A custom AI reviewer for multiple testing and enrichment in omics papers

PerfectPaper’s multiple testing agent reads an omics manuscript for the correction failures that draw methods objections: nominal p-values reported as significant without false discovery rate control, gene set enrichment run against the wrong background universe, ranking metrics left unstated, and pathway results driven by a handful of genes. Copy the brief below into a custom agent slot before running your review.

Genome-wide experiments test tens of thousands of hypotheses at once. Most reviewer objections in this literature are not about whether the statistics are sophisticated, but about whether the multiplicity was handled at all and whether the enrichment comparison was fair.

When to use this agent

• Your manuscript reports differentially expressed genes, peaks, regions or proteins
• You ran gene set enrichment, over-representation analysis, or pathway analysis
• You performed a CRISPR, drug or reporter screen with a hit list
• Your results section reports p-values for a large number of features
• A previous submission drew a reviewer comment about correction or significance thresholds
• Your enrichment figure is a bar chart of adjusted p-values with no background universe named anywhere in the manuscript
• A single pathway or hallmark term carries the biological interpretation in your discussion
• Your threshold appears in a figure legend but not in the methods

How often multiple testing and enrichment go wrong in published omics

Enrichment reporting fails in the large majority of published over-representation analyses, and the failure is usually the background gene list. Wijesooriya, Jadaan, Perera, Kaur and Ziemann catalogued 1,630 functional enrichment analyses from articles published in 2019 and examined 235 of them, drawn from 186 articles, in methodological detail. They reported in “Urgent need for consistent standards in functional enrichment analysis” (PLOS Computational Biology 2022;18(3):e1009935) that “95% of analyses using over-representation tests did not implement an appropriate background gene list or did not describe this in the methods” — in only 8 of 197 cases, 4.1%, was an appropriate background described. The same survey found that “failure to perform p-value correction for multiple tests was identified in 43% of analyses,” and that just 15% of the analyses examined were clear of every major issue the authors looked for.

Those two numbers describe the whole review problem. A manuscript can be methodologically ordinary in every other respect and still carry the single most common defect in the genomics literature, because the background universe is a parameter most software supplies by default and most methods sections never mention.

The arithmetic underneath is simple enough to state in one sentence. Testing 20,000 genes at a nominal threshold of p < 0.05 yields about 1,000 features expected by chance alone under the complete null, so a paper reporting 800 “significant” genes on nominal p has reported a list smaller than its own expected false positive count. The general mechanism is set out in what is multiple testing correction, and the objection it produces at review is covered in reviewer says you did not correct for multiple comparisons.

What “corrected” has to mean in a genome-scale results section

A corrected threshold names a method, a quantity being controlled and a level, and a results section that says only “adjusted p < 0.05” has named one of the three. The agent reads for all three, because the same numeral means different things under different procedures.

Benjamini and Hochberg introduced the false discovery rate in “Controlling the false discovery rate: a practical and powerful approach to multiple testing” (Journal of the Royal Statistical Society Series B 1995;57:289–300); Storey and Tibshirani extended it for genome-scale work in “Statistical significance for genomewide studies” (PNAS 2003;100:9440–9445); Ignatiadis, Klaus, Zaugg and Huber added covariate weighting in “Data-driven hypothesis weighting increases detection power in genome-scale multiple testing” (Nature Methods 2016;13:577–580). The distinction between controlling a proportion and controlling a probability is the one authors most often blur in a discussion paragraph, and it is set out in FDR versus family-wise error.

The agent does not prefer one of these. It reports which one was used, whether the same one was used everywhere, and whether the reader can tell.

Two conditions attach to that table that a methods paragraph rarely states. Benjamini-Hochberg’s FDR guarantee holds under independence or under positive regression dependency (PRDS) among the test statistics — the condition Benjamini and Yekutieli proved sufficient in “The control of the false discovery rate in multiple testing under dependency” (Annals of Statistics 2001;29(4):1165–1188); it is not unconditional. Most differential expression designs satisfy PRDS well enough in practice, but strongly negatively correlated features — competing isoforms sharing a fixed transcriptional budget, or compositional data such as microbiome relative abundances — can violate it, and the conservative remedy is the Benjamini-Yekutieli procedure, which divides the threshold by the harmonic sum of 1/i instead of assuming independence. Where only a family-wise guarantee is needed rather than an FDR, Holm’s step-down procedure (Holm, “A simple sequentially rejective multiple test procedure,” Scandinavian Journal of Statistics 1979;6(2):65–70) controls the same error rate as Bonferroni while rejecting at least as many hypotheses, so a manuscript pairing plain Bonferroni with a small, pre-specified confirmatory set has given up power for no added rigor.

Common questions about multiple testing correction

Can an AI reviewer catch it if I forgot to correct for multiple testing in my RNA-seq paper?

Yes. The agent reads every reported feature list against its stated threshold, and if a p-value cutoff is nominal rather than corrected, it names that directly rather than assuming the reported number is already controlled. It also recomputes, from the numbers given, how many features would be expected by chance alone at that threshold.

Will an AI reviewer tell me which FDR threshold my omics paper should use?

No, and neither will a good referee. There is no single correct threshold — 0.05 and 0.1 are both defensible, depending on the design and what follows the paper. What the agent checks is that whichever threshold you choose is stated once, applied the same way in every figure and results sentence, and never blurred into a nominal p-value.

My reviewer wrote “please correct for multiple comparisons” — what exactly do they want me to change?

Three things, stated together: the method (Benjamini-Hochberg, Storey q, Bonferroni or permutation), the level it controls, and the resulting count of features that pass. A revision that adds the word “corrected” without naming the method and the new count answers the sentence in the review letter, not the statistical objection underneath it.

The agent brief

The multiple testing and enrichment brief below is the full agent definition — paste it into a custom agent slot before running a review. Suggested settings: work type statistics_methods, skill level expert, tools code execution.

Name: Multiple testing and enrichment discipline

You are reviewing a genome-scale experiment for multiplicity and enrichment
rigour. Read the methods and results together; a correction described in
methods but contradicted by the results is the finding, not the correction.

Check, in this order:

1. Correction. For every set of features called significant, identify whether
   the threshold is a nominal p-value or a corrected one. Name the method
   (Benjamini-Hochberg, Storey q, Bonferroni, permutation) and the threshold.
   If a gene, peak or protein list is presented as significant on nominal p,
   say so and state how many features would survive correction if that can be
   determined from what is reported.

2. Consistency. Compare the threshold stated in methods with the thresholds
   used in each figure, table and results sentence. Report any mismatch,
   including thresholds that loosen for one analysis without explanation.

3. Enrichment background. For every enrichment or over-representation result,
   identify the background universe. The correct background is the set of
   features that could have been detected in this experiment, not the whole
   genome or the whole annotation database. If the background is unstated,
   say so and explain why the reported enrichment cannot be interpreted
   without it. If the background is all annotated genes while the experiment
   detected a subset, state that the enrichment is likely inflated.

4. Ranking metric. For GSEA or any rank-based method, identify the metric used
   to rank features. If unstated, report it. If the metric is a raw fold
   change on low-count features, note the instability.

5. Driver genes. Where an enrichment term is highlighted in the text, determine
   how many features drive it. A pathway conclusion resting on two or three
   genes should be reported as such.

6. Selective reporting. Note any analysis where the number of tests performed
   is not recoverable from the manuscript, including thresholds that appear
   only in a figure legend.

Use code execution to recompute counts and proportions from reported numbers
where the manuscript gives enough to do so. Do not assume values that are not
present. Where you cannot determine something, say what is missing and what
the author would need to add.

Report each issue with the exact location and a specific remedy. Do not
comment on grammar, formatting or citation style.

The ordering in the brief is load-bearing. Correction is checked before enrichment because an enrichment result computed on an uncorrected gene list inherits every false positive in that list, and reporting the background problem first would leave the author fixing the second-order defect while the first-order one stands. The closing instruction to ignore grammar and citation style exists for the same reason it exists in every well-scoped brief: an agent given a broad remit returns a broad report, and the multiplicity findings sink into it. How to write a custom reviewer agent covers the general form, and custom AI reviewer agents covers how briefs are slotted into a review.

What this agent catches

The multiple testing agent catches nine recurring failure patterns in omics manuscripts, listed here with the reason each draws a referee’s objection.

Why the background universe changes the answer

The background universe sets the null expectation for every over-representation p-value, so changing it changes the result without changing a single measurement. An over-representation test asks how surprising the overlap between your feature list and an annotated set is, given a universe of features that could have appeared in the list. Supply a universe that includes genes your assay never had a chance to detect, and the test compares your list against a population it was never drawn from.

The arithmetic is worth doing once, because it is small enough to check by hand. Take a differential expression list of 400 genes, an annotated pathway of 200 genes, and a whole-genome background of 20,000 genes. The expected overlap under the null is 400 × 200 / 20,000 = 4 genes. Observe 12, and the fold enrichment is 3.0.

Now use the correct background. The experiment was a bulk RNA-seq study of one tissue, and 12,000 genes passed the detection threshold; 150 of the pathway’s 200 genes were among them, and the other 50 are not expressed in this tissue at all. The expected overlap becomes 400 × 150 / 12,000 = 5 genes, and the same observed 12 gives a fold enrichment of 2.4. The point estimate has fallen by 20% and the p-value has moved with it, from one set of numbers the author never chose to another.

That is the mild case. The inflation grows with the gap between the annotation and the assay, and it grows fastest for tissue-restricted pathways — neuronal, immune and secretory sets are systematically over-called in experiments on tissues that express only part of them. It compounds when the feature list itself was produced without correction, because the numerator is then also wrong.

Two further background errors appear in submitted work. Using all genes in the annotation database rather than all genes in the platform is the same defect for arrays and targeted panels. Using the set of genes that passed a fold-change filter, rather than the set that could have passed it, treats the filter as if it were the assay.

Length and abundance bias sits alongside the background problem in RNA-seq specifically. Young, Wakefield, Smyth and Oshlack showed in “Gene ontology analysis for RNA-seq: accounting for selection bias” (Genome Biology 2010;11(2):R14) that longer transcripts accrue more reads and are therefore more likely to be called differentially expressed at any given power, which biases downstream gene ontology results toward categories enriched for long genes. Their remedy, implemented in goseq, weights each gene by its probability of selection rather than treating all genes as equally detectable.

Questions about enrichment background gene lists

Can an AI reviewer tell whether my enrichment background universe is the right one?

Yes — it checks whether a background was stated at all, and whether the one stated matches what your assay could detect. The correct background is your expressed or detected feature set, not every gene in the annotation database. Wijesooriya and colleagues (PLOS Computational Biology 2022;18(3):e1009935) found 95% of published over-representation analyses got this wrong or left it undescribed.

Reviewer 2 asked what background I used for my enrichment analysis and I honestly don’t know — what do I check?

Open the function call in your script, not the paper. Tools like clusterProfiler, topGO and DAVID default to every annotated gene in the organism unless a universe or background argument was set explicitly. If nothing was passed, that default is almost certainly what ran, and for a tissue- or assay-restricted experiment it is very likely too broad.

What background gene list should I use for GO or pathway enrichment — all annotated genes, or just my differentially expressed genes?

Neither. Use every gene your assay could plausibly have detected — commonly, genes passing a minimum expression or count filter in your dataset — not the whole annotated genome and not only the genes that already passed your significance filter. The second choice is the more common manuscript error, and it inflates enrichment by shrinking the null’s expected overlap.

Worked example: the pathway that rests on three genes

A hallmark term reported as significantly enriched can be carried by three genes out of two hundred, and nothing in the standard enrichment figure reveals which case you are looking at. Consider a manuscript whose discussion opens on oxidative phosphorylation, citing an adjusted enrichment p-value of 0.004 against the MSigDB hallmark collection — 50 gene sets described by Liberzon, Birger, Thorvaldsdóttir, Ghandi and colleagues in “The Molecular Signatures Database (MSigDB) hallmark gene set collection” (Cell Systems 2015;1(6):417–425).

Read the leading edge and the picture changes. The hallmark set contains roughly 200 genes; the differential list overlaps it at 12; the GSEA leading edge — the subset of genes contributing to the peak enrichment score — contains 5; and of those, 3 are nuclear-encoded complex I subunits, a tightly co-regulated module whose members move together in nearly every dataset. The term is real, the p-value is real, and the biological claim in the discussion is about mitochondrial reprogramming across the pathway.

The agent reports this as a driver-gene finding with the count attached: the term is supported by 5 leading-edge genes, 3 of which are co-regulated, and the discussion states a pathway-level conclusion. That is a specific, answerable comment. “The enrichment analysis is overinterpreted” is not, and it is what the author receives from a reviewer who noticed the same thing without the numbers. The same discipline applies to any claim whose apparent strength comes from an aggregate rather than from the measurements underneath it, which is the subject of the overclaim check agent.

Competitive tests, self-contained tests, and inter-gene correlation

A gene set test answers one of two different questions, and manuscripts routinely report the answer to one while making the claim that belongs to the other. Goeman and Bühlmann named the distinction in “Analyzing gene expression data in terms of gene sets: methodological issues” (Bioinformatics 2007;23(8):980–987). A self-contained test asks whether the genes in the set are associated with the phenotype. A competitive test asks whether the genes in the set are more associated with the phenotype than the genes outside it. Over-representation analysis and GSEA are competitive; the background universe is exactly the “outside” the competitive test needs, which is why an unstated background makes a competitive p-value uninterpretable rather than merely imprecise.

Goeman and Bühlmann’s second point is the one that survives least often into methods sections: the sampling unit is the subject, not the gene. Tests that build their null by sampling genes treat genes as independent draws, and genes in a pathway are correlated by construction — that is what makes them a pathway.

Wu and Smyth quantified the consequence in “Camera: a competitive gene set test accounting for inter-gene correlation” (Nucleic Acids Research 2012;40(17):e133). Even small average correlations between genes in a set inflate the variance of the set-level statistic, and a competitive test that assumes independence returns p-values that are too small: with a set of 100 genes at an average inter-gene correlation of 0.05, Wu and Smyth’s simulations drove the type I error rate of existing competitive tests to 0.28 at a nominal 0.01. camera estimates a variance inflation factor from the data and applies it. The agent does not require any particular method, but it reports when a competitive test is described as if it were self-contained, and when a permutation scheme shuffled genes where it should have shuffled sample labels.

What the ranking metric decides in GSEA

The ranking metric is an unconstrained parameter that changes the output of GSEA more than most other choices in the pipeline, and Wijesooriya and colleagues found it — along with the gene weighting parameter and the permutation type — left unstated in more than half of the GSEA analyses they examined. Subramanian, Tamayo, Mootha, Mukherjee, Ebert and colleagues describe the method in “Gene set enrichment analysis: a knowledge-based approach for interpreting genome-wide expression profiles” (PNAS 2005;102(43):15545–15550); the algorithm walks a ranked list, and the ranking is the input.

The metrics in circulation behave differently:

• Raw log fold change. Unstable for low-count features, where a change from 2 reads to 8 reads is a two-log-fold change with almost no evidence behind it. Low-abundance genes crowd the extremes of the ranked list.
• Shrunken log fold change. DESeq2’s shrunken log fold change, described by Love, Huber and Anders in “Moderated estimation of fold change and dispersion for RNA-seq data with DESeq2” (Genome Biology 2014;15(12):550) and exposed today as lfcShrink, pulls low-information estimates toward zero and produces a ranking that reflects evidence rather than arithmetic.
• Signed −log10 p, or the Wald or moderated t statistic. Combines direction and evidence, and is the ranking most GSEA implementations expect. The edgeR framework of Robinson, McCarthy and Smyth (Bioinformatics 2010;26(1):139–140) and limma supply the equivalents.
• Signal-to-noise ratio. The original GSEA default for two-class designs, and unstable at very small n per group.

When the metric is unstated, a reader cannot reproduce the figure from the deposited counts, which is a reproducibility finding rather than a statistical one. When the metric is a raw fold change and the manuscript has not filtered low-count features, the agent reports the instability with the location. Whether the resulting differences are large enough to matter biologically is a separate question, and the objection it produces is at reviewer says the effect size is not meaningful.

Permutation type is the parameter that travels with the metric and is just as easy to leave unstated. Phenotype permutation — shuffling sample labels and recomputing the enrichment score each time — preserves the correlation between genes, and GSEA’s own documentation states the threshold in the other direction: as a general recommendation, with fewer than seven samples per phenotype it advises gene-set rather than phenotype permutation. Gene-set permutation, the fallback for smaller studies, shuffles gene labels instead, which treats genes as independent draws — the same assumption Goeman and Bühlmann flagged above — and returns an anti-conservative p-value for exactly the correlated sets GSEA exists to detect. A methods section naming the ranking metric but not the permutation type has closed one gap and left the other open.

Questions about GSEA ranking metrics

Which GSEA ranking metric should I use — fold change, signal-to-noise, or the moderated t-statistic?

A signed statistic that combines direction with evidence is the safer default: DESeq2’s shrunken log fold change, or the moderated t-statistic from limma or edgeR. Raw log fold change ranks low-count, high-variance features at the list’s extremes on noise alone, and GSEA’s documentation puts three samples per phenotype as the bare minimum at which its default signal-to-noise metric can be computed at all — a floor, not a comfortable n.

Why did my significant pathways change when I switched from raw fold change to DESeq2’s shrunken fold change for ranking?

Because the raw metric was ranking genes on noise. A gene with almost no reads can show an enormous, unreliable fold change, and unshrunk ranking sends it to an extreme of the list regardless. Shrinkage pulls low-information estimates toward zero in proportion to their uncertainty, so the leading edge shifts toward genes supported by real evidence.

Screens and peak calls: where the correction already happened

Screen and peak-calling pipelines apply their own multiplicity correction, and the common failure is a second, undocumented threshold applied on top of it. The agent reads for the layer, not just the number.

CRISPR and drug screens. MAGeCK, described by Li, Xu, Xiao, Cong and colleagues in “MAGeCK enables robust identification of essential genes from genome-scale CRISPR/Cas9 knockout screens” (Genome Biology 2014;15(12):554), aggregates guide-level evidence to a gene-level score and reports an FDR. A manuscript that then selects hits by “FDR < 0.25 and at least three guides in the same direction and expressed in our cell line” has applied a composite filter whose overall error rate is not the reported FDR. That is reportable and usually defensible — it just has to be stated as the selection rule it is. The number of guides per gene and the behaviour of the non-targeting controls belong in the same methods paragraph.

ChIP-seq and ATAC-seq. MACS2 thresholds peaks on a q-value by default, at 0.05, so a peak set is already FDR-controlled before any downstream test runs. Papers that then report “peaks with p < 0.001” have moved to a less stringent, uncorrected criterion while sounding more stringent. Where replicates exist, the irreproducible discovery rate framework used in ENCODE ChIP-seq pipelines assesses reproducibility across replicates rather than significance within one, and it answers a different question — a peak can be highly significant in one replicate and irreproducible across two.

Differential binding and differential accessibility. Testing regions that were themselves selected as peaks means the feature set was chosen using the data, and the background for any downstream enrichment is the peak universe, not the genome.

Where the number of tests stops being recoverable

The number of tests performed is the denominator of every correction in the paper, and a manuscript that does not let a reader recover it cannot be checked at all. The agent reports irrecoverability as a distinct finding, separate from an incorrect correction, because the remedies differ: one needs a recomputation, the other needs a sentence.

The specific patterns it flags are concrete. A threshold that appears only in a figure legend, and differs from the methods. A results sentence reporting “the 43 significant metabolites” where no filtering step in the methods produces 43 of anything. Subgroup analyses — by sex, by subtype, by timepoint — corrected within each subgroup, with conclusions drawn across all of them, so that the family used for correction is narrower than the family of claims. Analyses described in the past tense in a discussion paragraph but absent from the results. Supplementary tables whose row counts do not match the counts in the main text.

None of these is proof of anything. Each is a place where the number of tests and the number of claims have come apart, which is the structure a reviewer names when writing about selective reporting — the general mechanism is in what is p-hacking, and the broader family of methods objections in reviewer objections about statistics.

How to check your own manuscript before submitting

Checking multiplicity and enrichment in your own manuscript takes four passes over text you have already written, and none of them requires re-running an analysis. Do them in this order.

Grep the manuscript for every threshold. Search for p <, p-value, FDR, q <, adjusted and significant, and write each hit into a list with its location. Every hit should agree with the methods, and every one that does not is a finding you can fix rather than defend.

Write one sentence naming the background universe for every enrichment result. If you cannot write it without opening the analysis script, the reader cannot recover it either. State the number of features in the background, not just its description: “14,218 genes with non-zero counts in at least three samples” is checkable, “expressed genes” is not.

Count the leading edge for every term you name in the discussion. Any term carrying a biological claim needs its overlap size and its driver count in the supplement. A term supported by three genes is publishable; a pathway-level conclusion drawn from three genes is not.

Reconcile the feature counts across the abstract, results, figures and supplementary tables. These disagree more often than authors expect, and a mismatch between a stated count and a table row count is the first thing a careful referee checks.

Name the ranking metric, the software and the version. One methods sentence — metric, tool, version, gene set collection and its release — closes the reproducibility objection outright.

What reviewers write when multiplicity is mishandled

Reviewers rarely open with the phrase “multiple testing”. They write the specific version, and these are the comments that decide the paper.

“The gene list in Figure 3 is thresholded at p < 0.05; please state whether this is nominal or adjusted, and report how many genes survive FDR control.” “The methods state an FDR of 0.05 but Figure 4 uses 0.1; please explain the difference or apply one threshold throughout.” “The background gene set for the enrichment analysis is not stated; the reported enrichment cannot be evaluated without it.” “Enrichment appears to have been computed against all annotated genes rather than against genes detected in this tissue.” “The ranking metric used for GSEA is not given.” “The oxidative phosphorylation conclusion rests on five leading-edge genes; please report the leading edge for each term discussed.” “Please provide the number of tests performed for each analysis reported.” “Corrections were applied within each subgroup but the conclusions are drawn across subgroups.” “The significance threshold appears only in the legend of Figure 2.”

The hardest of these to answer after submission is the background one, because the honest reply usually requires re-running every enrichment analysis in the paper with a different universe and finding out which terms survive. Doing it before submission takes an afternoon; doing it in revision means rewriting a discussion around whatever remains. How to write a response to reviewers covers the reply that works, which is the recomputed table rather than a paragraph explaining why the original background was defensible.

When the obvious remedy is unavailable

When nothing survives FDR correction, the remedy is a narrowed claim with the correction reported, not a looser threshold introduced without comment. That situation is common and it is survivable, and there are five moves available that do not involve pretending the correction did not happen.

Report the ranked list as hypothesis-generating, with the FDR stated. A paper that says “no gene reached an FDR of 0.05; the ten highest-ranked genes are given in Table 2 as candidates for follow-up” is making an honest, publishable statement. A paper that quietly switches to nominal p is making a claim it cannot support.

Pre-specify a looser threshold and say that you did. An FDR of 0.1 is used routinely in exploratory omics. It becomes a defect only when it appears without justification, or when it appears for one analysis and 0.05 for another.

Increase power without increasing the error rate. Independent filtering of low-count features and covariate weighting under IHW both raise the number of discoveries at a fixed FDR by spending the testing budget where evidence exists, and both are declared in the methods rather than applied silently.

Validate a small number of candidates by an independent method. Confirming three or four candidates by qPCR, targeted proteomics, immunostaining or a second cohort converts a screen result into a specific claim about specific features, and it does not depend on any threshold. See orthogonal validation.

State the direction and magnitude of what you found. Effect sizes with confidence intervals survive when thresholded lists do not, and they let a reader judge the result rather than accept a dichotomy. A post-hoc power calculation is not one of the moves available — reviewer asks for post-hoc power explains why it adds nothing to a completed analysis, and why power is an argument to make before the next experiment rather than after this one.

The limitation paragraph does the remaining work, and it should name the correction, the number of tests, and what the result can support. How to write a limitations section sets out the form.

How to report multiplicity so a reviewer can check it

Report multiplicity by naming, for each analysis, the number of tests, the correction method, the threshold, and the resulting count — in that order, in the methods, once. Four clauses close most of the objections above.

The enrichment analyses need one further sentence each: the gene set collection and its release version, the background universe with its size, the test used, and whether that test is competitive or self-contained. For rank-based methods, add the ranking metric and the permutation scheme, including what was permuted.

The supplement carries what prose cannot. Deposit the full results table with unadjusted and adjusted p-values for every feature tested, not only for the features that passed — a table of survivors makes the correction unverifiable, since the denominator is missing. For each enrichment term reported, deposit the overlap size, the set size, the background size and the leading-edge gene identifiers. Accession numbers for the underlying data belong in the same pass, which is the subject of the data availability and accessions agent.

What this agent does not do

The multiple testing agent reviews what the manuscript reports; it does not reanalyse data it has not been given. It reads the methods, results, figure legends and supplementary text, recomputes counts and proportions from numbers the manuscript states, and reports where those numbers do not reconcile. It does not open a counts matrix, refit a model, or re-run an enrichment analysis against a corrected background.

It also does not prescribe. It reports whether a threshold was stated, applied consistently and described accurately, rather than declaring an FDR of 0.05 correct and 0.1 incorrect. Choosing a threshold is a study-specific decision that depends on the design and on what follows the paper, and an agent that made it for you would be returning an instruction where a finding belongs.

How this differs from the built-in review

PerfectPaper’s standing team already includes specialists in statistical models and regression, confidence-interval and p-value coherence, and causal inference. Those read the paper as a statistician would. None of them is briefed on the conventions of genome-scale multiplicity or on what makes an enrichment background correct, because those are assay-specific rather than general statistical questions. That gap is what this agent fills.

PerfectPaper reads the threshold stated in the methods against every figure legend and every results sentence, and reports each mismatch with its exact location rather than a general warning about multiplicity.

Related agents in this cluster: batch effects and replicate structure and data availability and accessions. The full set is on AI peer review for genomics. An uncorrected batch effect and an uncorrected p-value inflate the same result, and what is a batch effect explains why the two findings usually arrive together. Where the effective sample size is smaller than the feature count suggests, the pseudoreplication agent is the one to run first, and single-cell manuscripts should start at my p-value shrinks with more cells.

Review my manuscript

Related agents

Four companion 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.

• 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.
• Metabolomics and flux reporting — use it for identification confidence, normalisation, and natural-abundance correction.
• Variant and gene nomenclature — use it when the paper reports specific variants that will be curated downstream.

Frequently asked questions

Is n=3 biological replicates enough for RNA-seq, or will a reviewer reject it on multiple-testing grounds?

n=3 is the field’s conventional working minimum, not a floor either DESeq2 or edgeR enforces — both will run on less, with wider intervals. Replicate count and multiple-testing correction are separate defects: a well-corrected n=3 analysis can be entirely sound, it will simply return fewer discoveries at a given FDR than the same design at n=6-8. A reviewer citing “sample size” is usually raising power, not correction.

Can an AI reviewer replace a statistician on a genome-scale analysis?

No. The agent reports what the manuscript does and does not state about multiplicity, and recomputes what the reported numbers allow. Study-specific decisions — which family of tests to correct over, which threshold suits the design, whether a covariate is admissible under IHW — still need a statistician who knows the study.

Can an AI reviewer run these checks on a proteomics or metabolomics manuscript?

Yes, for the multiplicity checks, which apply to any high-dimensional assay: nominal versus corrected thresholds, consistency across figures, background universes, and recoverable test counts all transfer unchanged. For metabolomics-specific reporting — identification confidence, normalisation, natural-abundance correction — pair it with the metabolomics reporting agent.

Can an AI reviewer check enrichment and FDR before I submit, rather than after review?

Yes, and before submission is when findings are cheapest to act on. You can run the same passes yourself: search the manuscript for every p <, FDR, q < and “significant”, and confirm each agrees with the methods. Name the background universe and its size for every enrichment result, and count the leading-edge genes behind every term you discuss.