absorbed into the covariate term.

Where it fits in the corpus

  • AACR 2026: no current dossier; relevant axis is bioinfo / AI methods
  • GBCC 2025: queued
  • iSEE (entry) — natural visualization substrate for the factorization output

Notes

Lieven Clement's group (UGent) is a frequent EuroBioC contributor — also behind msqrob2. omicsGMF complements that line of work by being agnostic to the input modality.