26 Capstone: Two End-to-End Case Studies

Every preceding chapter isolated one step. Here the steps become a study — twice, in two different fields, with the same toolchain.

This capstone runs two complete analyses from raw data to biological conclusion: a label-free proteomics experiment and an untargeted metabolomics experiment. Both use real, published datasets bundled in R packages, so you can reproduce every step without downloading anything.

The point of running them side by side is the thesis of this book made concrete: proteomics and metabolomics are different sciences that share one computational grammar in R. Watch how the same containers (SummarizedExperiment, QFeatures), the same normalization and missing-data logic, the same limma modelling, and the same reproducibility discipline serve both — you learn the toolchain once and apply it to either field.

26.0.1 From Pedagogy to Production

This chapter uses small, bundled datasets (UbiLength, msdata) so every code block runs without external downloads. For production-scale versions of the same workflows — running on real PRIDE and MetaboLights data with parameter tuning, multi-file processing, and cluster execution — the companion project r4ms_book/ provides complete, documented pipelines:

Case Study	Pedagogical Version (This Chapter)	Production Version (Companion)
LFQ proteomics	`DEP::UbiLength` (1,182 proteins)	`r4ms_book/analysis/proteomics_analysis.R` — PXD004886 (4,517 proteins × 22 samples, 11-site benchmark)
Tissue proteomics	—	`r4ms_book/analysis/tissue_atlas_analysis.R` — PXD010154 (33,812 proteins × 12 organs)
Clinical DIA	—	`r4ms_book/analysis/clinical_spectronaut_analysis.R` — PXD000547 (paired patient samples)
Untargeted metabolomics	`msdata::metabolomics()` (4 files)	`r4ms_book/analysis/mtbls38_standards_analysis.R` — MTBLS38 (51 compounds with known identity)

The production scripts are the same logic scaled up — they run the same limma models, produce the same volcano plots, and follow the same reproducibility discipline. When you move from learning the toolchain to applying it to your own data, the companion scripts are your starting template.

The One Mistake to Avoid

Assuming your data behaves like the example. The workflow transfers; the parameters do not. Re-tune feature detection, normalization, and imputation to your own study before trusting a single number.

26.1 Learning Objectives

By the end of this chapter you will be able to:

Assemble a full analysis from the individual steps taught in Chapters 4–25
Run a label-free proteomics study end to end: import → QFeatures → filter → normalize → impute → differential abundance → interpretation → report
Run an untargeted metabolomics study end to end: import → xcms feature detection → normalize → annotate → differential abundance → interpretation → report
Recognise which parts of the workflow are shared across omics and which are domain-specific
Produce a reproducible, deposit-ready result for each study

26.2 How to Read This Chapter

Each case study is a guided tour with pointers back to the chapter where each step is taught in depth. If a step is unfamiliar, follow the reference; here we keep commentary short and let the pipeline flow.

Stage	Proteomics tool	Metabolomics tool	Taught in
Import	`readQFeatures()`	`MsExperiment`, `Spectra`	Ch 4–5
Feature generation	search-engine PSMs	`xcms::findChromPeaks()`	Ch 7, 10
Container	`QFeatures`	`SummarizedExperiment`	Ch 5, 12
QC	detection rate, CV	TIC/BPC, QC-pool CV	Ch 6
Normalize	median / VSN	log + median / TIC	Ch 17
Missing data	filter + impute	filter + fill peaks	Ch 18
Identity	protein inference	adduct + mass search	Ch 11, 8–9
Model	`limma`	`limma`	Ch 20
Interpret	enrichment	pathway / networks	Ch 24
Report	Quarto + PRIDE	Quarto + MetaboLights	Ch 25

Note

All code is shown with #| eval: false so the chapter renders without the full analysis stack installed. Each block runs as written against the bundled example data once the packages in the setup chunk are available.

27 Case Study 1 — Label-Free Proteomics

27.1 The Study

The DEP package bundles the UbiLength dataset (Zhang et al.): a label-free experiment comparing HeLa cells in which the deubiquitinase Ubiquitin-specific protease pathway is perturbed (Ubi knockdown) against controls, quantified by MaxQuant LFQ. The biological question: which proteins change in abundance when the pathway is disrupted?

Design: 2 conditions (control vs. Ubi), replicates per condition
Input: a MaxQuant proteinGroups table (real output)
Goal: a ranked, FDR-controlled list of differentially abundant proteins, interpreted and reported

Code

library(DEP)
library(QFeatures)
library(SummarizedExperiment)
library(limma)
library(tidyverse)

data("UbiLength", package = "DEP")   # real MaxQuant proteinGroups table
dim(UbiLength)                        # proteins × columns

27.2 Step 1 — Build the Data Object (Ch 4–5, 12)

Filter contaminants and reverse hits, resolve duplicate identifiers, and construct a SummarizedExperiment/QFeatures with a clean experimental design.

Code

# Remove contaminant + reverse-database hits (standard MaxQuant filtering)
data <- UbiLength |>
  dplyr::filter(Reverse != "+", Potential.contaminant != "+")

# Unique protein identifiers
data <- DEP::make_unique(data, names = "Gene.names", ids = "Protein.IDs")

# LFQ intensity columns and an explicit experimental design
lfq_cols   <- grep("^LFQ.intensity", colnames(data))
experiment <- DEP::UbiLength_ExpDesign        # label / condition / replicate
se <- DEP::make_se(data, lfq_cols, experiment)
se

27.3 Step 2 — Quality Control (Ch 6)

Inspect protein counts per sample, the missing-value structure, and sample clustering before modelling.

Code

DEP::plot_frequency(se)     # proteins identified per sample
DEP::plot_numbers(se)       # per-sample protein counts
DEP::plot_missval(se)       # missing-value heatmap — is it structured (MNAR)?

27.4 Step 3 — Normalize and Handle Missing Data (Ch 17–18)

Variance-stabilise, then impute. The missingness here is largely left-censored (proteins below detection in one condition), so a minimum-based imputation is appropriate — exactly the MNAR reasoning from Chapter 18.

Code

se_norm <- DEP::normalize_vsn(se)
DEP::plot_normalization(se, se_norm)

# Filter proteins missing in too many replicates, then impute the rest
se_filt <- DEP::filter_missval(se_norm, thr = 0)
se_imp  <- DEP::impute(se_filt, fun = "MinProb", q = 0.01)   # MNAR-appropriate

27.5 Step 4 — Differential Abundance (Ch 20)

Fit the limma model and extract an FDR-controlled result. DEP wraps limma internally, so this is the empirical-Bayes moderation of Chapter 20 applied to the protein matrix.

Code

diff <- DEP::test_diff(se_imp, type = "control", control = "Ctrl")
dep  <- DEP::add_rejections(diff, alpha = 0.05, lfc = 1)

# Ranked results table
results <- DEP::get_results(dep)
results |>
  dplyr::filter(significant) |>
  dplyr::arrange(dplyr::desc(abs(Ubi_vs_Ctrl_ratio))) |>
  dplyr::select(name, Ubi_vs_Ctrl_ratio, Ubi_vs_Ctrl_p.adj) |>
  head(15)

27.6 Step 5 — Visualise and Interpret (Ch 20, 23)

A volcano plot for the headline result, then pathway enrichment on the significant set to move from a protein list to biology.

Code

DEP::plot_volcano(dep, contrast = "Ubi_vs_Ctrl", label_size = 3, add_names = TRUE)

# Enrichment on the significant proteins (Ch 24)
library(clusterProfiler)
library(org.Hs.eg.db)
sig_genes <- results |> dplyr::filter(significant) |> dplyr::pull(name)
ego <- clusterProfiler::enrichGO(
  gene = sig_genes, OrgDb = org.Hs.eg.db,
  keyType = "SYMBOL", ont = "BP", pAdjustMethod = "BH"
)
clusterProfiler::dotplot(ego, showCategory = 12)

27.7 Step 6 — Report and Deposit (Ch 25)

Export the figures and table, capture the environment, and prepare the ProteomeXchange deposit.

Code

dir.create("results", showWarnings = FALSE)
readr::write_csv(results, "results/proteomics_DE.csv")
sessioninfo::session_info() |> capture.output() |>
  writeLines("results/proteomics_session.txt")
# Deposit raw + mzIdentML + SDRF to PRIDE / ProteomeXchange (Ch 25)

Result. A reproducible, FDR-controlled list of proteins responding to the perturbation, with enriched pathways and a deposit-ready bundle — built entirely from the containers and models of Parts I–VI.

28 Case Study 2 — Untargeted Metabolomics

28.1 The Study

The faahKO package bundles a classic real LC–MS dataset: spinal-cord extracts from fatty-acid amide hydrolase knockout (FAAH⁻/⁻) mice versus wild-type, in which the knockouts accumulate fatty-acid amides. The question: which metabolite features distinguish knockout from wild-type? This is the canonical xcms demonstration dataset — and a complete study in its own right.

Design: 2 genotypes (KO vs. WT), replicates each
Input: real centroided .CDF/mzML files
Goal: an aligned feature matrix, differential features, and putative annotations

Code

library(xcms)
library(MsExperiment)
library(SummarizedExperiment)
library(faahKO)

cdf_files <- dir(system.file("cdf", package = "faahKO"),
                 recursive = TRUE, full.names = TRUE)
pheno <- data.frame(
  sample_name  = sub("\\.CDF$", "", basename(cdf_files)),
  sample_group = c(rep("KO", 6), rep("WT", 6))
)
mse <- MsExperiment::readMsExperiment(cdf_files, sampleData = pheno)
mse

28.2 Step 1 — Feature Detection with `xcms` (Ch 7)

The full CentWave → alignment → correspondence → gap-filling pipeline of Chapter 7, condensed.

Code

# 1. Peak detection (CentWave)
cwp <- xcms::CentWaveParam(peakwidth = c(20, 80), ppm = 25, snthresh = 10)
xd  <- xcms::findChromPeaks(mse, param = cwp)

# 2. Retention-time alignment (Obiwarp)
xd  <- xcms::adjustRtime(xd, param = xcms::ObiwarpParam(binSize = 0.6))

# 3. Correspondence (group peaks into features across samples)
pdp <- xcms::PeakDensityParam(sampleGroups = pheno$sample_group,
                              minFraction = 0.5, bw = 30)
xd  <- xcms::groupChromPeaks(xd, param = pdp)

# 4. Fill missing peaks so the matrix has no structural gaps
xd  <- xcms::fillChromPeaks(xd)

28.3 Step 2 — Extract the Feature Matrix (Ch 5, 12)

Turn the xcms result into the same SummarizedExperiment used on the proteomics side — the moment the two workflows converge.

Code

se_metab <- xcms::quantify(xd, value = "into")   # SummarizedExperiment
se_metab
# assay(): features × samples; rowData(): mz/rt; colData(): genotype

28.4 Step 3 — QC and Normalization (Ch 6, 17)

Code

# TIC per sample and log-transform + median normalization (Ch 17)
assay(se_metab) <- log2(assay(se_metab) + 1)
sample_medians  <- matrixStats::colMedians(assay(se_metab), na.rm = TRUE)
assay(se_metab) <- sweep(assay(se_metab), 2, sample_medians - mean(sample_medians))

# PCA to confirm genotype separates and no batch structure dominates
pc <- prcomp(t(na.omit(assay(se_metab))), scale. = TRUE)

28.5 Step 4 — Differential Features with `limma` (Ch 20)

The identical modelling code from the proteomics study — the payoff of a shared container.

Code

library(limma)
design <- model.matrix(~ 0 + factor(se_metab$sample_group))
colnames(design) <- c("KO", "WT")
fit <- limma::lmFit(assay(se_metab), design)
fit <- limma::contrasts.fit(fit, limma::makeContrasts(KO - WT, levels = design))
fit <- limma::eBayes(fit)
top <- limma::topTable(fit, number = Inf, adjust.method = "BH")
head(top, 15)

28.6 Step 5 — Annotate the Hits (Ch 8–9)

Convert the significant features’ m/z to putative identities via exact-mass search — the annotation funnel of Chapter 8.

Code

library(MetaboCoreUtils)
sig <- top |> tibble::rownames_to_column("feature") |>
  dplyr::filter(adj.P.Val < 0.05)
sig_mz <- SummarizedExperiment::rowData(se_metab)[sig$feature, "mzmed"]

# Expected m/z for common adducts of candidate fatty-acid amides
adduct_mz <- MetaboCoreUtils::mass2mz(
  MetaboCoreUtils::calculateMass("C18H35NO"),   # e.g. oleamide
  adduct = c("[M+H]+", "[M+Na]+")
)
# Match sig_mz against a reference library within tolerance (Ch 8–9)

28.7 Step 6 — Interpret, Report, Deposit (Ch 24–25)

Code

readr::write_csv(top, "results/metabolomics_DE.csv")
sessioninfo::session_info() |> capture.output() |>
  writeLines("results/metabolomics_session.txt")
# Deposit raw mzML + ISA-Tab metadata to MetaboLights (Ch 25)

Result. An aligned feature matrix, an FDR-controlled list of features separating knockout from wild-type (recovering the expected fatty-acid-amide accumulation), and a MetaboLights-ready deposit.

28.8 What the Two Studies Shared

Look back at the two pipelines. The science diverged — peptides vs. small molecules, protein inference vs. adduct deconvolution — but the computation converged the moment each produced a feature matrix:

Shared component	Proteomics	Metabolomics
`SummarizedExperiment` container	✅	✅
log + median / VSN normalization	✅	✅
Missing-data reasoning (MNAR-aware)	✅	✅ (gap-fill)
`limma` design → contrast → eBayes	✅ (identical code)	✅ (identical code)
BH-controlled ranked result	✅	✅
Quarto report + FAIR deposition	✅	✅

This is the argument for learning the R for Mass Spectrometry ecosystem rather than a single vendor pipeline: one set of skills, transferable across the whole field. A reader who has followed this book can now walk into either a proteomics or a metabolomics project and know exactly which container to build, which model to fit, and how to make the result reproducible.

28.9 Summary

A real label-free proteomics study (DEP::UbiLength) and a real untargeted metabolomics study (faahKO) were each carried from raw data to an interpreted, deposit-ready result.
The domain-specific work lives at the front (identification/feature detection); from the feature matrix onward, the workflows are essentially the same code.
Every step traces back to a chapter in Parts I–VI — the capstone is those chapters assembled into two complete analyses.

28.10 Exercises

Reproduce and extend. Run Case Study 1 end to end. Add a heatmap of the top-20 differential proteins (Chapter 17) and describe the two sample clusters.
Swap the model. In Case Study 1, replace the two-group contrast with a design that adjusts for replicate as a blocking factor (Chapter 21). Do the significant proteins change?
Full metabolomics annotation. In Case Study 2, take the top 10 differential features and complete the annotation to MSI Level 2 using a spectral library search (Chapter 9). How many can you name?
Add a classifier. Using the proteomics feature matrix, build a leakage-free classifier of condition following Chapters 22–23. Report the cross-validated AUC and compare its top features to the differential-abundance hits.
Make it a parameterized report. Wrap either case study in a parameterized Quarto report (Chapter 25) that takes the FDR threshold as a parameter, and render it at 0.05 and 0.01.

28.11 Session Information

Code

sessionInfo()

# Capstone: Two End-to-End Case Studies > *Every preceding chapter isolated one step. Here the steps become a study — twice, in two different fields, with the same toolchain.* This capstone runs two complete analyses from raw data to biological conclusion: a **label-free proteomics** experiment and an **untargeted metabolomics** experiment. Both use real, published datasets bundled in R packages, so you can reproduce every step without downloading anything. The point of running them side by side is the thesis of this book made concrete: **proteomics and metabolomics are different sciences that share one computational grammar in R.** Watch how the same containers (`SummarizedExperiment`, `QFeatures`), the same normalization and missing-data logic, the same `limma` modelling, and the same reproducibility discipline serve both — you learn the toolchain once and apply it to either field. ### From Pedagogy to Production This chapter uses small, bundled datasets (`UbiLength`, `msdata`) so every code block runs without external downloads. For production-scale versions of the same workflows — running on real PRIDE and MetaboLights data with parameter tuning, multi-file processing, and cluster execution — the companion project `r4ms_book/` provides complete, documented pipelines: | Case Study | Pedagogical Version (This Chapter) | Production Version (Companion) | |-----------|-----------------------------------|-------------------------------| | LFQ proteomics | `DEP::UbiLength` (1,182 proteins) | `r4ms_book/analysis/proteomics_analysis.R` — PXD004886 (4,517 proteins × 22 samples, 11-site benchmark) | | Tissue proteomics | — | `r4ms_book/analysis/tissue_atlas_analysis.R` — PXD010154 (33,812 proteins × 12 organs) | | Clinical DIA | — | `r4ms_book/analysis/clinical_spectronaut_analysis.R` — PXD000547 (paired patient samples) | | Untargeted metabolomics | `msdata::metabolomics()` (4 files) | `r4ms_book/analysis/mtbls38_standards_analysis.R` — MTBLS38 (51 compounds with known identity) | The production scripts are the same logic scaled up — they run the same `limma` models, produce the same volcano plots, and follow the same reproducibility discipline. When you move from learning the toolchain to applying it to your own data, the companion scripts are your starting template. ::: {.callout-warning title="The One Mistake to Avoid"} Assuming your data behaves like the example. The *workflow* transfers; the *parameters* do not. Re-tune feature detection, normalization, and imputation to your own study before trusting a single number. ::: ## Learning Objectives By the end of this chapter you will be able to: - Assemble a full analysis from the individual steps taught in Chapters 4–25 - Run a label-free proteomics study end to end: import → `QFeatures` → filter → normalize → impute → differential abundance → interpretation → report - Run an untargeted metabolomics study end to end: import → `xcms` feature detection → normalize → annotate → differential abundance → interpretation → report - Recognise which parts of the workflow are shared across omics and which are domain-specific - Produce a reproducible, deposit-ready result for each study ## How to Read This Chapter Each case study is a guided tour with pointers back to the chapter where each step is taught in depth. If a step is unfamiliar, follow the reference; here we keep commentary short and let the pipeline flow. | Stage | Proteomics tool | Metabolomics tool | Taught in | |---|---|---|---| | Import | `readQFeatures()` | `MsExperiment`, `Spectra` | Ch 4–5 | | Feature generation | search-engine PSMs | `xcms::findChromPeaks()` | Ch 7, 10 | | Container | `QFeatures` | `SummarizedExperiment` | Ch 5, 12 | | QC | detection rate, CV | TIC/BPC, QC-pool CV | Ch 6 | | Normalize | median / VSN | log + median / TIC | Ch 17 | | Missing data | filter + impute | filter + fill peaks | Ch 18 | | Identity | protein inference | adduct + mass search | Ch 11, 8–9 | | Model | `limma` | `limma` | Ch 20 | | Interpret | enrichment | pathway / networks | Ch 24 | | Report | Quarto + PRIDE | Quarto + MetaboLights | Ch 25 | ::: {.callout-note} All code is shown with `#| eval: false` so the chapter renders without the full analysis stack installed. Each block runs as written against the bundled example data once the packages in the setup chunk are available. ::: --- # Case Study 1 — Label-Free Proteomics ## The Study The `DEP` package bundles the **UbiLength** dataset (Zhang *et al.*): a label-free experiment comparing HeLa cells in which the deubiquitinase **Ubiquitin-specific protease** pathway is perturbed (`Ubi` knockdown) against controls, quantified by MaxQuant LFQ. The biological question: **which proteins change in abundance when the pathway is disrupted?** - **Design:** 2 conditions (control vs. Ubi), replicates per condition - **Input:** a MaxQuant `proteinGroups` table (real output) - **Goal:** a ranked, FDR-controlled list of differentially abundant proteins, interpreted and reported ```{r} #| eval: false library(DEP) library(QFeatures) library(SummarizedExperiment) library(limma) library(tidyverse) data("UbiLength", package = "DEP") # real MaxQuant proteinGroups table dim(UbiLength) # proteins × columns ``` ## Step 1 — Build the Data Object (Ch 4–5, 12) Filter contaminants and reverse hits, resolve duplicate identifiers, and construct a `SummarizedExperiment`/`QFeatures` with a clean experimental design. ```{r} #| eval: false # Remove contaminant + reverse-database hits (standard MaxQuant filtering) data <- UbiLength |> dplyr::filter(Reverse != "+", Potential.contaminant != "+") # Unique protein identifiers data <- DEP::make_unique(data, names = "Gene.names", ids = "Protein.IDs") # LFQ intensity columns and an explicit experimental design lfq_cols <- grep("^LFQ.intensity", colnames(data)) experiment <- DEP::UbiLength_ExpDesign # label / condition / replicate se <- DEP::make_se(data, lfq_cols, experiment) se ``` ## Step 2 — Quality Control (Ch 6) Inspect protein counts per sample, the missing-value structure, and sample clustering before modelling. ```{r} #| eval: false DEP::plot_frequency(se) # proteins identified per sample DEP::plot_numbers(se) # per-sample protein counts DEP::plot_missval(se) # missing-value heatmap — is it structured (MNAR)? ``` ## Step 3 — Normalize and Handle Missing Data (Ch 17–18) Variance-stabilise, then impute. The missingness here is largely **left-censored** (proteins below detection in one condition), so a minimum-based imputation is appropriate — exactly the MNAR reasoning from Chapter 18. ```{r} #| eval: false se_norm <- DEP::normalize_vsn(se) DEP::plot_normalization(se, se_norm) # Filter proteins missing in too many replicates, then impute the rest se_filt <- DEP::filter_missval(se_norm, thr = 0) se_imp <- DEP::impute(se_filt, fun = "MinProb", q = 0.01) # MNAR-appropriate ``` ## Step 4 — Differential Abundance (Ch 20) Fit the `limma` model and extract an FDR-controlled result. `DEP` wraps `limma` internally, so this is the empirical-Bayes moderation of Chapter 20 applied to the protein matrix. ```{r} #| eval: false diff <- DEP::test_diff(se_imp, type = "control", control = "Ctrl") dep <- DEP::add_rejections(diff, alpha = 0.05, lfc = 1) # Ranked results table results <- DEP::get_results(dep) results |> dplyr::filter(significant) |> dplyr::arrange(dplyr::desc(abs(Ubi_vs_Ctrl_ratio))) |> dplyr::select(name, Ubi_vs_Ctrl_ratio, Ubi_vs_Ctrl_p.adj) |> head(15) ``` ## Step 5 — Visualise and Interpret (Ch 20, 23) A volcano plot for the headline result, then pathway enrichment on the significant set to move from a protein list to biology. ```{r} #| eval: false DEP::plot_volcano(dep, contrast = "Ubi_vs_Ctrl", label_size = 3, add_names = TRUE) # Enrichment on the significant proteins (Ch 24) library(clusterProfiler) library(org.Hs.eg.db) sig_genes <- results |> dplyr::filter(significant) |> dplyr::pull(name) ego <- clusterProfiler::enrichGO( gene = sig_genes, OrgDb = org.Hs.eg.db, keyType = "SYMBOL", ont = "BP", pAdjustMethod = "BH" ) clusterProfiler::dotplot(ego, showCategory = 12) ``` ## Step 6 — Report and Deposit (Ch 25) Export the figures and table, capture the environment, and prepare the ProteomeXchange deposit. ```{r} #| eval: false dir.create("results", showWarnings = FALSE) readr::write_csv(results, "results/proteomics_DE.csv") sessioninfo::session_info() |> capture.output() |> writeLines("results/proteomics_session.txt") # Deposit raw + mzIdentML + SDRF to PRIDE / ProteomeXchange (Ch 25) ``` **Result.** A reproducible, FDR-controlled list of proteins responding to the perturbation, with enriched pathways and a deposit-ready bundle — built entirely from the containers and models of Parts I–VI. --- # Case Study 2 — Untargeted Metabolomics ## The Study The `faahKO` package bundles a classic **real** LC–MS dataset: spinal-cord extracts from **fatty-acid amide hydrolase knockout (FAAH⁻/⁻) mice versus wild-type**, in which the knockouts accumulate fatty-acid amides. The question: **which metabolite features distinguish knockout from wild-type?** This is the canonical `xcms` demonstration dataset — and a complete study in its own right. - **Design:** 2 genotypes (KO vs. WT), replicates each - **Input:** real centroided `.CDF`/mzML files - **Goal:** an aligned feature matrix, differential features, and putative annotations ```{r} #| eval: false library(xcms) library(MsExperiment) library(SummarizedExperiment) library(faahKO) cdf_files <- dir(system.file("cdf", package = "faahKO"), recursive = TRUE, full.names = TRUE) pheno <- data.frame( sample_name = sub("\\.CDF$", "", basename(cdf_files)), sample_group = c(rep("KO", 6), rep("WT", 6)) ) mse <- MsExperiment::readMsExperiment(cdf_files, sampleData = pheno) mse ``` ## Step 1 — Feature Detection with `xcms` (Ch 7) The full CentWave → alignment → correspondence → gap-filling pipeline of Chapter 7, condensed. ```{r} #| eval: false # 1. Peak detection (CentWave) cwp <- xcms::CentWaveParam(peakwidth = c(20, 80), ppm = 25, snthresh = 10) xd <- xcms::findChromPeaks(mse, param = cwp) # 2. Retention-time alignment (Obiwarp) xd <- xcms::adjustRtime(xd, param = xcms::ObiwarpParam(binSize = 0.6)) # 3. Correspondence (group peaks into features across samples) pdp <- xcms::PeakDensityParam(sampleGroups = pheno$sample_group, minFraction = 0.5, bw = 30) xd <- xcms::groupChromPeaks(xd, param = pdp) # 4. Fill missing peaks so the matrix has no structural gaps xd <- xcms::fillChromPeaks(xd) ``` ## Step 2 — Extract the Feature Matrix (Ch 5, 12) Turn the `xcms` result into the same `SummarizedExperiment` used on the proteomics side — the moment the two workflows converge. ```{r} #| eval: false se_metab <- xcms::quantify(xd, value = "into") # SummarizedExperiment se_metab # assay(): features × samples; rowData(): mz/rt; colData(): genotype ``` ## Step 3 — QC and Normalization (Ch 6, 17) ```{r} #| eval: false # TIC per sample and log-transform + median normalization (Ch 17) assay(se_metab) <- log2(assay(se_metab) + 1) sample_medians <- matrixStats::colMedians(assay(se_metab), na.rm = TRUE) assay(se_metab) <- sweep(assay(se_metab), 2, sample_medians - mean(sample_medians)) # PCA to confirm genotype separates and no batch structure dominates pc <- prcomp(t(na.omit(assay(se_metab))), scale. = TRUE) ``` ## Step 4 — Differential Features with `limma` (Ch 20) The *identical* modelling code from the proteomics study — the payoff of a shared container. ```{r} #| eval: false library(limma) design <- model.matrix(~ 0 + factor(se_metab$sample_group)) colnames(design) <- c("KO", "WT") fit <- limma::lmFit(assay(se_metab), design) fit <- limma::contrasts.fit(fit, limma::makeContrasts(KO - WT, levels = design)) fit <- limma::eBayes(fit) top <- limma::topTable(fit, number = Inf, adjust.method = "BH") head(top, 15) ``` ## Step 5 — Annotate the Hits (Ch 8–9) Convert the significant features' *m/z* to putative identities via exact-mass search — the annotation funnel of Chapter 8. ```{r} #| eval: false library(MetaboCoreUtils) sig <- top |> tibble::rownames_to_column("feature") |> dplyr::filter(adj.P.Val < 0.05) sig_mz <- SummarizedExperiment::rowData(se_metab)[sig$feature, "mzmed"] # Expected m/z for common adducts of candidate fatty-acid amides adduct_mz <- MetaboCoreUtils::mass2mz( MetaboCoreUtils::calculateMass("C18H35NO"), # e.g. oleamide adduct = c("[M+H]+", "[M+Na]+") ) # Match sig_mz against a reference library within tolerance (Ch 8–9) ``` ## Step 6 — Interpret, Report, Deposit (Ch 24–25) ```{r} #| eval: false readr::write_csv(top, "results/metabolomics_DE.csv") sessioninfo::session_info() |> capture.output() |> writeLines("results/metabolomics_session.txt") # Deposit raw mzML + ISA-Tab metadata to MetaboLights (Ch 25) ``` **Result.** An aligned feature matrix, an FDR-controlled list of features separating knockout from wild-type (recovering the expected fatty-acid-amide accumulation), and a MetaboLights-ready deposit. --- ## What the Two Studies Shared Look back at the two pipelines. The **science** diverged — peptides vs. small molecules, protein inference vs. adduct deconvolution — but the **computation converged** the moment each produced a feature matrix: | Shared component | Proteomics | Metabolomics | |---|---|---| | `SummarizedExperiment` container | ✅ | ✅ | | log + median / VSN normalization | ✅ | ✅ | | Missing-data reasoning (MNAR-aware) | ✅ | ✅ (gap-fill) | | `limma` design → contrast → eBayes | ✅ (identical code) | ✅ (identical code) | | BH-controlled ranked result | ✅ | ✅ | | Quarto report + FAIR deposition | ✅ | ✅ | This is the argument for learning the R for Mass Spectrometry ecosystem rather than a single vendor pipeline: **one set of skills, transferable across the whole field.** A reader who has followed this book can now walk into either a proteomics or a metabolomics project and know exactly which container to build, which model to fit, and how to make the result reproducible. ## Summary - A real label-free proteomics study (`DEP::UbiLength`) and a real untargeted metabolomics study (`faahKO`) were each carried from raw data to an interpreted, deposit-ready result. - The domain-specific work lives at the front (identification/feature detection); from the feature matrix onward, the workflows are essentially the same code. - Every step traces back to a chapter in Parts I–VI — the capstone is those chapters assembled into two complete analyses. ## Exercises 1. **Reproduce and extend.** Run Case Study 1 end to end. Add a heatmap of the top-20 differential proteins (Chapter 17) and describe the two sample clusters. 2. **Swap the model.** In Case Study 1, replace the two-group contrast with a design that adjusts for replicate as a blocking factor (Chapter 21). Do the significant proteins change? 3. **Full metabolomics annotation.** In Case Study 2, take the top 10 differential features and complete the annotation to MSI Level 2 using a spectral library search (Chapter 9). How many can you name? 4. **Add a classifier.** Using the proteomics feature matrix, build a leakage-free classifier of condition following Chapters 22–23. Report the cross-validated AUC and compare its top features to the differential-abundance hits. 5. **Make it a parameterized report.** Wrap either case study in a parameterized Quarto report (Chapter 25) that takes the FDR threshold as a parameter, and render it at 0.05 and 0.01. ## Session Information ```{r} #| eval: false sessionInfo() ```

26.0.1 From Pedagogy to Production

26.1 Learning Objectives

26.2 How to Read This Chapter

27 Case Study 1 — Label-Free Proteomics

27.1 The Study

27.2 Step 1 — Build the Data Object (Ch 4–5, 12)

27.3 Step 2 — Quality Control (Ch 6)

27.4 Step 3 — Normalize and Handle Missing Data (Ch 17–18)

27.5 Step 4 — Differential Abundance (Ch 20)

27.6 Step 5 — Visualise and Interpret (Ch 20, 23)

27.7 Step 6 — Report and Deposit (Ch 25)

28 Case Study 2 — Untargeted Metabolomics

28.1 The Study

28.2 Step 1 — Feature Detection with xcms (Ch 7)

28.3 Step 2 — Extract the Feature Matrix (Ch 5, 12)

28.4 Step 3 — QC and Normalization (Ch 6, 17)

28.5 Step 4 — Differential Features with limma (Ch 20)

28.6 Step 5 — Annotate the Hits (Ch 8–9)

28.7 Step 6 — Interpret, Report, Deposit (Ch 24–25)

28.8 What the Two Studies Shared

28.9 Summary

28.10 Exercises

28.11 Session Information

28.2 Step 1 — Feature Detection with `xcms` (Ch 7)

28.5 Step 4 — Differential Features with `limma` (Ch 20)