A custom AI reviewer for flow cytometry reporting

PerfectPaper’s flow cytometry agent reads a manuscript for the reporting gaps that make a cytometry result uninterpretable: a gating strategy that is never shown, missing viability or fluorescence-minus-one controls, unstated compensation, and percentages whose denominator changes between panels without being named. Copy the brief below into a custom agent slot.

The denominator problem is the one to look for first. A figure axis reading “% CD8+” can mean percentage of live cells, of CD45+ cells, of lymphocytes, or of the whole sample, and those numbers differ by an order of magnitude.

MIFlowCyt — the Minimum Information about a Flow Cytometry Experiment standard, published by Lee and colleagues for the International Society for Advancement of Cytometry (Cytometry Part A 2008;73A(10):926–930) — states the requirement in one clause: “It is important to define qualitative gate descriptors as well as to state denominators used to calculate percentages of events within a gate.” That clause has been in print since 2008. The reason it still goes unmet is mechanical rather than careless: inside the analysis software the denominator is unambiguous, because the parent gate sits directly above the child in the hierarchy. It becomes ambiguous the moment the plot is exported and the axis is labelled % CD8+.

When to use this agent

This agent is worth running whenever a manuscript’s central claim depends on immune population frequencies from flow cytometry, and specifically whenever any of the following hold:

• Your manuscript reports immune population frequencies from flow cytometry
• Your central claim is a change in a cell population between conditions
• Panels grew over the course of the project and were run by different people
• You report both frequencies and absolute counts
• You are submitting to an immunology journal with specialist reviewers
• Your gating strategy exists inside FlowJo or FCS Express but has never been exported as a figure
• A reviewer has already asked what your percentages are a percentage of
• The population carrying your claim is below 1% of its parent gate
• Samples were acquired across several months, on more than one cytometer, or after a panel redesign
• Your .fcs files are going into FlowRepository and the deposit needs a gating description

The agent brief

The brief below is the exact text to paste into a custom agent slot, with suggested settings: work type domain_review, skill level graduate, capability vision if your gating figure is a supplementary panel.

Name: Flow cytometry reporting

You are reviewing flow cytometry reporting in a manuscript. Read the methods,
every figure axis label, every figure legend, and the results text together.

Determine and report:

1. Denominators. For every reported percentage, determine the parent
   population it is a percentage of. Report every instance where the axis
   label, the legend and the results text do not agree, and every instance
   where the parent cannot be determined at all. Where two figures report the
   same population against different parents, report the inconsistency
   explicitly and note that the values are not comparable. Treat this as the
   highest priority item in this brief.

2. Gating strategy. Determine whether a full gating strategy is shown,
   including the sequence from initial scatter gate to each reported
   population. Report its absence. Report gating described only in words where
   the sequence is ambiguous.

3. Standard gates. Determine whether the analysis excludes debris, excludes
   doublets, and discriminates live from dead cells with a viability reagent.
   Report each that is absent. A frequency computed without dead-cell exclusion
   is unreliable for any population that dies differentially between groups.

4. Controls. Determine whether fluorescence-minus-one or isotype controls are
   described for gates placed on continuous markers. Report gates on markers
   with no clear negative population where no control is described.

5. Compensation. Determine whether compensation or spectral unmixing is
   described, and whether single-stain controls are mentioned. Report its
   absence for panels above three colours.

6. Reagents. Determine whether antibody clones are stated. A marker named
   without a clone cannot be reproduced; report each.

7. Frequency versus count. Determine whether the claim rests on a frequency or
   an absolute count, and whether the text distinguishes them. A frequency can
   fall while the absolute number rises if another population expands. Report
   every claim of a population decrease supported only by a frequency.

8. Cell numbers. Determine whether the number of events acquired is stated,
   particularly for rare populations.

Report each issue with its exact location and the specific addition needed.

The order of the brief is itself the finding. Item 1 sits first because a percentage without a named parent cannot be repaired by improving items 2 through 8 — there is no number to repair. Items 2 to 5 sit above 6 to 8 because a gating or control defect changes the value, while an unstated clone only stops someone else reproducing it. If you are adapting this brief for a different assay, see how to write a custom reviewer agent and the custom agent slots that hold it.

Frequently asked questions

What does an AI reviewer check in flow cytometry reporting?

An AI reviewer briefed for flow cytometry checks eight things: whether every percentage has a named parent population, whether a full gating strategy is shown from all events down, whether debris, doublets and dead cells are excluded, whether FMO or isotype controls back gates on continuous markers, whether compensation is described, whether antibody clones are stated, and whether a claim rests on a frequency or an absolute count.

How do I get an AI reviewer to check my flow cytometry methods before submission?

Paste a flow cytometry review brief into a custom agent slot, set the work type to domain_review, the skill level to graduate, and enable vision if the gating strategy is a supplementary figure; the agent then reads the methods, axis labels, legends and results text together and reports each issue with its location and the specific fix needed.

What this agent catches

This agent catches eleven distinct reporting failures, each one enough on its own to make a cytometry result uninterpretable to a specialist reviewer:

The denominator problem, worked through

A cytometry percentage is a ratio, and the ratio changes by whatever factor separates the two candidate parents. Take a manuscript reporting CD8+ T cells. Suppose CD45+ cells are 60% of live cells, CD3+ cells are 50% of CD45+, and CD8+ cells are 20% of CD3+. The same population is 20% of CD3+ cells, 10% of CD45+ cells and 6% of live cells. One axis label, three defensible numbers, a 3.3-fold spread between the widest pair — and nothing in the figure tells the reader which was plotted.

The damage compounds when the parent differs between panels of the same paper. Figure 2 reports the population against live cells; Figure 4, drawn a year later after the panel was extended, reports it against CD45+. The two figures are then not on the same scale, the text compares them anyway, and the discrepancy reads to a reviewer as either an error or an undisclosed reanalysis. The agent reports this case explicitly and states that the values are not comparable, rather than reporting each panel as internally fine.

There is a second, subtler version. A frequency reported against a parent that itself changes between groups is not a stable measure of anything. If treatment halves the myeloid compartment, every lymphocyte frequency expressed as a percentage of CD45+ rises without a single lymphocyte being made. Naming the parent is the minimum; choosing a parent that the treatment does not move is the actual methodological decision, and papers rarely say they made it.

The fix is one sentence in the legend and one convention applied throughout: report each population against a single, explicitly named parent, write it into the axis label in the form CD8+ (% of live CD45+), and state in the methods that all frequencies use that hierarchy. Where two hierarchies genuinely coexist, plot both.

There’s a reason no denominator is ever perfectly stable, and it is worth naming rather than working around. Percentages within a gating hierarchy are compositional data: they are constrained to sum to the whole of their parent, so sibling populations cannot vary independently of one another the way unconstrained counts can. Choosing a parent the treatment does not move reduces the problem; it cannot eliminate it, because some redistribution among sibling populations is close to the definition of an immune response. This is the same constraint that pushed microbiome analysis toward compositional-data methods rather than raw percentages, and the reasoning applies to a gating hierarchy for the same reason.

What a complete gating strategy figure contains

A gating strategy figure shows the ordered sequence of gates from the raw acquisition to every population you report, with each plot’s parent identifiable from the plot above it. MIFlowCyt’s requirement is directional and absolute: “The overall gating strategy shall be traceable up to the set of all events.” A supplementary panel showing three plots that begin at “lymphocytes” does not satisfy it, because the reader cannot see how lymphocytes were defined.

Nature Portfolio’s Reporting Summary makes the figure a checkbox. Its Flow Cytometry section carries five Methodology fields — sample preparation, instrument, software, cell population abundance and gating strategy — and one confirmation: “Tick this box to confirm that a figure exemplifying the gating strategy is provided in the Supplementary Information.” The same section asks authors to confirm four things about the plots themselves: “The axis labels state the marker and fluorochrome used (e.g. CD4-FITC)”; “The axis scales are clearly visible”; “All plots are contour plots with outliers or pseudocolor plots”; and “A numerical value for number of cells or percentage (with statistics) is provided.”

Those four are worth reading as a checklist because three of them are usually failed by the same figure. Density plots rendered without outliers hide the rare events that a rare-population claim depends on. Axis labels giving CD4 rather than CD4-FITC remove the information a reviewer needs to judge spillover. And a gate drawn without the percentage inside it makes the plot decorative.

MIFlowCyt goes further on gate definition than most manuscripts do: “Gating descriptions shall be clear and unambiguous, preferably by mathematical descriptions of each gate boundary.” A screenshot of the hierarchy meets that requirement more completely than a paragraph of prose, which is one reason the figure is the deliverable and the sentence is not.

This entire framework assumes sequential, manual gating. Where the analysis instead runs through an unsupervised method — FlowSOM, viSNE or a UMAP embedding of the marker space — a conventional gating tree does not apply, and MIFlowCyt’s traceability requirement is better met by naming the algorithm and its parameters, the markers fed into it, and, wherever a resulting cluster is asserted to correspond to a canonical population, the manual gate used to confirm that correspondence. A clustering analysis that reports a population percentage with no such confirmation has the same denominator problem in a different disguise.

Common questions about denominators and gating strategy

Can an AI reviewer tell me what my percentages are a percentage of?

An AI reviewer reconstructs the intended parent for every reported percentage from the axis label, the legend and the results text, and reports each place those three disagree or the parent cannot be determined. This is the agent’s first-priority item, because the same measurement reported against different parents gives very different numbers — the exact requirement MIFlowCyt states directly for gate descriptors and denominators.

Why did my CD8 percentage change when I gated off live cells instead of CD45?

Because the two gates are different parents, and the ratio between them is rarely one-to-one. If CD45+ cells make up 60% of live cells, a CD8+ population reported against live cells will read smaller than the same cells reported against CD45+ — same cells, same count, different denominator. State which parent you used, and report both if it matters.

Can an AI reviewer check my gating strategy?

An AI reviewer can check whether a gating strategy is present, complete and traceable from all events to each reported population, and whether the sequence described in the methods matches the sequence shown in the figure. It cannot open your .fcs files or judge whether a drawn gate encloses the right cells — that requires the raw data, not the manuscript.

Does an AI reviewer flag a missing gating strategy figure?

Absence of a full gating strategy is one of the eight items the agent reports, and most immunology journals expect that figure as supplementary material. Nature Portfolio’s Reporting Summary makes it a confirmation checkbox, asking authors to confirm that a figure exemplifying the gating strategy is provided in the Supplementary Information. The agent reports its absence because gating placement is where most of the interpretive latitude in cytometry lives.

Debris, doublets and dead cells: the three gates before any marker

Debris exclusion, doublet exclusion and viability discrimination are the three gates that sit before any marker gate, and their absence changes the number rather than merely the presentation. The agent reports each independently because they fail independently.

Doublets are the one most often skipped. Two cells traversing the laser together produce a single event whose area signal is roughly doubled while its height signal is not, so plotting FSC-A against FSC-H separates them; a doublet of a CD4+ cell and a CD8+ cell is recorded as a CD4+CD8+ event that exists nowhere in the biology. Any manuscript reporting a rare double-positive population without a doublet gate has an alternative explanation for it that a reviewer will raise first.

FSC-A against FSC-H is not the only working pair. A width-based gate — FSC-A against FSC-W, often followed by SSC-A against SSC-W — makes the same discrimination from the low pulse width that a single cell produces. Which pair is open to you is set by the instrument and by which pulse parameters were enabled for that detector at acquisition rather than by preference, so state the pair you used rather than leaving a reader to guess from an unlabelled biaxial plot.

Dead cells matter for a reason specific to antibody staining. Cells with compromised membranes bind immunoglobulin non-specifically and take up dye across every channel, which is why amine-reactive fixable viability dyes became standard practice — Perfetto and colleagues set out the approach in “Amine reactive dyes: an effective tool to discriminate live and dead cells in polychromatic flow cytometry” (Journal of Immunological Methods 2006;313(1–2):199–208), and their advantage over propidium iodide or 7-AAD is that they survive fixation and permeabilisation. Forward and side scatter alone do not identify dead cells reliably enough to substitute.

The consequence is asymmetric, which is what makes it a review finding rather than a technicality. If one arm of your experiment kills cells — a cytotoxic agent, an irradiation dose, a longer culture — dead cells accumulate differentially, they stain non-specifically, and the group difference you report is partly a difference in how much debris got counted. That is a confound in the same class as the batch problems covered by batch effects and confounded designs, arriving through a cytometer instead of a sequencer.

FMO, isotype and what actually places a gate

A fluorescence-minus-one (FMO) control is what actually places a gate on a marker with a continuous distribution, because nothing in the stained sample alone marks where the negative population ends. An FMO contains every antibody in the panel except one, so the gate for the missing marker is placed against the spread produced by all the others — the spread that actually exists in the tube being gated.

An isotype control answers a different question, and Maecker and Trotter’s tutorial is direct that the choice is contested: “The type of controls to include in various kinds of flow cytometry experiments is a matter of some debate and discussion” (“Flow cytometry controls, instrument setup, and the determination of positivity”, Cytometry Part A 2006;69A(9):1037–1042). The practical distinction is that an isotype tube estimates non-specific binding of an immunoglobulin of matched class, while an FMO tube estimates the fluorescence spread contributed by the rest of the panel. In a polychromatic panel the second dominates, and it is what a reviewer of a nine-colour panel is asking for.

Hulspas and colleagues decompose the background a control has to account for into “three major causes of background (autofluorescence, spectral overlap, and undesirable antibody binding)” (Cytometry Part B: Clinical Cytometry 2009;76B(6):355–364). An unstained tube addresses the first, an FMO addresses the second, an isotype or Fc-block addresses the third; no single tube addresses all three, which is why “we used isotype controls” is not a complete controls sentence for a large panel.

The need is not uniform across the panel, which is worth stating so the requirement does not read as blanket dogma. A lineage marker with a naturally bimodal, well-separated distribution — CD3, CD19, CD14 — rarely needs an FMO, because the negative population is visually obvious without one. The requirement bites hardest for markers with continuous or unimodal-shifted expression, where no natural gap exists to gate on: activation and exhaustion markers such as PD-1, CD25 or CTLA-4 are the cases a reviewer is actually asking about when they ask for an FMO.

Report which control was used for which gate, not which controls were run. A methods line reading “FMO controls were used to set gates for CD25, PD-1 and CD69; unstained and single-stain controls were acquired for every experiment” resolves item 4 completely. The broader question of which controls belong in a manuscript at all is covered in reviewer wants more controls.

Questions about controls and compensation

Will an AI reviewer notice that my gates were set without FMO controls?

Gates placed on continuous markers are checked specifically for a described control, and the agent reports every marker with no clean negative population where none is named. A fluorescence-minus-one control contains every antibody in the panel except one, so the gate is placed against the spread the rest of the panel actually produces — a different question from the one an isotype control answers.

Reviewer 2 wants FMO controls but I only ran isotype — what do I say?

Say plainly which question your isotype control answers — non-specific antibody binding — and which it doesn’t: the fluorescence spread the rest of your panel adds to the gate you’re placing. Where the marker’s expression is continuous rather than bimodal, offer to reanalyse against an internal negative population in the same tube and name that as the substitute, not as equivalent to an FMO.

Compensation, spectral unmixing and spillover spreading

Compensation is arithmetic, not judgement, and describing it is a reporting requirement rather than a courtesy. MIFlowCyt asks that “the type of compensation shall be described (e.g., no compensation, hardware compensation, computed compensation) and the spillover or compensation matrix shall be provided whenever available.” Single-stain controls, acquired on the same cells or on beads with the same fluorochrome-conjugate, are what the matrix is computed from; a panel above three colours compensated without them has no stated basis.

Compensating by eye until a plot looks square is the failure mode, and it has been documented for twenty-five years — Roederer’s “Spectral compensation for flow cytometry: visualization artifacts, limitations, and caveats” (Cytometry 2001;45(3):194–205) is the standard reference on why a correctly compensated plot can still look wrong, and Herzenberg and colleagues’ “Interpreting flow cytometry data: a guide for the perplexed” (Nature Immunology 2006;7(7):681–685) covers the display side. The manuscript-visible tell is a population sitting on the axis rather than distributed around a median, or a diagonal smear the text treats as a real double-positive.

Compensation does not remove the error it corrects for; it moves the median and leaves the variance. Nguyen and colleagues quantified this as the spillover spreading matrix, taking up the earlier theoretical prediction that spreading rises with the square root of fluorescence intensity and showing that the matrix is stable enough to serve as a quality-control metric, with non-trivial values showing a coefficient of variation under 30% across a two-month window (“Quantifying spillover spreading for comparing instrument performance and aiding in multicolor panel design”, Cytometry Part A 2013;83A(3):306–315). The practical consequence for a manuscript is that a dim marker measured in a detector receiving heavy spillover from a bright one has a resolution limit no amount of compensation improves — which is exactly the case where an FMO is not optional.

Compensation error also runs in a direction that is easy to miss because it looks like success rather than failure. Under-compensation inflates a population with spillover-positive events; over-compensation pushes true low-level positive cells into negative values, which reads as a clean, well-separated plot rather than as an error. A reviewer who asks to see an uncompensated or single-stain-control plot alongside the compensated one is checking for over-correction specifically, not repeating the under-compensation question.

Spectral cytometers substitute full-spectrum unmixing for pairwise compensation, and the reporting obligation is the same shape: name the unmixing algorithm, state whether autofluorescence was extracted as a separate parameter, and say whether single-stain reference controls were run on cells or on beads. Cells and beads give different reference spectra, and mixing them is a documented source of unmixing error.

Frequency versus absolute count

A frequency and an absolute count can move in opposite directions for the same biology, and the manuscript has to say which one carries the claim. Frequencies within a sample are constrained to sum to the whole, so any population’s percentage falls when another population expands, whether or not the first population lost a single cell.

Work the arithmetic once and the point stops being abstract. A control mouse yields 100,000 live cells of which 10,000 are CD8+, a frequency of 10%. A treated mouse yields 200,000 live cells — the treatment expanded the myeloid compartment — of which 15,000 are CD8+, a frequency of 7.5%. The frequency fell by a quarter. The absolute number rose by half. A results sentence reading “treatment reduced CD8+ T cells” is false, and a reviewer with the count in front of them will say so.

The two measures also fail differently under a statistical test, which is a separate finding from which one the biology supports. A percentage bounded at 0 and 100 violates the assumptions behind an unmodified t-test as it approaches either bound, which is why an arcsine-square-root or logit transform is standard practice before testing a cytometry frequency; a count, by contrast, is usually better served by a Poisson or negative-binomial model that reflects its own variance structure. A manuscript that runs a plain t-test on raw percentages near either bound has a second, independent problem.

Absolute counts require a counting method, and the method belongs in the methods. Single-platform counting adds a known number of fluorescent beads to a fixed sample volume and derives the count from the bead-to-cell ratio. Dual-platform counting multiplies a cytometry frequency by a cell concentration measured on a separate haematology analyser or haemocytometer, and inherits the error of both instruments. For tissue, the denominator is usually cells per milligram or per organ, and it depends on a digestion protocol whose recovery differs between inflamed and uninflamed tissue — which is itself worth stating.

The agent reports every claim of a population decrease supported only by a frequency. Where both are available, report both, and let the text say which one the conclusion rests on. Where only a frequency exists, narrow the claim to a change in composition rather than a change in cell number. That is the same discipline the overclaim check applies to text, reaching the same manuscript through a figure axis.

How many events you need for a rare population

The precision of a cytometry count is set by Poisson statistics, and the rule is one line: the coefficient of variation of a count of n events is approximately 1/√n. One hundred events in a gate gives roughly a 10% CV; 400 events gives 5%; 10,000 gives 1%. A gate containing 12 events carries a CV near 29%, and a between-group difference resting on it is not measurable however small the p-value looks.

The acquisition arithmetic follows directly. A population at 0.01% of acquired events needs 1,000,000 events acquired to record 100 of them. Manuscripts reporting antigen-specific T cells, circulating tumour cells, haematopoietic stem cells or residual disease routinely need six-figure or seven-figure event counts, and the number acquired is the fact that lets a reader judge whether the reported frequency means anything. This is why item 8 of the brief exists and why it is not merely bookkeeping.

Report the events acquired per sample and the events in the reported gate, not one or the other. “500,000 live events acquired per sample; 84–210 events in the tetramer+ gate” tells a reviewer everything they need. “Data were acquired on a five-laser cytometer” tells them nothing about precision.

The 1/√n rule is a floor, not the complete account, and it is worth being precise about why. It captures only the counting error of the population itself. A rare gate defined against an FMO or unstained background subtracts that background’s own count, and where the background and the positive gate are similar in size, the subtraction’s uncertainty can dominate the reported precision more than the positive count does. A reviewer asking for the FMO or background count alongside the positive-gate count is asking a Poisson-statistics question, not a formality.

Two related failures travel with rare-event work. The first is a threshold problem: a population at the limit of detection is defined by where the gate sits, and moving the gate a small distance changes the count by a large fraction — which makes the FMO control load-bearing rather than decorative. The second is multiplicity. A panel that reports 24 populations across 4 comparisons has made 96 comparisons, and the correction belongs in the paper; see multiple testing in omics and high-dimensional panels for how reviewers expect that handled.

Questions about frequencies, counts and rare populations

Can an AI reviewer catch a claim that rests on a frequency rather than a cell count?

Claims of a population decrease supported only by a frequency are reported as a distinct finding. All frequencies in a sample sum to the whole, so a population’s percentage falls when another expands — 10,000 CD8+ cells in 100,000 live cells is 10%, and 15,000 in 200,000 is 7.5%, a frequency that fell while the number rose by half. The agent asks for the absolute count.

Is it a problem that my rare population only has 30-40 events in it?

Yes — a gate with 30 to 40 events carries a Poisson coefficient of variation of roughly 16-18% (1/√n), so a between-group difference has to be large to survive that noise. Report the exact count alongside the percentage, and if the background or FMO gate has a similar event count, say so: the subtraction’s own uncertainty may dominate the result.

My frequency went down but I don’t think my treatment actually killed any cells — what’s going on?

Frequencies are constrained to sum to the whole of their parent gate, so any population’s percentage falls whenever a different population in the same parent expands — no cell death required. Pull the absolute count for the population in question; if it held steady or rose while the percentage fell, the biology is compositional, not a loss, and the results text should say so.

Antibody clones, panel tables and the reagents section

An antibody is identified by its clone, and a marker named without one cannot be reproduced. Two anti-CD11b antibodies with different clones can differ in epitope, in affinity, and in whether they detect the activated conformation, so “anti-CD11b” specifies an intention rather than a reagent.

Clone alone can still leave a gap, because a clone name does not guarantee which physical lot or conjugate a reader is holding. The community’s answer is the Research Resource Identifier (RRID) system, which resolves a clone-plus-vendor-plus-catalogue-number combination to one citable string, so citing one subsumes clone, vendor and catalogue number into a single lookup. Uptake is wide rather than universal, and it is worth being exact about that: RRIDs appeared in more than 5,000 articles across more than 380 journals by 2017 and in many more since, but a requirement at every journal is still something the field argues for rather than something it has (Kahn et al., “Antibody characterization is critical to enhance reproducibility in biomedical research”, eLife 2024;13:e100211). Check your target journal’s instructions rather than assuming either way. An RRID is also not a lot number — different lots of the same catalogue item carry the same identifier — so where lot-to-lot variation carries the result, report the lot as well.

The convention that survives review is a panel table carrying, per row: the marker, the clone, the fluorochrome, the vendor and catalogue number, the lot where it matters, the dilution or amount per test, and the detector or filter the fluorochrome was read in. That table also answers a reviewer’s spillover question in one glance, because it makes the fluorochrome-to-detector mapping explicit — the same information Nature Portfolio’s CD4-FITC axis-label rule is trying to surface.

Instrument and software belong in the same paragraph: the cytometer model and its laser and filter configuration, the acquisition software with version, the analysis software with version, and the FCS version written. The file format has a published specification — “Data File Standard for Flow Cytometry, version FCS 3.1” (Spidlen et al., Cytometry Part A 2010;77A(1):97–100), superseded by FCS 3.2 (Spidlen et al., Cytometry Part A 2021;99(1):100–102) — and naming the version you wrote is what lets someone else open your raw data at all.

Depositing the .fcs files closes most of this in one action. FlowRepository is the community repository for flow cytometry data, and a deposit carries the raw list-mode files, the compensation matrix and the gating hierarchy together, which satisfies items 1, 2 and 5 of the brief in a way no prose paragraph does. Cite the accession in the data availability statement rather than in a figure legend; see data availability and accession numbers for how reviewers check those.

How to check your own cytometry figures before a reviewer does

Flow cytometry reporting defects are visible in your own draft, and four passes find most of them without opening the raw data.

Pass one — write the parent beside every percentage in the paper. Go through each axis label, each legend and each results sentence and write the parent population you intended. Anywhere the three disagree, or anywhere you cannot write a parent from memory, is an item 1 finding. This pass usually takes twenty minutes and usually produces the largest single improvement available to the manuscript.

Pass two — draw the gating tree on one sheet of paper, from all events down. If the tree you draw has more levels than the supplementary figure shows, the figure is incomplete. If two branches reach the same population by different routes, the two numbers those branches produce are different numbers.

Pass three — for every claim of a decrease, find the absolute count. Open the export and check whether the count moved in the same direction as the frequency. Where it did not, or where no count exists, rewrite the sentence before a reviewer rewrites it for you.

Pass four — hand your methods paragraph to a colleague who did not run the experiment and ask them to reconstruct the panel. Whatever they cannot recover — a clone, the viability reagent, which gates were set on FMOs, how compensation was computed — is what a reviewer will ask for, and it is the same methods objection whether it arrives from the next bench or from a referee.

What reviewers say when cytometry reporting is mishandled

Reviewers rarely write “flow cytometry reporting”. They write the specific version, and these comments hold papers through a full revision round.

“Please state the parent population for each frequency reported in Figures 2 and 4; the axis labels are inconsistent between them.” “A full gating strategy, from all events to each reported population, should be provided as a supplementary figure.” “Were doublets excluded? The CD4+CD8+ population in Figure 3C is not otherwise explained.” “No viability dye is described; given the cytotoxicity of the treatment, dead-cell exclusion is essential before any frequency is interpreted.” “How were gates for PD-1 and CD25 placed? FMO controls should be shown.” “Was compensation checked for over-correction as well as under-correction? The CD25+ population in Figure 3 sits directly on the axis, which looks more like over-compensation than a genuine negative shift.” “If clusters were called using FlowSOM or a similar unsupervised method, please state the parameters and confirm each cluster labelled CD8+ against a conventional gate.” “Please state how compensation was computed and provide the spillover matrix.” “Antibody clones and catalogue numbers are not given for any reagent in the panel.” “The authors report a decrease in the frequency of this population; please provide absolute counts, since the myeloid compartment expands in the treated group.” “How many events were acquired, and how many events fall in the tetramer+ gate?” “Were samples acquired across multiple days, and if so how was instrument performance tracked?”

The last of those is the one authors handle worst, because the honest answer is often that acquisition ran over four months on whatever settings the core facility had that morning. Reporting a tracking-bead target, a stated PMT voltage regime, or an application-settings export closes it. Silence converts a methods query into a reproducibility query, and the two are answered very differently — see how to write a response to reviewers.

What to do when the obvious remedy is unavailable

The obvious remedy — reacquire the samples with FMO controls, a viability dye and one consistent gating hierarchy — is usually out of reach by the time a reviewer asks. Four situations recur, and three of them are fixable from the files you already have.

The gating hierarchy is inconsistent between figures. Regate. The .fcs files hold every event; the hierarchy is a saved workspace, not a property of the data. Reanalysing every sample against a single named parent is an afternoon of work and it converts the paper’s largest finding into a non-finding.

No FMO controls were acquired. You cannot manufacture one retrospectively, but you can bound the problem. Show the gate on a biologically negative population from the same tube, state plainly that gates were placed against an internal negative rather than an FMO, and restrict claims for continuous markers to comparisons of median fluorescence intensity between groups measured identically, which does not require the gate. Then say so in the limitations section with the direction of the residual uncertainty named.

No viability dye was used. Scatter-based exclusion is weaker but it is not nothing: report the debris gate you did apply, report viability measured by an orthogonal method if any exists (trypan blue at the bench, for instance), and state whether the groups differed in viability at the point of staining. Where one group was cytotoxically treated, the honest move is to say that dead-cell contamination would bias the reported difference and to name which direction.

Absolute counts were never collected. Frequency-only data can still support a compositional claim. Rewrite the sentence to describe what was measured — a change in the proportion of the parent gate — and remove the implication of a change in cell number. That is a smaller loss than defending an unsupported claim across two review rounds, and it is the same repair a reviewer recommends when the unit of analysis will not stretch to the claim.

In every case, name the limitation, its direction and its likely size. “Gates for continuous markers were set against internal negatives rather than FMO controls, which may place them conservatively and understate the positive fraction in both groups” is a specification. “Results should be interpreted with caution” is not.

How this differs from the built-in review

This flow cytometry agent differs from PerfectPaper’s standing review team by supplying domain knowledge the general reviewers do not have. The standing team includes figure-to-text consistency and table integrity specialists that check whether numbers agree across the manuscript, but they do not know that a cytometry percentage requires a named parent population, or that a frequency and an absolute count can move in opposite directions for the same biology. That domain knowledge is what turns a consistency check into a substantive review comment.

The agent is also not a reanalysis tool. It reads the description of a measurement against the claim the measurement supports; it does not open .fcs files, recompute a compensation matrix, or check whether a drawn gate encloses the population it claims to. Reviewing the reported analysis and auditing the raw data are different jobs, and this brief does only the first.

Where the same manuscript nests cells within animals — five mice, forty tubes, an n of forty in the legend — the finding belongs to a different agent. See what is pseudoreplication for the statistical consequence and reviewer says n is not independent for the comment it produces.

Where your central figure is an image rather than a plot, use image quantification integrity instead. Full set: AI peer review for preclinical cancer research.

For the underlying standards, two documents are worth having open while you revise: MIFlowCyt (Cytometry Part A 2008;73A(10):926–930), whose four top-level components are experiment overview, sample, instrument details and data analysis; and the third edition of “Guidelines for the use of flow cytometry and cell sorting in immunological studies” (Cossarizza et al., European Journal of Immunology 2021;51(12):2708–3145), which runs from page 2708 to page 3145 and is the field’s consolidated methods reference.

Review my manuscript

Related agents

Each of the agents below shares a review with this one; you can run three at a time. See AI peer review for preclinical cancer research for the full set of five preclinical agents.

• Unit of analysis and pseudoreplication — use it when a figure legend counts cells, tumours or wells rather than animals.
• In vivo experimental rigour — use it for randomisation, blinding, exclusions and the tumour volume formula.
• Image quantification integrity — use it when a representative image is standing in for a quantification.
• Model-to-human relevance — use it when an immune or microenvironment claim may exceed the host used.