Metabolomics MS: Platform Choice and QC Design

An untargeted metabolomics experiment is decided before a single spectrum is acquired. Two choices set what the rest of the study can support: which mass spectrometry platform generates the data — liquid chromatography (LC-MS) or gas chromatography (GC-MS), targeted or untargeted acquisition — and whether the injection sequence includes a pooled quality-control (QC) sample often enough to characterise and correct instrument drift. Neither choice is reversible after the run finishes. This guide covers both in the order they are actually made: platform and acquisition-mode selection first, then the pooled-QC design that makes batch correction possible afterward.

Targeted or untargeted: the question that decides everything else

Before choosing an instrument, decide what the experiment is asking.

Targeted metabolomics quantifies a pre-specified panel of known metabolites against authentic reference standards or calibration curves. It is built for hypothesis testing — confirming a specific pathway readout, validating a candidate biomarker, or reporting an absolute concentration — and commonly runs as multiple reaction monitoring (MRM) on a triple quadrupole. The transition-optimisation logic is the same one covered in our MRM transitions guide, applied to metabolite rather than peptide transitions.
Untargeted metabolomics is a global, unbiased survey of every detectable feature in a sample. It is hypothesis-generating rather than hypothesis-testing, and it depends on a high-resolution, accurate-mass (HRAM) analyser — an Orbitrap or a high-resolution time-of-flight — because confident feature annotation requires resolving and mass-measuring compounds that were never specified in advance. See how much resolving power and mass accuracy you actually need for the instrument side of that requirement.

The practical rule mirrors the discovery-versus-verification logic our proteomics mass spectrometry guide covers on the protein side: run untargeted acquisition to generate a candidate list of metabolites that differ between groups, then confirm the ones that matter with a targeted, standard-calibrated assay in more samples. Running a full untargeted acquisition to answer a four-metabolite question is an expensive way to get a less quantitatively rigorous answer than a targeted panel would have given directly.

Platform choice: LC-MS or GC-MS

Once the acquisition strategy is set, platform choice mostly comes down to what class of metabolite the study needs to see and how much identification confidence the downstream question requires.

LC-MS

Liquid chromatography coupled to mass spectrometry is the default platform for most untargeted metabolomics work today because it requires no chemical derivatization and, depending on the chromatographic mode, covers a broad polarity range: reversed-phase LC for nonpolar and mid-polarity metabolites (lipids, many drug and xenobiotic metabolites), hydrophilic interaction liquid chromatography (HILIC) for polar and ionic metabolites (amino acids, nucleotides, sugars) that reversed-phase columns retain poorly. Running both modes in the same study is common precisely because neither alone covers the full metabolome.

GC-MS

Gas chromatography-mass spectrometry requires derivatising most polar, non-volatile metabolites (commonly by silylation) before they are volatile and thermally stable enough to run on a GC column — an added sample-prep step LC-MS does not need. In exchange, GC-MS delivers two advantages that matter for identification confidence specifically: highly reproducible retention indices, and electron-impact (EI, typically 70 eV) fragmentation that is standardised enough to match directly against large, cross-instrument reference libraries such as NIST. That library-matching confidence is harder to get from LC-MS, where fragmentation is more method- and instrument-dependent. See GC-MS vs LC-MS for the full side-by-side.

Running both

Comprehensive metabolomics programs frequently run LC-MS (both chromatographic modes) and GC-MS in parallel on the same sample set rather than picking one. Coverage is not fully overlapping: GC-MS is strong on primary metabolism — sugars, amino acids, organic and fatty acids — with high identification confidence via library matching, while LC-MS reaches a much broader chemical space with less universal identification certainty. The decision to run one platform or both should be made against the specific metabolite classes the study needs to see, not by default.

Ionization and acquisition-mode considerations

LC-MS metabolomics typically uses electrospray ionization (ESI) in positive and/or negative mode; polarity switching within a single run extends metabolite coverage but adds duty-cycle cost, so the choice should be deliberate rather than left at an instrument default. GC-MS uses electron ionization, a harder ionization technique that produces the reproducible, information-rich fragmentation library matching depends on, at the cost of the molecular-ion information a softer ionization source would preserve. Whichever platform is chosen, the ionization settings should be fixed and documented before the injection sequence begins — changing source parameters mid-study reintroduces exactly the kind of systematic variation the QC design below exists to control for.

The pooled QC sample: what it is and why the study depends on it

A pooled QC sample is built by combining a small aliquot from every study sample (or a representative cross-section of them) into a single pool, then splitting that pool into many identical injections used throughout the acquisition sequence. It does three distinct jobs, and a study design that only does one of them is incomplete:

System conditioning. Several pooled-QC injections run first, before any study sample, to equilibrate the column and ion source and let signal intensity stabilise. Study samples acquired during that equilibration period are not reliable data.
Drift and batch monitoring. Because every QC injection is chemically identical, any systematic change in a feature’s measured intensity across the QC injections reflects instrument or method drift, not biology. This is the only direct evidence available for separating a real between-group difference in the study samples from a run-order artifact that happens to correlate with it.
Reproducibility filtering. A feature that is not consistently and reproducibly detected across the QC injections is not a reliable measurement, and is normally excluded from biological interpretation regardless of what its values look like in the study samples themselves.

Designing the QC injection sequence

Conditioning phase. Run several pooled-QC injections at the start of the sequence — enough that intensity has visibly stabilised — before the first study sample is acquired.
Regular interval. Inject pooled QC at a fixed interval throughout the batch, commonly framed in the field as roughly every five to ten study-sample injections. Frequent enough to characterise drift with enough resolution to correct it later; infrequent enough that QC injections do not consume the instrument time the study needs.
Randomised sample order. Acquire study samples in randomised order across groups, not block by condition. If drift does occur, randomisation spreads it across every group instead of confounding it with the biological comparison — a run-order confound (e.g. all controls acquired early, all cases acquired late) is one of the hardest failure modes to detect after the fact, because it can manufacture or mask a real group difference that looks statistically clean.
Process and solvent blanks. Include blanks alongside the QC injections to separate background and carryover signal from real sample signal — a QC-only design without blanks cannot distinguish the two.
Study-wide pooling for multi-batch studies. Where a study spans more than one acquisition batch, pool the QC across the whole study rather than per batch wherever feasible, so the same reference material links every batch together and supports cross-batch correction rather than just within-batch drift correction.

Using the QC data: reproducibility filtering and batch correction

Two distinct uses of the pooled-QC data happen after acquisition, and both depend on having enough QC injections spread across the run to be meaningful:

Reproducibility filtering

Compute the relative standard deviation (RSD, also reported as %CV) of each feature’s intensity across the pooled-QC injections. Features with high QC RSD are unreliable measurements and are commonly dropped before any biological analysis proceeds. Published practice in the field commonly uses RSD thresholds in roughly the 20–30% range for LC-MS features, tightened further for GC-MS given its typically better retention-time and peak-area reproducibility — but the exact cutoff varies by instrument, platform and study design, and should be set and reported explicitly for the specific dataset rather than borrowed silently from another paper.

Batch and drift correction

Once reproducibility has been characterised, the QC intensities measured across the run are commonly used to fit and remove the systematic drift trend from the study-sample data — a smoothed or regression-based curve fit through the QC values, applied as a per-feature correction to the surrounding study samples. This is the practical mechanism behind what the field broadly calls QC-based signal correction, and it is implemented in the open-source tools most untargeted metabolomics labs already use for feature detection and alignment, including pipelines built around XCMS, MZmine and MetaboAnalyst.

Reporting what was actually done

The Metabolomics Standards Initiative’s identification-confidence framework — ranging from a compound confirmed against an authentic reference standard down to an unidentified feature reported by mass and retention time alone — is the community-standard way to report how confidently each metabolite in the results was actually identified, and it should be stated explicitly rather than implied by how the results table is formatted. For the QC data and run metadata specifically, the Proteomics Standards Initiative’s mzQC format (final release v1.0.0, February 2024, maintained by the PSI Quality Control working group) exists as a structured, archivable format for exactly this kind of quality-control reporting and handover, despite its proteomics-oriented name — it is built to serve as a handover, report, and archive format for mass-spectrometry QC data generally.

Common design mistakes

QC only at the start and end of the run. Two QC points cannot characterise a drift curve across the middle of a long batch — there is nothing to fit a correction to.
Confounding batch with condition. Acquiring all controls in one batch and all cases in another makes any batch effect statistically indistinguishable from the biological effect the study is trying to measure.
Skipping the conditioning injections. Treating the first few study samples as real data when the system has not yet equilibrated introduces exactly the kind of artifact the QC design exists to catch.
Pooling from an incomplete sample set. A QC pool built from only part of the study (missing an arm or a timepoint) does not represent the full chemical space the study actually measures, which biases which features the QC can even evaluate.
Choosing the platform after the data exist. Deciding, post hoc, to treat an untargeted dataset as if it were a targeted, quantitatively validated one (or the reverse) does not add the coverage or the calibration the other platform would have provided from the start.

Frequently asked questions

What is a pooled QC sample in metabolomics?

A single reference sample created by combining a small aliquot from every (or a representative subset of) study sample, then injected repeatedly through an LC-MS or GC-MS acquisition sequence. Because every QC injection is chemically identical, it is the reference point used to detect instrument drift, assess feature reproducibility, and correct batch effects in the study-sample data.

How often should a pooled QC sample be injected during a run?

After an initial conditioning block of several QC injections at the start of the sequence, the field commonly injects QC at a regular interval through the batch — roughly every five to ten study-sample injections is a common framing — frequently enough to characterise drift with enough resolution to correct it, without consuming so much instrument time that it crowds out study samples.

Should I use LC-MS or GC-MS for untargeted metabolomics?

LC-MS is the more common default because it needs no derivatization step and, across reversed-phase and HILIC modes, covers a broad polarity range. GC-MS requires derivatization but offers more reproducible retention behaviour and stronger library-matching identification confidence via standardised electron-impact fragmentation, particularly for primary metabolism (sugars, amino acids, organic acids). Many comprehensive programs run both because their metabolite coverage does not fully overlap.

Can targeted and untargeted metabolomics use the same platform?

The two acquisition strategies can run on the same underlying chromatography, but the instrument and method optimised for one is rarely optimal for the other: untargeted work needs a high-resolution, accurate-mass analyser to identify unspecified compounds, while targeted quantitation is commonly run as MRM on a triple quadrupole optimised for sensitivity and dynamic range on a pre-declared list of transitions.

What RSD is acceptable for a metabolite feature measured in QC samples?

There is no single universal number; the field commonly works in roughly the 20–30% RSD range for LC-MS features (tighter for GC-MS), but the threshold should be set for the specific instrument, platform and study, and reported explicitly alongside the results rather than assumed from another paper’s cutoff.

Do GC-MS metabolomics samples need derivatization?

Most polar, non-volatile metabolites do — commonly by silylation — to make them volatile and thermally stable enough to run on a GC column without decomposing. This is the main added sample-prep step GC-MS carries that LC-MS does not.