Lecture notes. Written from the material of StatOmics Sga21, licensed CC BY-NC-SA 4.0. These are notes, not a transcript: the material has been reorganised and rewritten. This adaptation carries the same licence, and the original is linked above.

25. Case-Study Datasets for msqrob2¶

What this covers¶

This chapter sets up two real datasets used to practice a full differential-expression workflow in quantitative proteomics with msqrob2, on the simplest possible experimental design: two groups to compare. It assumes the reader already has the machinery from earlier chapters — peptide quantification, log-transformation, and the robust-regression summarization model that msqrob2 fits to turn peptide intensities into a protein-level fold change and test statistic — and is now asking the practical question that comes right after building a pipeline: how do you know the pipeline is telling the truth?

Why a spike-in experiment, specifically¶

Any differential-analysis pipeline can be run on real biological data and will return some list of "significant" proteins, but a real biological sample gives no way to check whether that list is right, because the true fold changes are unknown. A spike-in experiment fixes this by constructing a dataset where the ground truth is set by the experimenter rather than estimated: known proteins are added to a background at known, different concentrations across the groups being compared, so the analyst can check the pipeline's output against an answer that is already known before the data are collected.

Any spiked-in protein has a known, non-zero true fold change between the two groups.
Every background protein has a true fold change of zero — it wasn't touched, so any statistical significance found on it is a false positive.

This is exactly the design behind the CPTAC dataset below, and it is why the workflow is checked on it before being trusted on data without a known answer.

The CPTAC spike-in dataset¶

The case study is a subset of Study 6 of the Clinical Proteomic Technology Assessment for Cancer (CPTAC). The Sigma Universal Protein Standard mixture 1 (UPS1) — 48 human proteins — was spiked into a background of Saccharomyces cerevisiae strain BY4741, at two different concentrations:

6A: 0.25 fmol UPS1 protein/μL
6B: 0.74 fmol UPS1 protein/μL

So the comparison "6A vs 6B" has a known answer built in: the 48 UPS1 (human) proteins have a real, non-zero fold change between the two spike-in levels, while the yeast proteins making up the background are the same in both groups and should show no fold change. A pipeline that recovers the spiked proteins as differential and leaves the yeast proteins alone is behaving correctly; one that flags yeast proteins as differential is producing false positives that a real experiment, without this scaffolding, would hide.

The subset used here is restricted to one instrument and site — LTQ-Orbitrap W at site 56 — with three replicates at each of the two spike-in concentrations. The raw files were searched with MaxQuant version 1.5.2.8, with search settings as described in the original methods paper.

Two ways to obtain the data:

The raw instrument files, from the CPTAC data portal (Study 6).
The MaxQuant search output (peptides.txt), from a data archive; participants working in R/Rmarkdown can instead import it directly from the web within their script and skip the download.

The analysis itself is run either through the msqrob2gui Shiny app or through an Rmarkdown script, both built around the msqrob2 package.

The breast cancer dataset¶

The second case study has no built-in ground truth — it is a real clinical comparison, used to practice the same workflow on data where the analyst does not already know the answer. Eighteen oestrogen-receptor-positive breast cancer tissue samples, from patients treated with tamoxifen upon recurrence, were assessed by LC-MS (LTQ-Orbitrap, MaxQuant version 1.4.1.2, searched against the 2012-09 human canonical proteome). Nine patients had a good outcome and nine a poor outcome.

Three versions of the peptides.txt output are provided for this same cohort, at three different sample sizes:

3 vs 3
6 vs 6
9 vs 9

Exercises¶

Using the CPTAC 6A vs 6B data and either the GUI or the Rmarkdown script, run through the analysis workflow step by step, making sure you understand what each step is doing. Since the true fold change is known for both the spiked-in (UPS1) proteins and the yeast background proteins, check the output against that known answer — what do you observe?
Repeat the same analysis using median summarization in place of robust summarization. How do the results compare to the robust summarization, and why?

Sources¶

Notes: docs/omics-statistics/statomics/sga21/pda_tutorialPreprocessing.md (statOmics SGA21 course, "2. Preprocessing and statistical analysis with msqrob2 for experiments with simple designs", sections 2.1–2.3) — the entire chapter is drawn from this file; no slides or transcript were supplied for this session.
Referred to but not contained in the supplied material: the msqrob2 Bioconductor package and its methods papers (Goeminne et al. 2016; Goeminne et al. 2020; Sticker et al. 2020); the CPTAC Study 6 raw data and original description; the GUI walkthrough (cptac_robust_gui.html) and Rmarkdown script (cptac_robust.html) that the exercises are run in; the peptides.txt files for both datasets.

← 24. Experimental Design and Blocking · Contents · 26. Mass Spectrometry-Based Proteomics →