A custom AI reviewer for metabolomics and flux reporting

PerfectPaper’s metabolomics agent reads a metabolomics or isotope tracing manuscript for the reporting items that specialist reviewers ask about first: how samples were normalised, which internal standards were used, whether isotope data were corrected for natural abundance, and what confidence level supports each metabolite identification. Copy the brief below into a custom agent slot.

Metabolomics has reporting conventions that a general reviewer does not know and a metabolomics reviewer will not let pass. Most of the gap is in methods rather than in the science.

When to use this agent

PerfectPaper’s metabolomics agent is worth adding to a review whenever a manuscript’s claims rest on a mass spectrometry identification, an isotope tracing experiment, or a normalisation choice that has not been stated or defended — the seven signals below cover the cases seen most often.

• Your manuscript reports untargeted or targeted metabolite profiling
• You performed stable isotope tracing, flux analysis, or a Seahorse experiment
• Metabolite identities are asserted from mass spectrometry
• Your normalisation strategy is described in a single sentence
• You are submitting to a journal with metabolomics specialist reviewers
• Your figures name metabolites in full but the methods never state a confidence level
• A collaborator produced the mass spectrometry and you are writing the methods second-hand

The four levels of metabolite identification confidence

Metabolite identification confidence has a published four-level scale, and a specialist reviewer reads your results section against it whether or not you cite it. The Chemical Analysis Working Group of the Metabolomics Standards Initiative set it out in Sumner and colleagues, “Proposed minimum reporting standards for chemical analysis” (Metabolomics 2007;3:211–221), which defines four tiers: identified compounds, putatively annotated compounds, putatively characterised compound classes, and unknown compounds.

The clause that decides most manuscripts is the Level 1 requirement. The MSI states that “a minimum of two independent and orthogonal data relative to an authentic compound analyzed under identical experimental conditions are proposed as necessary” — retention time plus mass spectrum, or NMR plus MS — and adds that “literature values reported for authentic samples by other laboratories are generally believed insufficient”.

In practice, few untargeted studies reach Level 1 for more than a handful of the hundreds of features they report, because running an authentic standard for every annotated metabolite under identical conditions does not scale to a panel of that size. That gap is not a licence to call every annotation an identification — it is the reason the wording change to a putative annotation, covered under “When the obvious remedy is unavailable” below, is the response reviewers accept rather than a concession of failure.

Schymanski and colleagues later published a five-level companion scale for high-resolution mass spectrometry in Environmental Science & Technology (2014;48:2097–2098), running from Level 5 exact mass, through Level 4 unequivocal molecular formula and Level 3 tentative candidate, to Level 2 probable structure by library spectrum match and Level 1 confirmed structure by reference standard. Either scale is acceptable to reviewers. Using neither, and naming metabolites as though every assignment were Level 1, is the objection this agent exists to prevent. How to write a custom reviewer agent covers why the brief reports the level rather than instructing you to obtain standards.

Why accurate mass alone is not an identification

Accurate mass alone does not determine a molecular formula, and the arithmetic has been published. Kind and Fiehn generated more than 1.6 million molecular formulae in the range 0–500 Da and reported in “Metabolomic database annotations via query of elemental compositions: mass accuracy is insufficient even at less than 1 ppm” (BMC Bioinformatics 2006;7:234) that high mass accuracy below 1 ppm still leaves too many chemically possible formulae in the higher mass regions to exclude candidates.

Their measured remedy is the number worth quoting in your own methods: using isotopic abundance patterns as a single further constraint removes more than 95% of false candidates, and a mass spectrometer delivering 3 ppm mass accuracy with 2% error on isotope abundance outperforms a hypothetical instrument at 0.1 ppm that ignores isotope information.

Two consequences follow for a manuscript. First, a reported mass accuracy in ppm is not by itself evidence of identity, so a methods sentence that gives only the instrument’s ppm specification has not answered the identification question. Second, a formula is not a structure: isomers share a formula exactly, and glucose, fructose and galactose are indistinguishable by mass at any resolution. The agent reports metabolites named with a definite identity where only accurate mass supports them, and states the wording change — “putative annotation” rather than “identified” — that removes the objection without new experiments.

Common questions about metabolite identification confidence

Can an AI reviewer tell whether my metabolites are identified or only annotated?

Yes, to the extent the manuscript states its evidence. An identification requires an authentic standard run under the same conditions, or a spectral match to a library; an annotation rests on accurate mass or a predicted formula alone. The agent reads what supports each named metabolite and flags any asserted as identified — MSI Level 1 — when only accurate mass was measured.

Is matching retention time and accurate mass to a standard enough to call a metabolite identified?

No, not under the Metabolomics Standards Initiative’s Level 1 definition. Retention time plus accurate mass is one orthogonal property pair, but the standard must be run in your own laboratory under identical conditions — a published retention time from another lab does not count, because chromatography shifts between instruments and columns. Report the assignment as Level 2, a putative annotation, unless you ran the authentic compound yourself.

My advisor says accurate mass alone is enough to identify a metabolite — is that true for a manuscript under review?

No. Kind and Fiehn showed that below 1 ppm mass accuracy still leaves too many candidate molecular formulae to exclude alternatives in the mid-to-higher mass range, and a formula is not a structure — glucose, fructose and galactose share one exact mass. A specialist reviewer will ask for a second, independent property: isotope pattern, MS/MS fragmentation, or an authentic standard. Report accurate-mass-only assignments as putative annotations, not identifications.

The agent brief

Paste this into a custom agent. Suggested settings: work type domain_review, skill level expert, tools code execution.

Name: Metabolomics and flux reporting

You are reviewing a metabolomics or isotope tracing manuscript for reporting
completeness against the conventions specialist reviewers apply. Read the
methods, results and supplementary description together.

Determine and report:

1. Identification confidence. For every metabolite named in the results,
   determine what evidence supports the identification: an authentic standard
   run in the same batch, an MS/MS spectral match to a named library, accurate
   mass with isotope pattern, or accurate mass alone. Report metabolites named
   with a definite identity where only accurate mass supports it, and state
   that these should be reported as putative annotations rather than
   identifications.

2. Normalisation. Identify how sample-to-sample variation was normalised: to
   cell number, protein content, tissue mass, total ion current, an internal
   standard, or a quality control sample. Report analyses where normalisation
   is unstated. Note where the chosen normaliser could itself change with the
   experimental condition, which would bias every reported metabolite.

3. Internal standards and quality control. Determine whether internal
   standards were used and named, whether pooled quality control samples were
   run, and whether drift across the run was assessed. Report the absence of a
   quality control strategy in an untargeted study.

4. Isotope tracing. Where labelled substrate was used, determine whether
   mass isotopologue distributions were corrected for natural isotope
   abundance. Report the absence of that correction as a quantitative error
   rather than an omission. Determine whether isotopic steady state was
   established or assumed, and whether tracer enrichment in the medium was
   measured.

5. Flux claims. Where the text claims a change in flux, determine whether the
   evidence is a labelling pattern, a pool size, or a computed flux. A change
   in pool size alone does not establish a change in flux. Report every
   instance where the wording claims more than the measurement supports.

6. Coverage. Determine whether the number of features detected, the number
   annotated, and the number reported are all stated, and report the gap where
   they are not.

Use code execution to check internal consistency of any reported ratios,
percentages or enrichment values. Report each issue with its location and the
specific addition needed.

What this agent catches

PerfectPaper’s metabolomics agent catches nine recurring failure patterns, listed below with the specific reason a specialist reviewer objects to each one.

Normalisation: what the denominator does to every reported metabolite

Normalisation choice propagates into every metabolite in the paper, which is why a reviewer treats an unstated normaliser as a whole-manuscript problem rather than a methods detail. A metabolomics measurement is an intensity per injection; converting it to a biologically comparable quantity requires dividing by something, and that something is a measurement with its own biology.

Probabilistic quotient normalisation is the named method to reach for where dilution varies, introduced by Dieterle, Ross, Schlotterbeck and Senn in “Probabilistic quotient normalization as robust method to account for dilution of complex biological mixtures. Application in 1H NMR metabonomics” (Analytical Chemistry 2006;78:4281–4290). It estimates a single dilution factor per sample from the median of the quotients against a reference spectrum, which is more robust than total-intensity normalisation because a handful of large, condition-responsive features cannot dominate a median.

The median-based robustness has a boundary condition worth stating rather than assuming away: probabilistic quotient normalisation assumes most of the profile is unchanged between samples, so the median quotient reflects dilution rather than biology. That assumption has a published breaking point. PQN is sensitive to systematic variation shared across more than roughly half the profile, which is also the regime where the newer local-neighbour normalisation methods are proposed to take over, and it is the same threshold the table row above records. A second and more mundane source of bias is documented for 1H NMR urine specifically: excreted protein raises the spectral baseline, and the resulting PQN coefficients correlate with measured protein levels, so proteinuria can move the dilution estimate before any metabolite is read (Analytical Chemistry 2022). State which normaliser was used and, where a large fraction of the metabolome is expected to move together, whether that assumption was checked rather than inherited from the software default.

The failure the agent looks for is the condition-dependent normaliser, and it is directional rather than random. If a treatment reduces cell number by 30% and metabolites are normalised to cell number, every metabolite in the treated group is inflated by roughly the same factor, and the resulting figure shows a coordinated global increase that a reader will interpret as biology. The pattern is a technical signature closely related to a batch effect: a variable that tracks the condition, entering every measurement in one direction.

Common questions about metabolomics normalisation

Will an AI reviewer tell me which normalisation method my study should use?

No, and it should not. The correct normaliser depends on sample type and on what the treatment did to the cells or tissue. The agent checks only that the choice is stated and that the normaliser is not itself expected to change with the condition. Where dilution varies, as in urine or plasma, probabilistic quotient normalisation is the named method to reach for.

What does an AI reviewer report when the methods do not state a normalisation strategy at all?

PerfectPaper’s metabolomics agent reports the absence by location — which analyses, which figures — and states what has to be added: the normaliser used, and whether that normaliser could itself change with the experimental condition. Because the choice propagates into every reported metabolite, an unstated normaliser is reported as a whole-manuscript problem rather than a methods detail.

Should I normalise metabolomics data to total protein or to cell number?

Either can work, but both fail the same way: if the treatment changes protein synthesis, cell size, or how many cells survive to harvest, the normaliser moves with the condition and biases every metabolite in one direction. Total ion current or an internal standard avoids that dependency in some designs. Report which normaliser was used and whether the treatment could plausibly change it.

Internal standards, pooled quality control and run drift

Pooled quality control samples are the reporting item most often missing from untargeted work, and their absence makes drift and biology inseparable. A pooled QC is an aliquot of every study sample mixed together and injected repeatedly across the run; because its composition is identical every time, any variation it shows is instrumental.

Dunn and colleagues set out the large-scale workflow in “Procedures for large-scale metabolic profiling of serum and plasma using gas chromatography and liquid chromatography coupled to mass spectrometry” (Nature Protocols 2011;6:1060–1083), which describes quality control–based robust LOESS signal correction for combining data acquired across multiple analytical batches over months. Quality control–based robust LOESS signal correction, usually written QC-RLSC, is the literal method name a reviewer expects to see if you corrected drift, and naming it is worth more than a sentence about reproducibility.

Four items belong in the methods, and the agent reports each one that is absent:

• The internal standards, named individually. “Internal standards were added” is not a reported internal standard. Stable-isotope-labelled analogues of the analytes are the strongest choice in targeted work because they co-elute and experience the same ion suppression; a single standard for a broad untargeted panel does not.
• The pooled QC injection schedule. How many pooled injections, at what interval, and whether conditioning injections preceded the run.
• The acceptance filter applied to features. A widely used filter retains features whose relative standard deviation across pooled QC injections falls below 20%, with 30% applied in some workflows. Journals do not impose one threshold, so the number your study used has to be stated rather than assumed by the reader.
• The injection order. Randomised injection order is what separates drift from group. Running all controls then all treated makes acquisition order and condition the same column, which is the same defect the batch effects agent reads for in sequencing work.

Natural-abundance correction: the arithmetic a reviewer will do

Natural-abundance correction is not an optional refinement; without it, reported labelling fractions are wrong by an amount that scales with the number of atoms in the molecule. Carbon-13 occurs naturally at about 1.1%, so the M+1 peak of an unlabelled metabolite sits at roughly 1.1% × n relative to its monoisotopic peak, where n is the number of carbons: about 6.6% for a six-carbon metabolite such as citrate, and about 18% for sixteen-carbon palmitate.

That is the error before any tracer is added, and it is why an uncorrected comparison of labelling between a small metabolite and a large one is invalid. Three further contributions are routinely missed:

• Heteroatoms. Sulfur’s 34S sits at about 4.25% natural abundance, so one sulfur atom contributes far more to M+2 than two carbons do in a small molecule. Nitrogen, oxygen and hydrogen contribute smaller but non-zero terms.
• Derivatisation groups. In GC-MS, a trimethylsilyl group adds one silicon and three carbons to the correction formula. Silicon’s 29Si (about 4.7%) and 30Si (about 3.1%) contribute more to M+1 and M+2 than the three added carbons do, and derivatised metabolites often carry several such groups.
• Tracer purity. Commercial [U-13C6]glucose is typically supplied at 99 atom % 13C, so about 1% of the tracer carbon is 12C. Correcting for natural abundance and ignoring tracer purity leaves a residual bias in the opposite direction. The 99% figure is a catalogue default, not a measurement of the lot used — vendors ship a certificate of analysis with the actual atom % for that batch, and a correction that assumes 99% when the certificate reads lower silently understates enrichment.

Name the software. IsoCor, published by Millard and colleagues in Bioinformatics (2019;35:4484–4487), performs isotope correction for high-resolution MS labelling experiments. AccuCor, introduced in Su, Lu and Rabinowitz, “Metabolite spectral accuracy on Orbitraps” (Analytical Chemistry 2017;89:5940–5948), addresses a resolution-dependent complication specific to Orbitrap data: at high resolving power some natural-abundance isotopes are separated from the 13C peak rather than summed into it, so the correction matrix depends on the resolution setting used at that m/z. A methods sentence that names the tool, its version and the elemental formula used for correction closes this item outright.

Worked example: a labelling difference that does not survive correction

Consider a tracing experiment in which cells are cultured with [U-13C6]glucose and the manuscript reports that treatment increased M+2 citrate from 14% to 19%, described as increased glucose oxidation through the TCA cycle.

Citrate has six carbons, so before any tracer is present the natural-abundance M+2 fraction is small but the M+1 fraction is not, and the reported values are fractions of a distribution that also contains M+0 through M+6. If the fractions were computed from raw peak areas without correction, part of every mass isotopologue above M+0 is natural abundance rather than tracer. The correction redistributes that mass and typically shrinks the apparent labelling, and it shrinks it unequally across metabolites of different carbon number — which is exactly the comparison the paper is making when it places citrate, malate and aspartate on the same axis.

The second problem is the denominator on the tracer side. If the medium contained unlabelled glucose from serum or from a base formulation, or if glutamine also fed the TCA cycle, then the fraction of the citrate pool that could ever become labelled is below 100%, and a difference in that ceiling between conditions produces a difference in M+2 with no change in oxidation at all. Measuring tracer enrichment in the medium at harvest is what distinguishes the two, and it is a supplementary table rather than an experiment.

The third is the one that decides the sentence. A higher M+2 citrate fraction at a single time point is consistent with faster entry of glucose carbon into citrate, and it is equally consistent with a smaller citrate pool relabelling faster at unchanged flux. The wording that survives review names the measurement — “the fraction of citrate carrying two 13C atoms rose from 14% to 19%” — and the interpretation separately. The agent reports the gap between the two, in the same class of finding as the overclaim check.

Questions about isotope correction and flux claims

Can an AI reviewer tell whether my isotope data were corrected for natural abundance?

Yes — it reads for the correction step, the software and the elemental formulae used, and reports their absence as a quantitative error, not a missing sentence. Carbon-13 occurs naturally at about 1.1%, so the M+1 peak of an unlabelled metabolite sits at roughly 1.1% × n relative to its monoisotopic peak: about 6.6% for six-carbon citrate, about 18% for sixteen-carbon palmitate.

Can an AI reviewer tell whether my paper claims a flux change the measurement does not support?

Yes, and it is one of the checks the agent runs hardest. Pool size is a concentration and flux is a rate: a pool can stay constant while flux doubles, or change while flux is unaltered. The agent matches each rate verb in the results to the measurement behind it and reports the gap — a labelling percentage or a pool size is not, by itself, a flux.

Why did my M+2 citrate percentage change after I ran it through IsoCor?

Because the isotopologue fractions are shares of one distribution, correcting any of them moves all of them. Citrate has six carbons, so carbon-13’s 1.1% natural abundance alone contributes roughly 6.6% to the uncorrected M+1 fraction before any label is present, plus a smaller contribution to M+2. IsoCor subtracts that expected background using the molecule’s elemental formula and renormalises what remains, which typically shrinks the reported labelled fraction — a smaller number is the corrected one, not an error.

Pool size, labelling and flux are three different measurements

Pool size, labelling and flux answer three different questions, and conflating them is the most common substantive error in metabolism manuscripts. Pool size is a concentration, in moles per cell or per milligram of protein. Labelling is a dimensionless fraction of a pool carrying tracer atoms. Flux is a rate, in moles per unit time, and no single static measurement yields one.

Stationary 13C metabolic flux analysis computes fluxes by fitting mass isotopologue distributions to a metabolic network model, and it requires isotopic steady state — the labelling pattern no longer changing with time. Upper glycolytic intermediates approach steady labelling faster than TCA-cycle intermediates, and macromolecular pools slower still, so a single harvest time cannot be assumed to be at steady state for every metabolite reported. Where steady state is not reached, isotopically non-stationary MFA is the method, implemented in INCA (Young, Bioinformatics 2014;30:1333–1335). It is also the more demanding experiment: fitting a dynamic labelling curve needs enough time points across the approach to steady state to constrain the fit, and a sampling schedule built for a stationary design — a handful of time points, or one — will still return a flux solution from the software without adequately constraining it.

Good practice for reporting flux studies has been set out, though no agreed standard exists. Crown and Antoniewicz reviewed publishing practice in “Publishing 13C metabolic flux analysis studies: a review and future perspectives” (Metabolic Engineering 2013;20:42–48), found no consensus among researchers and editors on minimum data standards, and proposed a checklist to serve as one. A specialist reviewer of a flux paper asks for the network model and its atom transitions, the goodness of fit of the flux solution to the measured isotopologue data, and confidence intervals on each reported flux — not a standard error from a single local fit.

Seahorse extracellular flux measurements are rates, but not the rates they are usually called. Mookerjee, Gerencser, Nicholls and Brand showed in “Quantifying intracellular rates of glycolytic and oxidative ATP production and consumption using extracellular flux measurements” (Journal of Biological Chemistry 2017;292:7189–7207) that medium acidification has a substantial respiratory component, because CO2 produced by the TCA cycle hydrates to bicarbonate and protons. Extracellular acidification rate is therefore not a glycolysis rate without accounting for that contribution, and oxygen consumption rate is not an ATP production rate without correcting for proton leak and non-mitochondrial oxygen consumption. Reported maximal respiration also depends on an FCCP dose that was titrated for that cell type; an untitrated dose understates it.

Where the manuscript then reasons from a flux change to a mechanism, the objection moves out of metabolomics and into the general class covered by causal language discipline and the reviewer comment at the finding is correlative, not causal. A metabolite that changes with a phenotype is not a driver of it, and a reviewer will say so.

Coverage: features detected, annotated and reported

A feature is not a metabolite, and the difference between the two counts is roughly an order of magnitude. Mahieu and Patti demonstrated the scale in “Systems-level annotation of a metabolomics data set reduces 25 000 features to fewer than 1000 unique metabolites” (Analytical Chemistry 2017;89:10397–10406): an untargeted LC-MS data set from Escherichia coli containing about 25,000 detected features resolved to roughly 2,961 unique analytes once adducts, isotopologues, in-source fragments, multimers and other degenerate signals were grouped, and to fewer than 1,000 once an orthogonal 13C-credentialing step removed non-biological features as well.

The degeneracy is not evenly distributed across compound classes. Lipids are disproportionately affected, because a fragment produced in the ion source is frequently isomeric with a genuine member of a different lipid class rather than with nothing at all. Hu and colleagues catalogue these ion-source artefacts across ionisation methods in Mass Spectrometry Reviews (2022), and two of their examples show the shape of the problem: triacylglycerols lose a fatty acyl chain in the source to give diacylglycerol-like ions, and phosphatidylcholine and lysophosphatidylcholine fragments carry the same masses as endogenous phosphatidylethanolamine and lysophosphatidylethanolamine species. An untargeted lipidomics feature list can therefore inflate one class with fragments of another before any statistics are run, which is why a lipid annotation is checked against retention time and not mass alone.

The reporting consequence is three numbers, not one. State how many features were detected after peak picking, how many passed quality control filtering, how many were annotated and at what confidence level, and how many are discussed in the results. A paper that reports “over 20,000 metabolites were profiled” has reported a feature count under a metabolite label, and a specialist reviewer will say so in the first paragraph.

The same degeneracy matters statistically. Testing 25,000 features that collapse to 1,000 compounds means the tests are heavily correlated, so a Bonferroni correction across features is conservative while an uncorrected nominal p is anticonservative — the trade-off set out at FDR versus family-wise error. Repeated injections of one sample are not independent observations either, which is pseudoreplication in a mass spectrometer.

How to check your own metabolomics manuscript before submitting

Checking metabolomics reporting takes four passes over the methods, and none of them requires new data. Run them in this order.

Build an identification table. One row per metabolite named in the results, with columns for the evidence type, the library or standard used, and the MSI or Schymanski level. Any row whose evidence column reads “accurate mass” and whose results text reads “identified” is a wording change, made once, before a reviewer makes it for you.

Write the normalisation sentence and then attack it. State the normaliser, then ask whether the treatment could change it. If it could, either report a second normalisation and show the conclusion holds under both, or state the direction the bias would push every metabolite.

Reconcile every reported percentage. Mass isotopologue fractions for one metabolite must sum to 100%. Enrichment values, percentage changes and fold changes in the text must match the figures. This is the check the agent runs with code execution, and it is the fastest source of findings in a metabolomics manuscript because the numbers usually pass through a spreadsheet before reaching the text.

Read the results for verbs the methods do not support. “Increased flux”, “enhanced oxidation”, “shifted metabolism toward” and “reprogrammed” all assert rates. Match each to the measurement behind it, and where the measurement is a pool size or a single-time-point labelling fraction, rewrite the sentence to name what was measured.

Frequently asked questions

Can an AI reviewer check metabolomics reporting standards in my manuscript?

Yes, when it is briefed on the conventions. PerfectPaper’s metabolomics agent reads the methods, results and supplementary description together and reports identification evidence per metabolite, the normalisation strategy, the quality control design, natural-abundance correction, flux wording and the feature-to-metabolite counts. It reports what is missing and where; it does not run your data or replace a specialist reviewer’s judgement.

Can an AI reviewer check my metabolomics reporting before I submit, rather than after review?

Yes, and that is the point of running it. Build a table of every metabolite with its identification evidence and confidence level; write the normalisation sentence and ask whether the treatment could change the normaliser; confirm every mass isotopologue fraction sums to 100%; and match each rate verb in the results to the measurement behind it. The agent runs those four passes and reports each gap with its location.

Reviewer 2 says my metabolite identifications are not rigorous enough — what should I check before responding?

Build the identification table first: one row per named metabolite, its evidence, and its MSI or Schymanski level. A metabolite reported as identified but supported only by accurate mass is very likely reviewer 2’s objection, and downgrading the wording to “putatively annotated” answers it without new experiments. If the paper’s central claim rests on a few compounds, running an authentic standard for those converts them to Level 1.

What reviewers write when metabolomics reporting is thin

Reviewers rarely write “reporting standards” as their first phrase. They write the specific version, and these are the comments that decide the paper.

“Metabolites are named throughout without any statement of identification confidence; please provide MSI levels and state which assignments rest on authentic standards.” “It is not clear whether the reported mass isotopologue distributions were corrected for natural isotope abundance; please state the software and the elemental formulae used.” “The normalisation strategy is not described; per-cell values cannot be interpreted without knowing what the treatment did to cell number.” “No pooled quality control samples appear to have been run, so instrumental drift cannot be distinguished from the reported group differences.” “Samples appear to have been acquired in group order; please provide the injection sequence.” “The authors report increased glycolytic flux on the basis of ECAR alone.” “Tracer enrichment in the medium is not reported, so the labelling ceiling is unknown.” “Isotopic steady state is assumed rather than demonstrated at the single harvest time used.” “Over 20,000 metabolites are reported; this appears to be a feature count.” “The raw data are not deposited in MetaboLights or Metabolomics Workbench.”

The natural-abundance comment is the one authors find hardest to answer late, because it requires the raw isotopologue areas rather than the processed fractions, and those are sometimes only in a collaborator’s instrument software. Drafting the reply is covered at how to write a response to reviewers; what works is the corrected table and the tool version, not a paragraph explaining that the effect would be small.

When the obvious remedy is unavailable

Buying authentic standards for every named metabolite and re-running the samples is frequently impossible, and the honest alternative is a downgraded claim with the evidence stated, not silence. When the extracts are exhausted, the instrument time is gone, the collaborator has moved, or the standards for a class do not exist commercially, the manuscript still has five moves.

Downgrade the language, not the finding. Replacing “we identified” with “we putatively annotated, at MSI Level 2” concedes nothing scientifically and removes the objection entirely. The biology in the paper is unchanged; only the epistemic claim about each assignment moves.

Correct natural abundance retrospectively. Correction is a computation on the raw isotopologue peak areas, not a new experiment. If the raw files exist, IsoCor or AccuCor can be run on them after submission, and the corrected table becomes a revision rather than a resubmission.

Report the normaliser you have, plus a second view. Where cell counts were not recorded, total protein or total ion current can usually be recovered from data already collected, and showing the conclusion under two normalisations is a stronger response than defending one.

Validate the specific compounds the argument rests on. Standards for the three or four metabolites carrying the paper’s conclusion are a much smaller undertaking than standards for the whole panel, and they convert the central claims to Level 1 while the rest remain annotations. This is the metabolomics form of orthogonal validation, and it is the response reviewers accept most often.

Write the limitation with a direction. Name which way an uncorrected or condition-dependent normaliser would push the estimate and why, rather than writing that results should be interpreted with caution. How to write a limitations section sets out the form; a limitation that names the mechanism and the sign reads as an analysis.

How to report it so a reviewer can check it

Report metabolomics as a sample-level table plus a metabolite-level table, deposited with the raw data. Those two artefacts answer most of the questions above before a reviewer has to ask, and they are the format the agent expects to find.

The sample table carries one row per injection: sample identifier, group, biological unit, extraction batch, injection order, acquisition date, and whether the row is a study sample, a pooled QC or a blank. Injection order and acquisition date are the two columns that let a reviewer check for drift, and they are the two most often omitted.

The metabolite table carries one row per reported compound: name, molecular formula, adduct, m/z, retention time, identification evidence, library or standard used, confidence level, and the QC relative standard deviation. For tracing studies, add the elemental formula used for natural-abundance correction and the correction software with its version.

mzTab-M is the community data standard for sharing quantitative metabolomics results, published by Hoffmann and colleagues in Analytical Chemistry (2019;91:3302–3310). Raw and processed data belong in MetaboLights or Metabolomics Workbench, with the accession stated in the manuscript rather than a data-availability sentence carrying no identifier — the gap covered by the data availability and accessions agent. For animal tracing studies, the housing, diet, fasting state and infusion protocol belong alongside them.

How this differs from the built-in review

PerfectPaper’s metabolomics agent differs from the standing review team in domain depth, not in analytic ability. The standing team includes methodology and study-design specialists and a numeric integrity specialist that recomputes reported arithmetic, but neither is briefed on identification confidence conventions or on natural-abundance correction, which are unique to mass spectrometry work and are where most metabolomics reviewer comments land.

The distinction is domain knowledge rather than analytic ability. A general statistical reader can see that a percentage does not sum to 100 and cannot see that a six-carbon metabolite’s M+1 peak was never corrected, because nothing in the manuscript looks wrong — the number is present, formatted, and plausible. That is the class of defect a briefed specialist agent finds and an unbriefed one does not.

If your paper combines metabolomics with transcriptomics, pair this with multiple testing and enrichment. Full set: AI peer review for genomics.

Review my manuscript

Related agents

Four sibling agents share a review with this one; you can run three at a time. See AI peer review for genomics for the full set of five omics agents.

• Multiple testing and enrichment — use it when a gene list is called significant on nominal p, or enrichment runs against the wrong background.
• Data availability and accessions — use it before submission, since a missing accession is what gets a paper returned without review.
• Batch effects and replicate structure — use it when samples crossed runs or days, or when n counts passages rather than animals.
• Variant and gene nomenclature — use it when the paper reports specific variants that will be curated downstream.