Skip to content

scp + distributional inference (Beerland)

Interpretable distributional inference for single-cell proteomics — a methods extension layered on top of the scp Bioconductor package (UCLouvain-CBIO single-cell-proteomics framework) that models the full distribution of per-cell protein abundance rather than a point estimate, surfacing uncertainty and skew alongside the mean for interpretability.

  • Speaker / lead: Lucas Beerland (affiliation TBD from slides — likely UCLouvain-CBIO Gatto-lab orbit given the scp connection)
  • Underlying framework: scp (Christophe Vanderaa, Laurent Gatto, UCLouvain-CBIO)
  • scp source: github.com/UCLouvain-CBIO/scp
  • Standalone package: no public package name surfaced as of the talk — methods extension presented; package release status TBD
  • Methodological lineage: the talk explicitly builds on "Lindsey's Method" (a distributional regression technique); not the bioRxiv distributional R package per se
  • Status at EuroBioC2025: methods short-talk — work-in-progress, no public package URL yet

Talk at EuroBioC2025

  • Speaker: Lucas Beerland
  • Day / session: Day 1 (Wed Sep 17, 2025) — afternoon short-talk session, 2:45 PM
  • Talk title: "Interpretable Distributional Inference in Single-Cell Proteomics"
  • Slides (Zenodo DOI): to verify after publish
  • Video: not recorded
  • Abstract / schedule: eurobioc2025 schedule

What it does

Single-cell proteomics (SCP) is heteroskedastic, sparse, and right-skewed: a point-estimate-only summary collapses a lot of the cell-to-cell biological signal. The Beerland talk argues for distributional inference — modeling the per-cell protein-abundance distribution (mean, variance, skew, etc.) jointly — using a Lindsey's-method-flavored approach so that downstream differential analyses can flag cells/proteins where the shape of the distribution shifts even when the mean does not. The package is positioned as a method-layer on top of scp's SingleCellExperiment/QFeatures data container rather than a replacement.

How it works

Core idea. Underneath: scp organizes single-cell proteomics data as a hierarchical QFeatures container nesting SingleCellExperiment assays at PSM, peptide, and protein levels, with explicit links between levels so preprocessing decisions propagate upward through aggregation. On top: the Beerland method models each per-cell protein-abundance distribution rather than a point estimate, using a Lindsey's-method-flavored distributional-regression scheme to estimate mean, variance, and higher moments jointly so downstream tests can detect shape shifts (not just mean shifts).

Inputs / outputs. scp input: MS quantification tables (e.g., MaxQuant evidence.txt with PSM IDs, reporter ion intensities across TMT channels, PIF/PEP quality metrics) plus a sample annotation table mapping channels to single cells / carrier / reference / blank — ingested via readSCP(). scp output: a multi-level QFeatures with PSM → peptide → protein assays. Distributional-layer output: per-cell distributional parameters (e.g., location, scale, skew) and contrast tests on those parameters. Exact output object format not specified — package release status TBD.

Key innovation. Treating the per-cell protein-abundance distribution as the first-class statistical object rather than the per-cell mean. The Lindsey's-method route fits the distribution via a Poisson-regression-on-binned-counts trick, which keeps the estimator interpretable and avoids prescribing a fixed parametric family for SCP's heavy-tailed signal.

Parameters worth knowing. Upstream (scp): aggregation choice (aggregateFeaturesOverAssays() PSM → peptide → protein), normalization scheme, contaminant/reverse-sequence filters. Distributional-layer: choice of basis / bin grid for Lindsey's method, which moments to model, and how to contrast distributions across groups — full parameter list pending the package release / paper.

Canonical example. scp ships mqScpData (a MaxQuant evidence table with TMT-11 and TMT-16 multiplexed single-cell samples) and sampleAnnotation as the worked example for the container layer. The distributional-inference layer's canonical demonstration is pending the speaker's slides / forthcoming preprint; the EuroBioC2025 talk used SCP data in the same Gatto-lab orbit as the substrate.

Where it fits in the corpus

  • AACR 2026: axis = single-cell & spatial omics (proteomics flank); distributional methods for SCP are an emerging methodology area
  • scp upstream: scp is the canonical Bioconductor SCP processing package — Beerland's work sits atop it
  • GBCC 2025: no direct SCP-distributional counterpart
  • Nextflow Summit 2026: not mentioned

Notes

No standalone package name was surfaced from the EuroBioC2025 schedule entry or the public summary — the talk is methods-first, with the implementation likely to surface as either an scp extension or a separate small package. Recheck the Zenodo DOI / a follow-up preprint when posted. Listed here as a flagged work-in-progress (external / pre-release) rather than a confirmed Bioconductor package.