Skip to content

msqrob2

Robust statistical inference for quantitative LC-MS proteomics — robust linear mixed model framework for differential abundance assessment in mass spectrometry proteomics; supports raw peptide intensities or summarized protein values, with stabilization through ridge regression, empirical Bayes variance estimation, and robust M-estimation; the hurdle workflow handles missing data without imputation assumptions.

Workshop at EuroBioC2025

  • Presenter: Christophe Vanderaa (UCLouvain — Lieven Clement's collaborator)
  • Day / session: Day 2 (Thu Sep 18, 2025) — afternoon hands-on workshop (parallel with OSTA, mia, Workshop Platform, iSEE)
  • Workshop title: "msqrob2 for proteomics workflows"
  • Format: hands-on tutorial, attendees follow vignettes in their own R session
  • Workshop materials: to verify — typically a per-event GitHub repo from the Clement group
  • Video: workshops not recorded; the Bioconductor YouTube playlist covers plenaries only
  • Abstract / schedule: eurobioc2025 schedule

What it does

msqrob2 frames LC-MS proteomics differential abundance as a robust linear mixed model on either peptide-level or protein-level intensities, stabilizing variance via ridge penalties and empirical Bayes shrinkage, and handling outliers through M-estimation. The hurdle workflow is the methodological signature: rather than imputing missing peptide intensities (and propagating untestable assumptions), it splits the inference into a presence/absence component and an intensity component combined into a single test.

How it works

Core idea. Fit a robust linear mixed model to peptide intensities with ridge regression on peptide-specific effects and empirical Bayes shrinkage on the residual variance, then test contrasts. M-estimation downweights outliers; ridge regularizes the many peptide-level random effects; empirical Bayes borrows variance information across proteins.

Inputs / outputs. Input: a QFeatures object that holds the peptide-level intensity matrix, sample metadata, and the peptide-to-protein mapping. Output: per-protein effect estimates, standard errors, t-statistics, p-values, and FDR for each contrast specified in the design — attached back to the QFeatures rowData so the chain is traceable.

Key innovation. The Stage1 / Stage2 decomposition splits the problem cleanly: Stage1 aggregates peptide-level information into protein-level summaries while estimating MSqRob parameters; Stage2 performs hypothesis tests on those protein summaries. The hurdle workflow then handles missingness without imputation by combining a presence/absence component with an intensity component into a single test, avoiding the untestable MAR/MCAR assumptions standard imputation requires.

Parameters worth knowing. Design formula (fixed and random effects). Ridge penalty for peptide-specific random effects (regularization strength). Choice of Stage1 aggregation (peptide-to-protein summarization) vs running msqrob directly on peptide intensities. Whether to enable the hurdle path for proteins with substantial missingness. Key functions: msqrob(), msqrobAggregate(), hypothesisTest().

Canonical example. The CPTAC vignette uses a Clinical Proteomic Technology Assessment for Cancer spike-in dataset where UPS1 proteins are added at known differential concentrations into a constant background — ground truth for benchmarking effect-size recovery and FDR control. The workflow loads the data into QFeatures, runs the Stage1 + Stage2 pipeline, and recovers the spiked UPS1 proteins at the expected fold-changes.

Where it fits in the corpus

  • AACR 2026: no current dossier; proteomics is an emerging axis at AACR via translational/biomarker tracks
  • omicsGMF (entry) — same Clement-group lineup; matrix-factorization sister to msqrob2's regression framing
  • ISMB 2026: scaffold
  • GBCC 2025: queued
  • Nextflow Summit 2026: not mentioned (different ecosystem)

Notes

msqrob2's hurdle approach for proteomics missing data is the methodological signature — explicitly avoiding imputation rather than working around it; sister to omicsGMF in the Clement-group methods stack.