AI peer review for genomics and epigenetics papers

PerfectPaper reviews genomics, epigenetics and metabolism manuscripts with a standing team of specialist agents, and lets you add up to three custom agents of your own for the conventions that team does not cover. This page contains five prebuilt agent briefs for omics work, each a text brief you copy into the agent field.

The five cover multiple testing and enrichment, data availability and accessions, batch effects and replicate structure, metabolomics reporting, and variant nomenclature. The standing team already covers statistics, methodology, figures, tables, citations and prior work. The five below exist because omics manuscripts fail for reasons that are specific to the assay rather than to statistics in general, and a generalist reviewer does not know to look for them.

The five agents

The five prebuilt agents for omics manuscripts are multiple testing and enrichment discipline, data availability and accessions, batch effects and replicate structure, metabolomics and flux reporting, and variant and gene nomenclature — each one a ready-made brief you paste into the agent field unchanged.

Multiple testing and enrichment discipline — nominal p-values reported without correction, enrichment run against the wrong background universe, ranking metric unstated. A recurring reason an omics paper draws a methods objection.

Data availability and accessions — missing GEO, SRA, dbGaP or MetaboLights accessions, unstated genome build, absent tool versions and parameters. A recurring reason an omics paper is returned before review.

Batch effects and replicate structure — batch confounded with condition, batch never modelled, technical replicates presented as biological.

Metabolomics and flux reporting — normalisation and internal standards unstated, no natural-abundance correction on isotope tracing, identification confidence level never given.

Variant and gene nomenclature — non-HGVS variant strings, unversioned transcript references, deprecated gene symbols.

Why omics manuscripts fail for assay-specific reasons

Omics manuscripts fail at review on conventions that are measurable, published, and almost never taught alongside general statistics. Each of the five agents above exists because someone has counted how often the corresponding item goes wrong, and the counts are large.

Enrichment reporting is the clearest case. Wijesooriya, Jadaan, Perera, Kaur and Ziemann screened 186 open-access articles published in 2019, describing 235 enrichment analyses, and 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 it in the methods, and that failure to correct for multiple tests appeared in 43% of analyses. An extension of the survey to 1,630 analyses found these problems were not associated with journal or article-level bibliometrics. Only 4.1% of over-representation analyses — 8 of 197 — described an appropriate background. This failure is specific to over-representation testing (ORA), which requires an explicit background gene list; rank-based methods such as GSEA use the whole ranked gene list and have no comparable background-list step, though an unstated or arbitrarily chosen ranking metric is their equivalent failure mode. A custom agent briefed on “the background gene list” is briefed on an ORA-shaped problem and will not catch a GSEA manuscript’s version of it unless the brief names both.

Data availability is the second. Gabelica, Bojčić and Puljak analysed every article published in January 2019 across 333 open-access BioMed Central journals — 3,556 in all — and reported in “Many researchers were not compliant with their published data sharing statement” (Journal of Clinical Epidemiology 2022;150:33–41) that among the 1,792 whose statements indicated willingness to share, 1,669 (93%) either did not respond or declined, and only 123 (6.8%) supplied the data. “Available from the corresponding author on reasonable request” is a measured non-answer, not a compliant statement.

Batch structure is the third. Leek and colleagues examined nine public datasets in “Tackling the widespread and critical impact of batch effects in high-throughput data” (Nature Reviews Genetics 2010;11:733–739) and found batch effects in all nine, with 32.1% to 99.5% of measured features showing a statistically significant association with processing date alone. Processing date is only one batch variable among many, and it is rarely the one a methods section names. Correction methods — ComBat’s empirical-Bayes adjustment (Johnson, Li and Rabinovic, Biostatistics 2007;8(1):118–127) and removal-of-unwanted-variation (RUV) approaches among them — can only separate batch from biology when the two are not fully aliased; when every sample in one condition was also processed in a single batch, no algorithm can partition the two from the expression data alone, because nothing in the data isolates one variable from the other. In that case the fix is disclosure of the confound, not a correction step.

None of these is a statistics failure in the ordinary sense. Each is a convention of the assay, which is exactly the gap a custom agent is for.

Common questions about omics review conventions

Is n=3 enough for RNA-seq?

Three is the historical default, but it is thin. Schurch and colleagues generated 48 biological replicates per condition in yeast (RNA 2016;22:839–851) and found three replicates reliably detect only the largest fold changes; they recommend at least six per condition, rising to twelve when the goal is identifying differentially expressed genes across all fold-change sizes. A reviewer citing n=3 is citing this study.

Why did my reviewer say my enrichment analysis used the wrong background gene list?

Most enrichment tools default to every annotated gene in the genome as the background, not the genes your assay could have detected. Wijesooriya and colleagues found 95% of over-representation analyses in a 186-article survey did not specify or use an appropriate background (PLOS Computational Biology 2022;18(3):e1009935); only 4.1% did. The correct background is the set of genes that passed detection in your tissue, not the full annotation.

Reviewer 2 says my sample size is too small — what do I say?

Cite the design literature for your assay rather than defending the number itself. For bulk RNA-seq, Schurch and colleagues (RNA 2016;22:839–851) is the standard reference recommending six to twelve replicates per condition. For single-cell data, note that thousands of cells from one donor is still n=1 for any donor-level claim, so the response should address biological replication, not cell count, and should offer a revision plan rather than a rebuttal.

What each agent reads and what it reports

Each of the five agents reads a defined region of the manuscript and returns findings with locations rather than general warnings. The table below states what each one is briefed on, the settings the brief is written for, and the single failure it catches most often.

The work type and skill level are not decoration. Work type routes the agent into the review contract that matches its job, and skill level sets the register of the finding — a graduate-level verification agent reports that a dbGaP study accession such as phs000178 is cited without its version and participant-set suffix; an expert-level statistics agent reports that the family used for correction is narrower than the family of claims.

How custom agents work

A review takes up to three custom agents alongside the standing team. Each takes a name, a set of instructions up to 6,000 characters, a work type, a skill level, and an optional set of tools — web search, scholarly search, code execution, and vision for figure reading. The instructions are yours; PerfectPaper supplies the review contract and output format around them, so a brief describes what to look for rather than how to format findings.

Three slots is a real constraint, which is why these briefs are narrow. An agent asked to check everything checks nothing in particular.

Every brief in this cluster fits inside 6,000 characters with room to add your own items, and the additions worth making are the conventions of your own subfield: a required control your field expects in every figure, a reporting standard your target journal enforces, a mistake a reviewer caught in your last submission. How to write a custom reviewer agent sets out the structure that works and the instructions that make an agent worse — chiefly instructions that invite the agent to adjudicate rather than report. PerfectPaper can also propose an agent for your specific paper from its title and abstract, which you then edit before running. The full set of prebuilt briefs across fields is at custom reviewer agents.

Tools are worth giving sparingly. Two of the five briefs here use web search, to confirm a current HGNC symbol or a repository’s accession format. The other three are more reliable without any tool, because a checking task that reads only the manuscript cannot be talked out of a finding by something it retrieved.

Questions about custom agents for omics review

How many custom agents can I add to one genomics review?

Three. Each takes a name, instructions up to 6,000 characters, a work type, a skill level and an optional tool selection, and all three run alongside PerfectPaper’s standing specialist team rather than replacing it. Three narrow agents outperform one broad agent asked to check everything, which is why every brief in this cluster is written narrow rather than general.

Can an AI reviewer for genomics search the literature?

Yes, if you give it the web search or scholarly search tool. Two of the five briefs in this cluster use search — to confirm a current HGNC gene symbol or a repository’s accession format — while the other three are more reliable without any tool, because a checking task that reads only the manuscript cannot be talked out of a finding by something it retrieved.

Choosing among them

Which of the five agents to run depends on the assay and the manuscript’s stage — pre-submission or revision — not on severity alone; a smaller scientific problem caught earlier can matter more than a larger one caught later. For a sequencing paper with a differential expression result, the multiple testing agent earns its slot first, then batch effects. For a paper being prepared for submission rather than revised, data availability moves first, because it is the check that prevents a return without review. For metabolomics, take the metabolomics agent and one of the other two. Variant nomenclature is worth a slot only when the paper reports specific variants, in which case it is fast and unambiguous.

The ordering rule underneath those recommendations is stage, not severity. A missing accession is a smaller scientific problem than a confounded batch design and a larger practical one, because it is caught by editorial staff at the technical check, before the handling editor reads the science. That check runs from a checklist rather than from judgement, and a blank in the box returns the file with the status unchanged — the mechanism behind most cases of being desk rejected without review on an omics paper with no scientific defect at all. Pre-submission, spend a slot on the gate that opens first.

At revision the ordering inverts. A reviewer has already read the paper, the editorial check has already passed, and the findings that matter are the ones that decide whether the claim survives. That is when the multiple testing and batch agents earn both remaining slots.

Worked example: three slots on a bulk RNA-seq manuscript

Consider a manuscript reporting 2,431 differentially expressed genes between 12 tumour and 12 adjacent-normal samples, with a hallmark enrichment figure and a discussion built on oxidative phosphorylation. Three slots, allocated in order.

Slot one, multiple testing. The agent reads every threshold in the paper against the methods. Testing 20,000 genes at a nominal p < 0.05 yields about 1,000 features expected by chance under the complete null, so the first question is whether the 2,431 are adjusted, and the second is what “adjusted” names — Benjamini-Hochberg controls the expected proportion of false positives among rejections, Bonferroni controls the probability of one or more, and the same numeral means different things under each. Benjamini-Hochberg’s guarantee was proved under independence and under positive regression dependency (Benjamini and Yekutieli, Annals of Statistics 2001;29(4):1165–1188), and overlapping gene sets are the positively dependent case it already covers — two pathways sharing genes draw on the same evidence, so their statistics move together rather than apart. The case that falls outside the proof is a dependency structure nobody has characterised, and the more conservative Benjamini-Yekutieli modification from that same paper is the one that holds under arbitrary dependence at the cost of power. The agent does not adjudicate which was right; it flags a methods section that names a correction without naming the assumption the correction rests on. The mechanism is set out at what is multiple testing correction. The agent then asks for the background universe: 12,000 genes passed detection in this tissue, not the full annotation the enrichment tool used by default, and swapping one for the other moves both the fold enrichment and its p-value without changing a single measurement.

Slot two, batch effects. The methods give two sequencing dates. The tumours were accrued first and sequenced in March; the normals came from a second consenting wave in June. The volcano plot is clean, the dispersion estimates are unremarkable, and the principal component plot separates the groups sharply along PC1 — which is exactly what a March-versus-June effect looks like, because in this dataset the two variables are the same column. Nothing in the figures distinguishes them, and no batch-correction algorithm can rescue this design after the fact — ComBat or RUV can adjust for a batch that varies somewhat independently of condition, but here the two variables share every sample, so there is no contrast left to estimate one apart from the other. The tell is a methods sentence and a supplementary table of run identifiers the results never mention, and the agent reads those two artefacts against each other. See what is a batch effect for the general mechanism.

Slot three, data availability. The RNA-seq has a GSE accession. The ATAC-seq generated for Figure 4 does not, the genome build is given as “the human genome” rather than GRCh38.p14, and three tools are named without versions. Each of those is a one-sentence fix before submission and a revision cycle after it.

The paper is not badly done. It carries three findings that a generalist review would not raise, and all three are recoverable at the desk.

Common questions about batch effects and replicate structure

My PCA plot splits by batch, not by condition — is my experiment ruined?

Not necessarily, but it needs to be disclosed, not hidden. A split along a variable fully confounded with your comparison of interest looks identical in the plot to a split along the biology — Leek and colleagues found batch effects in nine of nine public datasets examined (Nature Reviews Genetics 2010;11:733–739). Publish the batch-by-condition cross-tabulation so reviewers see the confound rather than have to infer it from the figure.

Are technical replicates the same as biological replicates for statistics purposes?

No. A technical replicate measures the same biological sample twice and estimates measurement noise; a biological replicate measures independent samples and estimates the variance a p-value is meant to capture. Presenting technical replicates as biological ones understates true variance and inflates apparent significance — it is one of the three failure modes the batch effects and replicate structure agent is briefed to catch.

I’m not a statistician, and reviewer 2 just said my RNA-seq p-values need multiple testing correction — can you explain what that means and whether my results are still valid?

It means your p-values need adjusting because testing thousands of genes produces many false positives by chance alone: testing 20,000 genes at p<0.05 yields roughly 1,000 “significant” hits with no real effect behind them. Benjamini-Hochberg (controlling the expected false-positive proportion among hits) and Bonferroni (controlling the chance of any false positive at all) are the two standard corrections, and your results likely remain valid once the correct one is reported.

Worked example: three slots on a 13C tracing manuscript

A tracing manuscript reports that treatment raised M+2 citrate from 14% to 19% and describes the result as increased glucose oxidation through the TCA cycle. The allocation changes completely.

Metabolomics and flux reporting takes the first slot, and does most of the work. Carbon-13 occurs naturally at about 1.1%, so an unlabelled metabolite’s M+1 peak sits at roughly 1.1% × n relative to its monoisotopic peak: about 6.5% for six-carbon citrate and about 17% for sixteen-carbon palmitate. Without natural-abundance correction, part of every isotopologue above M+0 is natural abundance rather than tracer, and the correction shrinks apparent labelling unequally across metabolites of different carbon number — which is precisely the comparison a figure makes when it places citrate, malate and aspartate on one axis. The agent asks for the correction software by name, IsoCor or AccuCor, with its version and the elemental formula used.

The same agent reads the identification claims against the Metabolomics Standards Initiative scale of Sumner and colleagues (Metabolomics 2007;3:211–221), where a Level 1 identification requires two orthogonal properties matched to an authentic standard run in your own laboratory. Accurate mass alone does not qualify: Kind and Fiehn generated more than 1.6 million molecular formulae in the 0–500 Da range (BMC Bioinformatics 2006;7:234) and showed that isotopic abundance patterns remove more than 95% of false candidates that sub-1-ppm mass accuracy leaves standing. Level 2 — a putatively annotated compound matched to a public spectral library or literature data without an in-house authentic standard — is the level most papers actually support, and is a legitimate claim when labelled as such; the finding the agent raises is a Level 1 claim resting on Level 2 evidence, not the use of Level 2 itself.

Slot two goes to the overclaim check rather than to statistics. A higher M+2 fraction at one time point is consistent with faster entry of glucose carbon into citrate, and equally consistent with a smaller citrate pool relabelling faster at unchanged flux. Pool size, labelling and flux are three measurements, and the sentence that survives review names the measurement and the interpretation separately. That gap is the subject of the overclaim check agent.

Slot three goes to data availability, because a metabolomics deposit lands in MetaboLights or Metabolomics Workbench rather than GEO, those repositories are less familiar to authors who mostly submit sequencing data, and the technical check treats a missing accession identically either way.

Where the standing team stops and a custom agent starts

PerfectPaper’s standing team stops at general statistical and methodological review: it includes specialists in statistical models and regression, confidence-interval and p-value coherence, causal inference, figures, tables, numeric integrity, citation verification, prior work and journal fit, and none of them is briefed on assay-specific convention. Those specialists read the paper as a statistician and a methodologist would.

The boundary is easiest to see on a concrete item. A general statistics reviewer will notice that a p-value and a confidence interval disagree, and will not notice that an enrichment analysis was run against all annotated genes rather than the 12,000 detected in this tissue, because the second is not a statistical error at all — it is a wrong parameter supplied by software default and never mentioned in the methods. Likewise, a general reviewer will question an n of 3 and will not know that Schurch and colleagues, generating 48 biological replicates per condition in Saccharomyces cerevisiae (RNA 2016;22:839–851), recommended at least six per condition for RNA-seq, rising to at least twelve where differentially expressed genes at all fold changes need to be identified.

Single-cell work has the same shape with an extra trap. Thousands of cells from one donor is n = 1 for any claim about donors, and a p-value computed across cells shrinks toward zero as more cells are captured whether or not the biological effect exists — see my p-value shrinks with more cells. Hicks, Townes, Teng and Irizarry found, across the eight of 15 public scRNA-seq datasets that compared predefined biological groups (Biostatistics 2018;19:562–578), a standardised contingency coefficient between instrument run and biological condition ranging from 82.1% to 100%. The standard remedy is pseudobulk aggregation: summing or averaging counts to one value per biological sample before testing, which restores the sample as the unit of replication and removes the artificial statistical power that cell count otherwise supplies.

When the obvious remedy is unavailable

The remedy an omics finding implies is frequently impossible by the time the finding arrives, and the honest alternative is a narrowed claim with a stated direction, not a quieter version of the original one.

When nothing survives correction, report the ranked list as hypothesis-generating with the false discovery rate stated. “No gene reached an FDR of 0.05; the ten highest-ranked genes are given in Table 2 as candidates for follow-up” is publishable. A silent switch to nominal p is not.

When the batch confound cannot be undone — samples exhausted, cohort closed, consent lapsed — publish the condition-by-batch cross-tabulation. Reviewers respond very differently to a design they were shown than to one they discovered. Gilad and Mizrahi-Man’s reanalysis of Mouse ENCODE comparative expression data (F1000Research 2015;4:121) is the published cautionary case: flowcell and lane assignment was nearly completely confounded with species annotation, and after accounting for batch the data clustered by tissue rather than species, reversing the headline finding from the same reads.

When the data cannot be deposited publicly, consent that forbids release is a reason to use a controlled-access repository, not a reason to have no accession. dbGaP and EGA both issue an accession that appears in your paper while access remains governed by a data access committee, and the derived layer — summary statistics, count matrices, quality-control metrics — is usually shareable when the reads are not.

When a claim rests on a confounded or uncorrected result, validate a small number of candidates by an independent method. Confirming three or four features by qPCR, targeted proteomics or a second cohort processed in one batch converts a threshold-dependent list into a specific claim about specific features; see orthogonal validation. Then write the limitation with a direction rather than a caution, in the form set out at how to write a limitations section.

What these agents do not do

The five agents review what the manuscript reports; none of them reanalyses data it has not been given. They read methods, results, figure legends, tables and supplementary text, recompute counts and proportions from numbers the manuscript states, and report where those numbers do not reconcile. They do not open a counts matrix, refit a model, re-run an enrichment analysis against a corrected background, or confirm that an accession resolves.

They also do not prescribe. The multiple testing agent reports which correction was used, whether it was used consistently, and whether a reader can tell — it does not declare an FDR of 0.05 correct and 0.1 incorrect, because that choice depends on the design and on what follows the paper. The nomenclature agent checks HGVS syntax, which is checkable from the manuscript alone, and is instructed not to assert that a variant exists in any database, because a genuinely novel variant is frequently the point of the paper. HGVS syntax and ACMG/AMP pathogenicity classification (Richards and colleagues, Genetics in Medicine 2015;17:405–424) are different axes entirely: a variant can be written in flawless HGVS format and still be unclassified or scored a variant of uncertain significance, and the nomenclature agent checks only the former, never the latter.

Manuscripts are reviewed as .docx. The agents do not read LaTeX source.

Related

Preclinical cancer agents · Pseudoreplication · Ethics and data sharing · Study section lens · FDR versus family-wise error · Unpublished research handling

Review my manuscript

Frequently asked questions

Can AI peer review a genomics manuscript?

Yes, on the reporting and reasoning the manuscript contains. PerfectPaper’s standing specialist team reviews statistics, methodology, figures, tables, citations and prior work, and up to three custom agents add the omics-specific conventions the standing team is not briefed on: enrichment backgrounds, accessions and genome build, batch structure, metabolite identification level, and HGVS variant syntax.

Does AI peer review replace a human reviewer for genomics work?

No. The agents find absences and inconsistencies reliably and report them with locations, but they cannot tell you that the dataset you deposited is not the dataset you analysed, or that a pipeline parameter chosen two years ago was wrong. PerfectPaper is expert-led and AI-assisted: it prepares a manuscript for human review rather than standing in for one.

My paper got desk rejected and the email just said “data availability incomplete” — what does that actually mean and how do I fix it before I resubmit?

It usually means one accession is missing from a paper with several datasets — commonly, sequencing data carries a GEO number but a companion assay, such as ATAC-seq, does not. Editorial staff check this from a checklist before an editor reads the science, so the fix is mechanical: list every dataset generated, an accession for each, the genome build, and every tool version, then resubmit.