Injection Order: The Quietest Trap in LC-MS Metabolomics

In mass-spectrometry metabolomics, the thing that quietly changes your conclusions is often not the statistical test. It is the order in which you injected the samples.

Everything below is drawn from published literature. No client data is used.

A story about bathroom scales

A scale that has just been switched on takes a while to settle. Now imagine measuring a class of students like this:

Weigh everyone in Group A first, then everyone in Group B.

Group B comes out 2 kg heavier on average. Is Group B actually heavier, or had the scale simply settled by then?

You cannot tell. "Group" and "measurement order" have become the same variable.

The same thing happens in LC-MS metabolomics and proteomics more often than people realise, and it is almost invisible if you only look at the final data table.

1. Instruments drift

An LC-MS run lasts hours. Over that time:

• the column surface gradually gets coated with sample matrix, so separation behaviour shifts
• the ion source accumulates contamination, so sensitivity changes
• solvent composition and temperature move slightly

So the first sample injected and the last sample injected were not measured under the same conditions. This is signal drift.

The beginning of a batch is the least stable part. A 2023 review notes that the data from the first four to eight injections are usually not yet stable [3], which is why it recommends starting the sequence with five to ten QC injections used purely for equilibration [3]. Broadhurst et al. leave the exact number to each lab to determine for its own platform; their worked example uses eight [1].

A common misconception: a blank is not a conditioning injection. If anything it does the opposite. Broadhurst et al. warn that a single process blank disturbs the platform's equilibrium and de-conditions the column, so no study sample should directly follow a blank. Typically four or five pooled QC injections are needed after a blank to re-condition the system [1].

2. The real problem is not drift, it is confounding

Drift is always present to some degree. It becomes a problem when drift runs parallel to your experimental groups.

If you inject group by group (all controls first, all treated samples afterwards), you get this:

	Position in run	Experimental group
Early	instrument less stable	control
Late	instrument stable	treated

Now, when the treated group looks different, you cannot tell whether that is the treatment or the position in the run.

A tutorial on batch effects in large-scale proteomics puts it bluntly [2]. Their example is sample-preparation batches, but injection position is the same kind of technical factor:

"It is also important to stress that batch factors should be already considered in the experimental design phase, to ensure that the data are not biased beyond repair, something that can happen when biological groups are completely confounded with sample preparation batches." (emphasis added)

The same paper pins down the exact condition behind the familiar worry that "correction flattens the biological signal", and names it for what it is [2]:

"Another common advice is to not correct batch effects to avoid 'flattening' the biological signal. This can occur when the biological groups are confounded with the batch factors. However, confounding is a serious experimental design error and can be prevented by balancing the biological groups between the expected batches." (emphasis added)

Once groups and run order coincide, you lose either way:

• Correct for drift: the variance you remove is the group contrast. Real differences get flattened.
• Do not correct: group differences stay mixed with drift.
• Let the data decide: it cannot. The contrast needed to separate the two was never built into the design.

So this is not a problem you solve at the analysis stage. No batch-correction algorithm, however good, can recover a contrast that was never measured independently.

3. How common is this?

A scoping review took untargeted metabolomics papers published in the first half of 2021 and had each of the 109 papers reviewed by at least five reviewers [4]. The findings are refreshingly candid:

• 17 of the 109 papers reported no QC samples or other QC approach
• Among the 72 papers that did report pooled QC samples, the stated purpose varied: repeatability/reproducibility 60%, LC conditioning 24%, filtering low-quality features 22%, batch or drift correction just 10%, supporting metabolite identification 3%
• 31 papers displayed QC samples on a PCA plot, but only one used a numerical metric to assess technical variation. The rest judged by eye

The review is direct about the limits of judging by eye [4]:

"PCA models can show trends in data, but even relatively large analytical variation can seem insignificant if the biological variation is appreciably larger than the technical variation."

Read that in reverse: the smaller your biological effect, the more easily technical drift drives your result.

And on the absence of an agreed threshold [4]:

"There is no widely accepted metric for delineating acceptable from unacceptable data quality; it is up to individual investigators to determine if their data are reproducible or if corrections are required."

4. Measure it, don't eyeball it

Two numbers quantify instability. Both require pooled QC samples.

① QC RSD: relative standard deviation across repeated injections of the same pooled QC.

Target: below 20% for LC-MS and below 30% for GC-MS [3]. Broadhurst et al. give the same 20% or 30% thresholds, chosen depending on the sample type [1].

② D-ratio: per metabolite, the spread among QC samples divided by the spread among study samples.

Target: below 50% [1, 3], and preferably much lower [1]. 0% is perfect; 100% means what you measured is essentially noise [3].

A D-ratio near 100% means re-measuring the same sample moves the value as much as switching to a different subject does. Comparing groups on such a feature is not meaningful.

Both numbers belong in the report. Because there is no universally agreed pass mark [4], the honest approach is to show the numbers and let the reader judge, rather than quietly deciding for them.

5. Five things to fix at the design stage

Most of these cost nothing but a different injection sequence.

1. Randomise sample order across groups. Do not run one group as a block. This single change breaks the link between group and run position, which is exactly the problem described here.
2. Inject pooled QC samples generously. One every 3 to 10 samples [3]; at least five per batch, and at least eight if you intend to apply a nonlinear drift correction [1].
3. Put two QC injections at the start and two at the end of each batch [1]. It is insurance against a failed injection, and the QCs at either end weigh heavily on the shape of the correction curve [1].
4. Add conditioning injections before the batch begins: five to ten [3], or a number you have determined for your own platform [1].
5. Never inject a study sample directly after a blank. Re-condition with four or five QC injections first [1].

Pooled QC samples do not consume extra study material. They are aliquots of a pool made from the samples you already have. The only real cost is instrument time.

If the study is large enough to need multiple batches, make sure every group appears in every batch (a blocked design). Märtens et al. note that in a blocked design only samples within one batch can be compared, which reduces statistical power, and that batch effects therefore need to be removed before samples can be compared across batches [3].

6. What if the run is already finished?

• Drift present, but not aligned with groups: correction helps. With eight or more QC samples you can apply a QC-based nonlinear correction [1] such as QC-RSC or LOESS-based signal correction [3].
• Drift aligned with groups: correction cannot recover the contrast. What you can do: ⓐ compute QC RSD and D-ratio and report them alongside the results, ⓑ separate the features that pass those criteria from those that do not, and ⓒ state plainly that the finding is not confirmatory.
• Cross-checking is a practical fallback. A metabolite that changes in the same direction in both ionisation modes, or a finding that survives a re-run in randomised order, is far less likely to be an artefact of drift.

One more caution: do not use "we found more differential features after correction" as evidence that the correction worked. The batch-effect tutorial explicitly advises against judging correction by improved clustering by biological condition or a higher number of differentially expressed proteins [2].

Closing

A large part of metabolomics data quality is decided not during analysis but at the moment someone writes the injection sequence.

This is not about having a better instrument. Drift occurs on excellent instruments too, and if the run order tracks your experimental groups, the ambiguity is identical.

So when designing the experiment, treat "in what order will these samples be injected?" as part of the statistical design. It is among the cheapest decisions available, and one of the few that changes how much your result is actually worth.

References

1. Broadhurst D, Goodacre R, Reinke SN, Kuligowski J, Wilson ID, Lewis MR, Dunn WB. Guidelines and considerations for the use of system suitability and quality control samples in mass spectrometry assays applied in untargeted clinical metabolomic studies. Metabolomics 2018;14:72. doi:10.1007/s11306-018-1367-3
2. Čuklina J, Lee CH, Williams EG, Sajic T, Collins BC, Rodríguez Martínez M, et al. Diagnostics and correction of batch effects in large-scale proteomic studies: a tutorial. Molecular Systems Biology 2021;17:e10240. doi:10.15252/msb.202110240
3. Märtens A, Holle J, Mollenhauer B, Wegner A, Kirwan J, Hiller K. Instrumental Drift in Untargeted Metabolomics: Optimizing Data Quality with Intrastudy QC Samples. Metabolites 2023;13:665. doi:10.3390/metabo13050665
4. Broeckling CD, Beger RD, Cheng LL, Cumeras R, Cuthbertson DJ, Dasari S, et al. Current Practices in LC-MS Untargeted Metabolomics: A Scoping Review on the Use of Pooled Quality Control Samples. Analytical Chemistry 2023;95:18645–18654. doi:10.1021/acs.analchem.3c02924

All quotations were checked word for word against the full text of the cited papers; bold inside quotations is added for emphasis.