A custom AI reviewer for variant and gene nomenclature

PerfectPaper’s nomenclature agent reads a genomics manuscript for variant and gene naming errors: variant strings that do not follow HGVS syntax, transcript references given without a version, protein and coding descriptions that disagree, and gene symbols that have been superseded or that collide with an alias. Copy the brief below into a custom agent slot.

This is the narrowest agent in the set and the fastest to act on. Nomenclature errors are unambiguous, they propagate into databases, and they are trivially fixable once found.

When to use this agent

The nomenclature agent earns a slot on manuscripts that name specific sequence variants, and on essentially no other kind of paper. The trigger is a variant string in the text, a table or a figure — not the presence of sequencing.

• Your manuscript reports specific sequence variants
• Variants are described in text, a table and a figure, and were typed more than once
• The paper has co-authors who used different conventions
• You are reporting clinical or germline variants that will be curated downstream
• Gene symbols in the paper were chosen from older literature
• A variant list was assembled in a spreadsheet before being pasted into the manuscript
• The paper reuses variant calls generated on a different genome build from the one the analysis reports

The agent brief

The agent brief below configures a reviewer for HGVS syntax, versioned reference sequences, and current HGNC gene symbols; paste it into a custom agent slot. Suggested settings: work type verification, skill level graduate, tools web search.

Name: Variant and gene nomenclature

You are reviewing a genomics manuscript for sequence variant and gene symbol
nomenclature. Examine the text, every table, and every figure caption.

Check and report:

1. HGVS syntax. Every variant should follow HGVS form with an explicit
   reference sequence and a prefix indicating the coordinate system: c. for
   coding, g. for genomic, p. for protein, n. for non-coding, m. for
   mitochondrial. Report variants written without a prefix, written in legacy
   forms, or using a substitution arrow instead of the accepted syntax.

2. Reference sequence. Every coding variant should name the transcript it is
   numbered against, with a version suffix. Report transcripts given without a
   version, and report variants given with no reference sequence at all.
   Where different variants in the same gene use different transcripts,
   report the inconsistency.

3. Protein notation. Protein-level changes should use three-letter amino acid
   codes in HGVS form. Report one-letter forms, and report any protein
   description that is inconsistent with the coding description given
   alongside it, showing both.

4. Internal consistency. Compare every variant as it appears in the abstract,
   the text, each table and each figure. Report any variant whose position,
   reference allele or alternate allele differs between appearances. This is
   the highest-value check in this brief; report these first.

5. Gene symbols. Report gene symbols that are not current HGNC-approved
   symbols, including withdrawn symbols and aliases still common in older
   literature. Report symbols that are ambiguous between species where the
   species is not stated. Note that human symbols are conventionally
   upper-case and mouse symbols title-case, and report inconsistent usage.

6. Genome build. Where genomic coordinates are given, determine whether the
   build is stated. Coordinates without a build are uninterpretable; report
   them.

Use web search only to confirm current HGNC symbol status where a symbol looks
superseded. Do not attempt to validate that a specific variant exists in any
database, and do not assert clinical significance.

Report every instance with its exact location and the corrected form.

The last paragraph of the brief is the load-bearing constraint, and it is there for a measured reason rather than caution. An agent invited to confirm that a variant exists in a public database will report a genuinely novel variant as spurious, and novelty is frequently the point of the manuscript. The same failure runs the other way: a variant that is absent from ClinVar today may be absent because nobody has submitted it, not because the description is wrong. Syntax is checkable from the manuscript alone; existence is not. Keep the agent on the checkable side of that line and it produces findings you can act on without adjudicating each one.

What this agent catches

The nomenclature agent’s findings fall into nine recurring categories, listed below with the concrete downstream consequence each one causes rather than an abstract severity label.

What HGVS nomenclature actually requires

The HGVS Nomenclature is the sequence-variant description standard maintained by the HGVS Variant Nomenclature Committee, and its recommendations are specific enough that most manuscript defects are checkable against a short list of rules. The recommendations state that “a sequence variant is defined in the context of a reference sequence which must be referred to by means of a unique sequence identifier”, and that “it is mandatory to indicate the type of reference sequence file using a prefix preceding the variant description.”

The prefix set is closed: c for coding DNA, g for linear genomic, m for mitochondrial DNA, n for non-coding DNA, o for circular genomic, p for protein, and r for RNA. The reference sequence is separated from the description by a colon, in the form the recommendations give as NC_000011.9:g.12345611G>A — accession, version, colon, prefix, description. Nucleotides are written in capitals, and a substitution uses >, not an arrow, not a slash, and not the “A to G” phrasing that survives in older clinical prose. Deletions are written without the deleted nucleotides, so c.76del rather than c.76delA.

Those five rules account for the majority of what the agent reports on a first pass, because each is a surface property of a string. A manuscript that writes A1234G, 1234A→G, Arg→Trp at 123 or IVS4+2T>C is not making a scientific error; it is making a string that no HGVS-aware tool will parse and no curator can resolve without asking you which transcript you meant. The genomics agent set treats this as the most mechanically resolvable category of finding in the whole cluster for exactly that reason: the repair is a string rewrite once the location is known.

Why an unversioned transcript is the defect that changes numbers

A coding variant with no versioned transcript accession is the single most consequential nomenclature failure in a manuscript, because it is the one that silently changes the position a reader resolves. The HGVS recommendations are explicit that RefSeq and Ensembl require version numbers: NM_004006.3 is a valid reference and NM_004006 alone is not.

The mechanism is that c. numbering counts from the A of the initiation codon of one specific transcript. A gene with several annotated isoforms has several coding coordinate systems, and the same genomic change carries a different c. number in each. Change the transcript and you change the number without changing the variant; change the transcript version and you may change the number too, because transcript records are revised as annotation improves. This is why the ACMG and AMP joint consensus recommendation (Richards et al., Genetics in Medicine 2015;17:405–424) states that “a reference transcript for each gene should be used and provided in the report when describing coding variants,” and that the transcript “should either represent the longest known transcript and/or most clinically relevant transcript.”

The practical remedy is MANE. The Matched Annotation from NCBI and EMBL-EBI project defines a MANE Select transcript as “one transcript at each locus across the genome that is representative of biology at that locus”, where “each MANE transcript represents an exact match in exon coordinates, coding region, and sequence between a RefSeq transcript and its counterpart in the Ensembl/GENCODE annotation.” The project reports that MANE, currently at version 1.5, has “near-total coverage of protein-coding genes that are annotated in RefSeq and Ensembl gene sets and can be accurately represented on GRCh38.” Naming the MANE Select transcript with its version removes the ambiguity for almost every human protein-coding gene, and it makes your numbering match the numbering a curator will use.

MANE is a GRCh38 annotation, and that is the qualification a manuscript still built on GRCh37 needs to state rather than assume. A MANE Select transcript’s exon and coding coordinates are defined against GRCh38; citing “the MANE Select transcript” for a variant called against an older GRCh37 pipeline does not by itself guarantee the same c. numbering, because the transcript record the two builds’ annotations point to is not always identical. MANE publishes mappings of its transcripts onto GRCh37.p13 and onto T2T-CHM13v2.0 for exactly this situation, with the caveat that users “will need to account for sequence differences” between those assemblies and GRCh38 when mapping variants across. State the build the calls came from alongside the transcript, and let the transcript resolve the coding numbering rather than the genomic one. LRG accessions cover a much smaller set of loci than MANE, and the project is now closed: no new LRGs are being generated, the site’s last update was 31 March 2021, and it directs readers to the Ensembl and RefSeq transcripts “specified by the MANE collaboration” for all genes where those exist. Existing LRGs stay valid for the genes that already have one, chiefly those with an established diagnostic history, so treat them as a legacy set to honour where your field already uses them — not a general substitute, and not somewhere to send a gene that has no LRG today.

Where a gene already carries one, LRG accessions (LRG_#, with LRG_#t# for transcripts and LRG_#p# for proteins) take no version suffix, because the sequence is fixed by definition and will never be revised — that permanence is exactly the property a diagnostic laboratory wanted from them. Whichever you pick, pick one per gene and use it in every table. The inconsistency the agent reports most often is not a wrong transcript but two transcripts for the same gene in the same paper, usually because two co-authors generated their variant lists in different years.

Common questions about HGVS syntax and transcripts

Do I actually have to add the version number, like the .3 in NM_004006.3?

Yes. The HGVS recommendations treat NM_004006 and NM_004006.3 as different references, because RefSeq and Ensembl both revise transcript records over time, and a later version can carry a different sequence at the same accession. Reviewers and curators cannot resolve an unversioned accession to one specific sequence, so the omission is reported as a defect rather than a style preference.

My reviewer said my variants aren’t in HGVS format — what exactly do they want changed?

Almost always three things: a coordinate-system prefix (c., g., p., m. or n.) before every variant, a named and versioned reference sequence before that prefix, and capital-letter nucleotides with > for a substitution rather than an arrow or “A to G” wording. Fixing those three items resolves most HGVS comments without touching the underlying result.

What’s actually the difference between a c. variant and a g. variant?

c. numbers a coding DNA change against one specific transcript’s coding sequence, counted from that transcript’s initiation codon; g. numbers a genomic change against a chromosome or contig accession, independent of any transcript. The same substitution can carry two different numbers under the two systems, and a paper that mixes both without saying which is which invites the exact “which position do you mean” question a curator will ask.

Why does a database like ClinVar show a different position for my variant than my manuscript does?

Almost always because the database and the manuscript are numbered against different reference transcripts, or because a genomic coordinate is missing its build. Name the same versioned transcript or chromosome accession in both places before concluding the positions disagree — once they are numbered against the same reference and still differ, only then is one of the two actually wrong.

Genome build: naming hg38 is not naming a reference sequence

Genomic coordinates require a versioned reference sequence, and a build name is not one. Assemblies are patched between releases, so naming a build leaves, in the HGVS recommendations’ words, an ambiguity as to “whether you mean the original assembly or a specific patched version.” The recommended practice is to give the versioned RefSeq chromosome accession instead — NC_000001.10 names one sequence and nothing else.

The failure lands in supplementary tables and is invisible to the author. A position such as chr7:5,527,151 names different sequence under hg19 and under hg38, so a reader who guesses the wrong build maps your result to the wrong place, and neither the coordinate nor the surrounding prose gives any signal that a guess was made. Chromosome naming compounds it, because UCSC-style chr1 and Ensembl-style 1 are not interchangeable and a pipeline that mixed them usually failed on a subset of intervals rather than erroring. This is the same class of defect that data availability and accession review catches at the methods level; the nomenclature agent catches the instances that appear in variant tables, where the build is often stated nowhere on the page a reader has open.

Mitochondrial variants have their own reference and their own trap. The HGVS recommendations name the preferred human mtDNA reference as “the Homo sapiens mitochondrion, complete genome (GenBank NC_012920.1)”, and mitochondrial descriptions take the m. prefix rather than g.. Legacy mitochondrial notation of the form A3243G is still common in clinical and mitochondrial-disease literature and is not HGVS; the HGVS form names the reference sequence and uses m. with the > substitution operator. A manuscript reporting nuclear and mitochondrial variants in one table frequently uses HGVS for the first and legacy notation for the second, which is one of the more reliable findings the agent returns.

A build name also hides which conventions produced the number in front of it, and converting a variant caller’s output into an HGVS description is a second, separate source of off-by-one errors the agent cannot resolve from the manuscript text alone. Both formats count from one — the VCF specification defines POS as “the reference position, with the 1st base having position 1” — so the trap is not the indexing base, and an author who has been told it is will go looking for the wrong thing. It is the representation. VCF requires that a simple insertion or deletion include the base before the event in both REF and ALT, with POS pointing at that padding base, so a deletion’s POS sits one position to the left of the first deleted nucleotide, and copying it into a g.del description unadjusted shifts the variant by one. The two standards also disagree about where an indel belongs in a repeat: normalised VCF is left-aligned, while HGVS applies the 3’ rule, assigning “the most 3’ position possible of the reference sequence”. In a homopolymer or a tandem repeat the same deletion therefore carries a legitimately different position in the two formats, and neither number is wrong in its own system. These are defects in how a number was converted rather than in how it is written, so the agent cannot flag them from the manuscript alone — the check belongs in the pipeline that generates the description, before it reaches the table. Hg19 and hg38 are not the only builds in circulation, either: the T2T-CHM13 assembly is a complete, gapless human genome released as a third reference, and a manuscript that calls variants against it should name T2T-CHM13 explicitly rather than leave a reader to assume hg38.

Protein descriptions and the parentheses that mean “predicted”

Protein-level descriptions carry a distinction most manuscripts drop: HGVS uses parentheses to mark a consequence that was predicted rather than observed. The recommendations give the predicted form as a reference sequence, a colon, p. and the change enclosed in parentheses — NP_003997.1:p.(Trp24Cys) is the syntax summary’s own example — and reserve the unparenthesised form for a change that was actually examined at the protein level. Manuscripts drop the parentheses and, often enough, the p. prefix along with them.

The distinction is not pedantry, because the two claims differ in what evidence supports them. p.Arg1459Ter asserts that the protein product was examined. p.(Arg1459Ter) asserts that a coding change was observed and that this is what translation software says it does. A paper that reports the parenthesised form has made a smaller and defensible claim; a paper that drops the parentheses has, in nomenclature terms, claimed a protein-level observation it did not make. Reviewers who work in variant curation read the distinction that way, and the same instinct sits behind the overclaim check applied to results prose.

Three further protein defects recur. One-letter amino acid codes — E412K, R123W — are compact and permitted, but three-letter codes are the recommended form, and a bare one-letter string carrying no p. prefix and no reference sequence is not a description a validator can resolve. Protein descriptions that contradict the coding description printed beside them are the higher-value finding, because one of the two is wrong and the manuscript does not say which; the brief instructs the agent to show both rather than guess. And legacy protein numbering that counts residues from the mature protein rather than from the initiator methionine shifts every position in a gene by a fixed offset, which produces a description that is internally coherent, historically familiar to specialists in that gene, and off by one against every database a curator will consult.

A stop codon is its own small exception to the three-letter rule, and it is more often over-corrected than got wrong. HGVS names a translation-terminating change with Ter in the three-letter system — p.Arg1459Ter — or, equivalently, with the symbol *, so p.Arg1459* is not an error; the two are the same claim in two accepted forms. The * is the portable one — the recommendations allow it “in both one- and three-letter amino acid code descriptions”, which is why p.Trp24* is a valid description rather than a mixed one. Ter is not portable in the same way, because it is itself the three-letter spelling, so pairing it with a one-letter residue — p.W24Ter — is the actual defect, and that is the direction worth reporting.

Gene symbols: current, previous, and autocorrected

A gene symbol in a manuscript should be the current HGNC-approved symbol, and “current” is a moving target with dates attached. The HGNC has approved nearly 43,000 symbols, approximately 19,000 of them for protein-coding genes, and states that “each symbol is unique and we ensure that each gene is only given one approved gene symbol.” The committee is “committed to maintaining stable gene symbols where possible”, so changes are uncommon — which is exactly why authors miss the ones that happen.

Two families of changes are worth knowing by name because they are recent, large, and driven by a data-integrity problem rather than by biology. SEPT9 became SEPTIN9 with a symbol change dated 4 April 2019 in the HGNC record, which lists SEPT9 among its previous symbols. MARCH5 became MARCHF5 with a change dated 9 July 2019, its record listing MARCH5 and RNF153 as previous symbols. Bruford and colleagues set out the reasoning in “Guidelines for human gene nomenclature” (Nature Genetics 2020;52:754–758). A paper written from literature published before those dates will carry the old symbols throughout, and every one of them is a symbol a symbol-based literature search may not reach.

The reason those families were renamed is the reason to take the whole category seriously. Ziemann, Eren and El-Osta screened 35,175 supplementary Excel files from 18 journals published between 2005 and 2015 and found gene name errors in 987 files spanning 704 published articles, concluding in Genome Biology (2016;17:177) that “approximately one-fifth of papers with supplementary Excel gene lists contain erroneous gene name conversions.” Abeysooriya and colleagues repeated the scan over PubMed Central for 2014 to 2020 and reported in PLOS Computational Biology (2021;17(7):e1008984) that gene name errors “continued to accumulate unabated in the period after 2016”, appearing in 30.9% — 3,436 of 11,117 — of articles with supplementary Excel gene lists. Narrowing to the 37 journals that had published at least 50 such articles, they found those journals “accounted for 67.9% of affected publications”. The problem is not confined to obscure venues, and it is not shrinking.

The conversion is silent and happens at paste time: a symbol is read as a date or a floating-point number, the cell displays something that looks deliberate, and no error is raised. The agent reads the manuscript text and its tables, so it sees the converted value as a symbol that is not in the HGNC vocabulary — which is precisely what it is. Species case convention is the smaller companion check: human symbols are conventionally upper case and mouse symbols title case, so Trp53 and TP53 are not typographic variants of one another, and a paper that mixes them without naming the species leaves an orthologue question open for every downstream reader. Where the paper’s claims turn on which organism was studied, that overlaps with mouse model relevance.

HGNC governs only the human symbol, which is the qualification a multi-species manuscript needs stated rather than assumed. The mouse symbol is assigned separately, by the International Committee on Standardized Genetic Nomenclature for Mice and maintained at Mouse Genome Informatics, and other species have their own committees again — so Trp53 and TP53 are two separately maintained symbols for orthologous genes, not one symbol written two ways. The committees coordinate deliberately rather than working in isolation: the mouse guidance is that symbols “should agree with human gene symbols when practical”, and the HGNC describes working “in close collaboration with staff at other gene nomenclature resources”. But agreement is a goal, not a guarantee, and Trp53 against TP53 is the standing example of a difference that is settled, permanent, and easy for a copy-editor to mistake for a typo worth fixing. A manuscript that reports a human finding alongside a mouse model is naming two authorities at once, and the agent checks each symbol against the vocabulary for the species the sentence around it claims.

Common questions about gene symbol errors

Why did Excel turn some of my gene names into dates, and how do I fix it before I submit?

Spreadsheet software reads symbols such as SEPT2 or MARCH1 as a day-and-month pattern and silently reformats the cell, with nothing on screen flagging the change. Reformatting the cell back to text afterward does not recover the original characters — re-enter the symbol from a source never opened in a spreadsheet, or set the column to text format before the first paste, then re-export.

Is it SEPT9 or SEPTIN9 now — which one do I use in a new manuscript?

SEPTIN9. HGNC changed the approved symbol from SEPT9 to SEPTIN9 on 4 April 2019, as part of a wider renaming of the whole septin gene family, and SEPT9 is now listed only as a previous symbol in the HGNC record. Use SEPTIN9 in anything written after that date, and keep SEPT9 only where you are quoting older literature verbatim.

My earlier paper used MARCH5 — does a new manuscript need to rename it?

Yes. HGNC renamed MARCH5 to MARCHF5 on 9 July 2019, alongside the rest of the membrane-associated RING-CH family, for the same spreadsheet-autocorrect reason the septins were renamed. A manuscript submitted now should use MARCHF5; citing the earlier paper does not require changing its title, only the symbol used in your own new text.

Worked example: one table row, six defects

Consider a row of a variant table of the kind that appears in submitted manuscripts every week, giving a gene, a variant, a protein change and a position: SEPT7 | 1234G>A | E412K | chr7:35,890,000 (SEPTIN7 is the same class of spreadsheet-autocorrect symbol, is 437 aa — UniProt Q16181 — so residue 412 and c.1234 both exist, and sits on the forward strand at GRCh38 chr7:35,800,932-35,907,110, which contains 35,890,000).

The agent reports six distinct defects from that row alone. The gene symbol is a previous symbol; the current HGNC-approved symbol is SEPTIN7, changed on 4 April 2019. The variant has no prefix, so nothing states whether 1234 is a coding, genomic or non-coding position. The variant has no reference sequence, so even reading 1234 as coding leaves it numbered against an unnamed transcript. The protein change is in one-letter code rather than three-letter HGVS form, and it carries no p. prefix and no parentheses, so it reads as an observed protein change when it is almost certainly a translation prediction. The genomic coordinate is given with no build and no versioned chromosome accession, so it resolves to different sequence under GRCh37 and GRCh38. And the coding and genomic descriptions are not cross-checkable against each other, because neither names the sequence it is numbered against.

The repaired row names each of those: the current symbol SEPTIN7; the coding change as the gene’s MANE Select transcript accession with its version, a colon, and c.1234G>A; the predicted protein consequence as p.(Glu412Lys); and the genomic position, once the author has confirmed which build the coordinate came from, as a versioned chromosome accession in the form NC_000007.14:g.35890000G>A rather than a bare chr7 coordinate. That last repair is the only one the author has to answer a question to make, which is itself the argument for stating the build in the methods. The whole row takes about twenty extra characters. It is the difference between a table a curator can ingest and a table a curator has to email you about.

A seventh defect usually appears once the same variant is compared across the paper. The abstract, the results text, the table and the figure legend were each typed at a different time, and a transposed digit or a flipped allele in one of the four is the most common finding this agent returns. That is why the brief ranks internal consistency as its highest-value check and instructs the agent to report those first.

How to check your own manuscript before the agent does

Run the check as a reconciliation rather than a read-through, because most of these defects are mismatches between two places in the paper rather than errors visible in either place alone.

1. Extract every variant string in the manuscript into one list, with its location — abstract, results, each table, each figure legend, each supplementary file. A variant that appears once cannot be inconsistent; a variant that appears four times usually is.
2. Assign one reference transcript per gene, with its version, and write it down before you edit anything. Prefer the MANE Select transcript where one exists.
3. Validate each description mechanically. VariantValidator states that it validates “HGVS sequence variation descriptions, accurately mapping between transcript and genomic variants”, and Mutalyzer performs the same class of check; both will tell you that a description is unparseable, which is the finding you want before a reviewer supplies it.
4. Check every gene symbol against the current HGNC vocabulary, including symbols that appear only in a figure axis label or a supplementary sheet. A symbol that resolves as a previous symbol rather than an approved one is the flag.
5. Open every supplementary spreadsheet and look for dates and floating-point numbers in a column that should contain symbols. Then fix the source, not the display format, and re-export.
6. State the genome build and the annotation release once, prominently, in the methods, and state the versioned chromosome accession alongside any coordinate table.

The agent performs steps 1, 2, 4 and 6 against the manuscript in a single pass and reports each instance with its location and the corrected form. Steps 3 and 5 are yours, because they require running the files rather than reading them. If you are assembling several briefs for one submission, how to write a custom reviewer agent covers the brief structure this one uses, and custom AI reviewer agents covers how the slots work.

What reviewers and curators actually write

Reviewers rarely write the words “nomenclature error” first. They write the specific version, and each of these comments sends the manuscript back for another revision cycle.

“Variants are not described according to HGVS nomenclature; please provide the reference transcript with version number for all coding variants.” “The transcript used for numbering is not stated, and the positions given do not correspond to the canonical transcript.” “Table 2 and the abstract report different positions for the same variant.” “Genomic coordinates are given without a genome build.” “Please use current HGNC-approved gene symbols; several symbols in Table 1 have been superseded.” “Protein consequences are reported without indicating whether they were observed or predicted.” “The mitochondrial variants are described in legacy format; please use HGVS with reference to NC_012920.1.” “Supplementary Table 4 contains dates in the gene symbol column.”

The last one is the comment authors find most embarrassing and the easiest to prevent, and it is the one most likely to arrive from a data curator rather than from a scientific reviewer. Answering any of these is a short response to reviewers with a manuscript edit attached — provided the underlying facts are still recoverable, which is the reason to run the check before the variant caller’s output has been archived and the person who ran it has left.

What this agent does not do

The nomenclature agent checks description, not biology. It does not assert that a variant is pathogenic, benign or of uncertain significance, and the brief instructs it not to; clinical interpretation requires curated evidence, segregation data and functional work that no reading of a manuscript can supply. It does not confirm that a variant exists in any database, for the reason given above. It does not decide which transcript should be canonical for your gene — it reports that two transcripts are in use where one should be, and leaves the choice to you and to the MANE Select record.

It also does not judge whether the variants you found support the claim you make about them. That is a separate question about evidence and inference, and it belongs with multiple testing and enrichment when the variant list came from a screen, and with ethics and data sharing review when the variants are germline and identifiable. Manuscripts are read as .docx. Because the paper is unpublished when you run this check, how PerfectPaper handles unpublished research is worth reading first.

How this differs from the built-in review

The standing team includes a terminology and abbreviation consistency specialist, which catches inconsistent usage in general prose. It is not briefed on HGVS syntax, on transcript versioning, or on HGNC symbol status, all of which are formal conventions with specific correct forms rather than matters of consistent style.

That distinction is the whole argument for spending a slot. A consistency specialist reading A1234G in three places sees three consistent strings and reports nothing; a nomenclature specialist sees a string with no prefix and no reference sequence and reports it three times with locations. Consistency is a property of the manuscript. Conformance is a property of the manuscript measured against an external standard, and it takes a reviewer who has been given the standard. For what journals check on arrival, see journal policies.

This agent is quick to run and worth a slot only when the paper reports specific variants. Otherwise spend the slot on multiple testing or data availability. Full set: AI peer review for genomics.

PerfectPaper reads the abstract, the results text, every table and every figure caption together, and reports each variant string that does not name the reference sequence it is numbered against, along with every gene symbol the HGNC has superseded.

Review my manuscript

Related agents

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

• Multiple testing and enrichment — use it when a gene list is called significant on nominal p, or enrichment runs against the wrong background.
• Data availability and accessions — use it before submission, since a missing accession is what gets a paper returned without review.
• Batch effects and replicate structure — use it when samples crossed runs or days, or when n counts passages rather than animals.
• Metabolomics and flux reporting — use it for identification confidence, normalisation, and natural-abundance correction.

Frequently asked questions

Can I just paste my manuscript into ChatGPT and ask it to check the HGVS nomenclature instead of using a dedicated agent?

A general-purpose assistant can catch the obvious cases — a missing prefix, an arrow instead of > — but nothing tells it that SEPT9 became SEPTIN9 in 2019, or that a coordinate needs a stated build, unless a brief and, where needed, a symbol-status lookup are supplied. Treat either tool’s output as instances to verify against the standard, not as a certified proofread.

Does an AI nomenclature reviewer decide whether my variant is pathogenic?

No, and the brief instructs it not to. It checks that each variant is described unambiguously and consistently against a named reference sequence. Clinical interpretation requires curated evidence, segregation data and functional work that no reading of a manuscript can supply.

Will an AI reviewer flag mouse gene symbols as errors?

No. It checks that case usage is internally consistent and that the species is stated, since human symbols are conventionally upper case and mouse symbols title case, and several symbols are shared across species. Trp53 in a mouse paper is correct; Trp53 and TP53 used interchangeably with no species named is the finding.

Can an AI reviewer tell if my variant actually exists, or does it only check the spelling?

Spelling and structure only, by design. The brief instructs the agent not to query variant databases, because a variant absent from ClinVar today may simply be unsubmitted rather than wrong, and a manuscript’s whole contribution is sometimes a variant nobody has recorded yet. Existence and clinical significance are separate questions that require curated evidence the agent is deliberately not given.

I’ve typed the same variant into a table, the abstract and a figure legend — how would I even notice if one of them is wrong?

Line them up rather than reading each in isolation: extract every occurrence into one list with its location, and compare position, reference allele and alternate allele across all four. A transposed digit or a flipped allele reads as correct in every place it appears alone, which is exactly why this is the highest-value check in the brief and the one authors miss reading start to finish.