Bio Metabolomics Xcms Preprocessing

Version Compatibility

Reference examples tested with: xcms 4.x+ (MsExperiment/XcmsExperiment containers), Spectra 1.12+, CAMERA 1.58+

Before using code patterns, verify installed versions match. If versions differ:

• R: packageVersion('xcms') then ?CentWaveParam to verify parameter names and defaults

If code throws ImportError, AttributeError, or TypeError, introspect the installed package and adapt the example to match the actual API rather than retrying.

A feature table is only meaningful alongside its full processing specification: which xcms version, every *Param value, and the fill/filter ordering. The table is a parameterized hypothesis about which molecules exist, not the data.

XCMS Untargeted LC-MS Preprocessing

"Turn my raw LC-MS files into a feature table" -> Detect chromatographic peaks per file, align retention times across runs, group corresponding peaks into features, fill gaps, then collapse adduct/isotope redundancy.

• R: readMsExperiment() -> findChromPeaks() -> adjustRtime() -> groupChromPeaks() -> fillChromPeaks() (xcms)

The Single Most Important Insight -- The Feature Table Is a Model-Dependent Artifact, Not Ground Truth

Every cell in the table is the output of a detection + grouping + filling model with chosen parameters. Two analysts with different centWave/grouping settings produce materially different tables from identical raw files, so "not detected" is a statement about the parameters, not the sample. Three consequences reorganize the whole workflow: (1) preprocessing parameters silently set the detection floor - a compound absent from results may be present in the raw data but excluded by noise/prefilter/peakwidth/snthresh; (2) fillChromPeaks integrates whatever signal sits in a feature window even when no peak exists, fabricating a positive number where the honest answer is "below detection"; (3) one compound yields 5-15 features (adducts, isotopologues, in-source fragments, multimers), so a 10,000-feature table is plausibly ~1,000 compounds (Mahieu 2017). Report parameters as part of the result, inspect EICs and alignment of every hit, and collapse redundancy before annotation.

API Generations -- Use Modern, Not Legacy

Parameters are objects, not loose args: findChromPeaks(data, param = CentWaveParam(...)), never findChromPeaks(data, ppm=..., peakwidth=...).

Decision Tree by Scenario

Peak Detection

Goal: Detect chromatographic peaks in each centroided file.

Approach: Build a CentWaveParam with ppm and peakwidth set from the actual instrument and chromatography (see Quantitative Thresholds), then call findChromPeaks.

library(xcms)
# spectraFiles: centroided mzML paths; pd: data.frame with one row per file
raw <- readMsExperiment(spectraFiles = mzml_files, sampleData = pd)

# ppm is across-scan centroid scatter (~2-3x measured error), NOT the spec mass accuracy.
# peakwidth is c(min, max) in SECONDS, measured from EIC base-widths of known peaks.
cwp <- CentWaveParam(ppm = 10, peakwidth = c(2, 20), snthresh = 10,
                     prefilter = c(3, 1000), noise = 1000, mzdiff = -0.001,
                     integrate = 1L, mzCenterFun = 'wMean')
xdata <- findChromPeaks(raw, param = cwp)
nrow(chromPeaks(xdata))

Retention-Time Alignment

Goal: Remove cross-run RT drift so the same compound lands at the same RT in every sample.

Approach: Choose obiwarp (no prior peaks) or peakGroups (anchor-based); align to a pooled QC, never to file #1. Regroup afterward because RTs changed.

# obiwarp: full-profile warping. binSize here is the m/z profile bin (default 1),
# distinct from PeakDensityParam$binSize and MatchedFilterParam$binSize.
xdata <- adjustRtime(xdata, param = ObiwarpParam(binSize = 0.6))

# peakGroups alternative needs an initial correspondence and good universal anchors:
# xdata <- groupChromPeaks(xdata, param = pdp_anchor)
# xdata <- adjustRtime(xdata, param = PeakGroupsParam(minFraction = 0.85, span = 0.4,
#     subset = which(sampleData(xdata)$sample_type == 'QC'), subsetAdjust = 'average'))
plotAdjustedRtime(xdata)

Correspondence (Grouping)

Goal: Match peaks across samples into consensus features.

Approach: Peak-density grouping in m/z slices; bw is the dominant knob and must reflect residual post-alignment RT scatter, not raw peak width.

pdp <- PeakDensityParam(sampleGroups = sampleData(xdata)$sample_group,
                        bw = 5, minFraction = 0.5, minSamples = 1, binSize = 0.025)
xdata <- groupChromPeaks(xdata, param = pdp)
nrow(featureDefinitions(xdata))

Gap-Filling

Goal: Integrate signal for features missing a detected peak in some samples.

Approach: fillChromPeaks with ChromPeakAreaParam; treat filled values as imputations, not measurements.

xdata <- fillChromPeaks(xdata, param = ChromPeakAreaParam())
filled <- chromPeakData(xdata)$is_filled   # logical flag; lives in chromPeakData, not chromPeaks
feat <- featureValues(xdata, value = 'into')        # features x samples matrix
defs <- featureDefinitions(xdata)                   # mzmed / rtmed / npeaks per feature

Redundancy Collapse

Goal: Group the same compound's adducts/isotopes/fragments back toward compound spectra before annotation.

Approach: CAMERA in order groupFWHM -> groupCorr -> findIsotopes -> findAdducts (isotopes before adducts). Correlation grouping needs enough samples to be meaningful and can over- or under-merge - verify against the table size.

library(CAMERA)
xsa <- xsAnnotate(as(xdata, 'xcmsSet'))
xsa <- groupFWHM(xsa, perfwhm = 0.6)
xsa <- groupCorr(xsa)
xsa <- findIsotopes(xsa, mzabs = 0.01, ppm = 10)
xsa <- findAdducts(xsa, polarity = 'positive')
peaklist <- getPeaklist(xsa)

QC Feature Filtering (Preprocessing/QC Bridge)

Goal: Drop features that fail conventional QC, operationalizing Broadhurst 2018 inside the xcms object.

Approach: filterFeatures with RsdFilter (CV in QCs), DratioFilter (sd_QC/sd_sample), PercentMissingFilter, BlankFlag. Drift correction and the full QC pipeline live in metabolomics/normalization-qc.

qc <- sampleData(xdata)$sample_group == 'QC'
study <- sampleData(xdata)$sample_group %in% c('Control', 'Treatment')
xdata <- filterFeatures(xdata, filter = RsdFilter(threshold = 0.3, qcIndex = qc))
xdata <- filterFeatures(xdata, filter = DratioFilter(threshold = 0.5, qcIndex = qc, studyIndex = study))

Per-Method Failure Modes

ppm set to the spec sheet

• Trigger: Setting ppm = 3 because the Orbitrap datasheet says 3 ppm.
• Mechanism: centWave ppm is across-scan centroid scatter, which exceeds time-averaged mass accuracy; too tight fragments one ion into short ROIs that each fail prefilter.
• Symptom: Features vanish entirely (not degrade); the better the instrument spec, the worse it looks.
• Fix: Set ppm to ~2-3x the empirical per-scan centroid scatter, not the datasheet number.

peakwidth mismatch

• Trigger: Copying the default c(20, 50) onto modern UHPLC.
• Mechanism: Lower bound too high discards sharp 2-5 s peaks; upper bound too low clips broad HILIC/tailing peaks. No warning is emitted.
• Symptom: Real peaks silently absent from the table.
• Fix: Measure base-width FWHM from 5-10 known EICs; set peakwidth ~ c(0.5x min, 2x max).

prefilter/noise/snthresh as the trace guillotine

• Trigger: Tuning snthresh while prefilter[I] already kills the trace.
• Mechanism: These are three serial gates on the same low-intensity signal; the lowest wins. On high-baseline instruments the default I=100 may both under-filter noise and kill trace metabolites.
• Symptom: Trace metabolites never appear regardless of snthresh.
• Fix: Lower prefilter[I] first for trace work; the lowest gate dominates.

bw too coarse / alignment-coupled

• Trigger: Copying bw = 30 onto UHPLC, or choosing bw independently of alignment quality.
• Mechanism: On UHPLC, bw=30 merges chromatographically resolved co-eluting compounds; with poor alignment a tight bw instead splits one compound across features.
• Symptom: Averaged-away differences (over-merge) or duplicate split features (under-merge).
• Fix: Set bw from residual post-alignment RT scatter (often 2-6 s on UHPLC); inspect EICs of merged/split features.

gap-filling fabricates intensities

• Trigger: Feeding a naively filled table straight into a t-test.
• Mechanism: Missingness is MNAR (below LOD); filling integrates the noise floor into a positive number, inflating the absent group's mean and shrinking the fold-change being tested.
• Symptom: "Significant" features that are mostly filled in one group.
• Fix: Track is_filled; report per-feature filled fraction; for inference use unfilled values with MNAR-aware imputation (QRILC/GSimp), reserving the fill for dense exploratory PCA.