How do you translate GFF3 genome annotation documentation and human-readable feature descriptions without changing where genes, transcripts or coding sequences are located? Generic Feature Format version 3 is a structured way to represent genomic features and relationships between them. It looks like a simple tab-separated file, yet its coordinates, feature ontology terms, attribute IDs, parent links and coding-sequence phases are the instructions that let software reconstruct gene models. A fluent translation that changes those structures can produce the wrong model of a genome.
Searches for GFF3 translation, GFF3 file format explained, genome annotation localization, GFF3 ID and Parent attributes, Sequence Ontology feature types, CDS phase 0 1 2, GFF3 percent encoding, gene versus mRNA features and GTF to GFF3 conversion often reveal a common confusion: human-facing gene descriptions and machine-readable annotations are not the same layer. A laboratory may want a Japanese-language tutorial, a repository may need translated Notes, and a genome browser may display localized feature labels. None of those aims automatically authorizes rewriting the underlying coordinates or reference relations.
This international guide explains the nine GFF3 columns, 1-based inclusive coordinates, feature types, attribute syntax, hierarchical gene models, multi-parent exons, discontinuous features, sequence directives, embedded FASTA and validation. It is written for annotation teams, life-science editors, educators, software developers and translators. Specific genome projects may use extensions or conventions different from the core Sequence Ontology specification; always read the generating database’s format notes and validate real files with the intended parser.
The sixty-second answer
Keep the original nine GFF3 columns, sequence identifiers, 1-based inclusive positions, ontology-backed types, strand, CDS phase and graph relationships intact. Translate authorized explanatory material, display labels and human-readable notes in a separate layer wherever possible. Validate percent encoding, ID and Parent graphs, embedded reference sequence and edge cases before distributing a localized annotation guide.
1. GFF3 is a feature-relationship format
GFF3 describes annotated intervals on reference sequences and relationships among features such as genes, transcripts, exons and coding regions. It is not just a list of gene names to translate.
For example: A genome portal localizes the caption for a gene model while leaving the GFF3 records and parent relationships stable.
The main failure mode: Changing a displayed gene label is different from changing a feature ID that downstream software uses to connect components.
Check it this way: Keep the original annotation file immutable and make translated explanatory surfaces independently versioned.
2. Nine columns carry distinct meanings
Each normal feature line is tab-delimited into seqid, source, type, start, end, score, strand, phase and attributes. Their positions are part of the format contract.
In practice: A Spanish training table prints the nine canonical names alongside natural-language descriptions and a synthetic sample line.
What goes wrong: Inserting an additional translated column or replacing tabs with variable spaces can cause parser failures or shift field interpretation.
The acceptance test: Count fields and validate tab separation across all data lines, especially after spreadsheet or word-processor export.
3. The seqid is a reference key
Column one identifies the reference sequence on which the feature lies. The value must agree with the sequences the annotation project uses, not merely with a human-readable chromosome name.
Picture a real handover: A record on NC_000001.11 can be described in German as a human chromosome reference, but the accession itself remains unchanged.
An unsafe shortcut: Renaming the seqid into a translated term detaches coordinates from their intended sequence and may break browser display.
For a trustworthy release: Compare seqid values with sequence-region directives, reference FASTA identifiers and annotation provenance.
4. Source identifies where the annotation came from
The source column typically describes an annotation method, database or program. Its exact interpretation may be project-specific, so not every apparent word should be translated.
To make it concrete: A project retains a source token identifying its gene prediction tool while documenting the method in several languages.
A tempting confusion: Turning a program name into a common translated phrase can make separate annotation sources appear identical.
Before publishing: Maintain source-token mappings in a glossary and consult the generating pipeline before permitting any change.
5. Feature type belongs to a controlled vocabulary
The third column identifies the kind of feature, commonly using Sequence Ontology terms or accessions. Types such as gene, mRNA, exon and CDS are not casual synonyms.
For example: A Portuguese explanation of an exon can use familiar language while preserving the feature-type token required by downstream tools.
The main failure mode: Renaming type=mRNA to a colloquial phrase may prevent a parser from recognizing the model or lead to unsupported inference.
Check it this way: Verify feature types against the relevant Sequence Ontology definition and any documented project extensions.
6. Start and end are one-based inclusive
GFF3 feature coordinates conventionally use one-based inclusive start and end positions. A feature from 100 through 109 covers ten reference bases, not nine.
In practice: An educational figure marks positions 100 to 109 on a short synthetic chromosome and labels both ends as included.
What goes wrong: Moving coordinates into a zero-based API without conversion leads to off-by-one boundaries, even if the language of the report is polished.
The acceptance test: Check length as end minus start plus one and document any intentional coordinate-system conversion.
7. Coordinates must match the exact assembly
A GFF3 seqid and coordinate pair refers to an underlying assembly or reference sequence version. A familiar species name does not settle which assembly is meant.
Picture a real handover: Two teams translate annotations for the same species but keep separate reference accession releases to prevent false alignment of features.
An unsafe shortcut: A translated report that drops assembly version information can make gene boundaries appear directly comparable when they are not.
For a trustworthy release: Record assembly provenance and verify each seqid against the delivered reference sequences before comparison.
8. Score is optional contextual evidence
The sixth field can contain a score or a dot for unavailable information. Its meaning depends on the producing annotation method and is not a universal confidence percentage.
To make it concrete: A gene predictor outputs a numeric score; its localized manual explains that the score follows the tool’s documented model.
A tempting confusion: Calling an unlabelled score ‘chance this gene is real’ may overstate or mischaracterize what the software measured.
Before publishing: Keep numeric values unchanged and inspect source documentation before interpreting any score scale.
9. Strand records orientation
The seventh column indicates strand according to the GFF3 convention, commonly plus, minus or other permitted states. Orientation matters for transcript and coding-feature interpretation.
For example: A genome browser uses translated labels ‘forward’ and ‘reverse’ while preserving + and – tokens in the data.
The main failure mode: Flipping a minus sign to a visually nicer dash or treating it as arithmetic subtraction can reverse a gene model.
Check it this way: Validate strand tokens and compare a known minus-strand transcript with the reference annotation.
10. CDS phase is not simply the reading-frame number
The eighth column supplies a phase value for CDS features, indicating how many bases must be skipped from the beginning of that CDS segment to reach the next codon boundary. It is not an arbitrary codon annotation.
In practice: A split coding sequence has CDS phase values that depend on cumulative coding length and strand orientation, not just its genomic start modulo three.
What goes wrong: Describing phase 1 as ‘reading frame 1’ without defining the actual rule can teach an incorrect transformation.
The acceptance test: Reconstruct the coding sequence using the documented phase semantics and validate translation with a suitable annotation tool.
11. The dot has a specific missing-value role
A dot can indicate that a value is undefined or not applicable in score, strand or phase contexts as permitted. It should not be silently replaced with zero or a guessed category.
Picture a real handover: A noncoding feature has phase ‘.’, while a CDS record has a numeric phase; the translation keeps that distinction.
An unsafe shortcut: Changing phase ‘.’ to 0 invents a coding-frame assertion for a feature that may not have one.
For a trustworthy release: Check missingness per column and verify that CDS phase is present where required by the chosen validator.
12. Attributes use key-value syntax
Column nine contains semicolon-separated attribute entries of the form tag=value, with list and escaping rules. Keys such as ID, Parent, Name and Note have defined roles.
To make it concrete: A French guide translates what each attribute means but keeps raw syntax like ID=tx1;Parent=gene1 inside data examples.
A tempting confusion: Replacing semicolons with local punctuation or putting prose directly between attribute pairs destroys parseable structure.
Before publishing: Parse attributes with a GFF3-aware library rather than naive word-processing operations.
13. ID defines a feature identity
The ID attribute identifies features within a GFF3 file under rules including certain discontinuous-feature cases. Child features use IDs to connect to their parents.
For example: A transcript with ID=tx1 is cited as tx1 in its exon relationships, regardless of the reader’s preferred language.
The main failure mode: Translating tx1 into an ordinary-language label breaks Parent references and can join or separate the wrong records.
Check it this way: List all IDs and verify expected uniqueness and allowed repeated-ID cases before releasing a file.
14. Name is a display label, not necessarily a stable key
The Name attribute is often intended for human display, but its use is project-specific. It may include a well-established gene symbol, not a phrase to translate.
In practice: A repository may display an approved local-language explanation next to a gene symbol while retaining the official symbol in Name.
What goes wrong: Replacing an official symbol with a translated description can break exact searches, literature matching and external database joins.
The acceptance test: Identify authoritative nomenclature and make display-language expansions separate from stable symbols.
15. Parent encodes relationships
Parent identifies one or more parent features and supports gene-to-transcript and transcript-to-exon hierarchies. The attribute is a graph edge, not a family relationship in ordinary prose.
Picture a real handover: A gene ID=gene1 has an mRNA child Parent=gene1; several exons then refer to the transcript ID rather than directly to the gene.
An unsafe shortcut: Replacing Parent IDs with translated words or pointing an exon to the wrong parent can change splice structures.
For a trustworthy release: Build a graph from IDs and Parent links and confirm all referenced parents exist and carry appropriate relationships.
16. Genes and transcripts are different objects
A gene region can contain multiple transcript models. Gene, mRNA and other transcript types describe different biological or annotation concepts, not interchangeable display aliases.
To make it concrete: A localized browser shows one gene with two transcript rows and uses clear terminology to explain alternative transcription.
A tempting confusion: Collapsing transcripts into a single gene line loses splice variants and coding relationships that may affect downstream analyses.
Before publishing: Compare feature counts and parent-child edges for each gene before and after any authorized change.
17. Exons belong to transcript structures
Exons are portions of a transcript model and may be shared among transcripts. An exon can have relationships that differ across isoforms, so labels alone are not enough.
For example: A synthetic gene has two transcripts sharing an exon and differing at another. The translation explains the model without creating duplicate new identities.
The main failure mode: Treating every exon as uniquely owned by one transcript can distort a model that legitimately has multiple parents.
Check it this way: Validate all Parent lists and inspect representative multi-isoform genes in a genome browser.
18. CDS is not interchangeable with exon
CDS features represent coding portions of transcript models. Exons may include untranslated regions, and coding sequences can be split across multiple exons.
In practice: A lesson shows that a transcript exon contains a 5-prime untranslated part and a CDS segment with its own phase.
What goes wrong: Calling every exon coding sequence can imply the wrong translated protein and make model boundaries inaccurate.
The acceptance test: Check CDS intervals separately from exon intervals and reconstruct coding sequence with the correct strand and phase.
19. The UTR concepts need precise terminology
Five-prime and three-prime untranslated regions have orientation relative to the transcript, not merely low versus high genomic coordinates. Strand must be considered.
Picture a real handover: For a minus-strand gene, the five-prime end may occur at a larger genomic coordinate than the three-prime end.
An unsafe shortcut: Translating 5-prime as ‘left’ or 3-prime as ‘right’ works poorly across strands and genome-browser orientations.
For a trustworthy release: Use transcript-oriented explanations and verify examples for both strand directions.
20. Multiple parents can be valid
Some exon or other part-of features may legitimately list several parents separated by commas. This represents shared components and is not automatically a format error.
To make it concrete: An exon belongs to tx1 and tx2 and has Parent=tx1,tx2 in a synthetic example; each transcript must preserve the relationship.
A tempting confusion: A spreadsheet tool may split comma-separated parents into separate cells and export only one, silently deleting graph edges.
Before publishing: Parse list-valued attributes as structured arrays and compare all parent links after conversion.
21. Discontinuous features may span lines
Some GFF3 features can occupy more than one genomic interval and may be represented by multiple lines under the specification’s rules. Repeated IDs can therefore have permitted meanings.
For example: A single biological feature has two disjoint intervals but one shared identity, distinguished from two unrelated features.
The main failure mode: A simplistic validator that rejects every repeated ID without context may flag valid cases, while accepting all duplicates is also wrong.
Check it this way: Check repeated IDs against the feature and discontinuity rules in the versioned specification.
22. Alias is not the same as ID
An Alias attribute may list alternative names, but those names do not automatically replace the feature’s unique identity. Synonyms can be useful for readers across disciplines.
In practice: A gene has an official stable identifier and several historical aliases; documentation distinguishes all of them.
What goes wrong: Replacing the ID with whichever synonym sounds familiar in a target language risks ambiguous database matching.
The acceptance test: Keep the canonical identifier fixed, and document aliases with their source and intended display context.
23. Note is closer to natural-language text
A Note attribute may carry free-form explanatory material and is a plausible localization candidate when the project’s rules allow it. Even then, its value must obey GFF3 escaping and provenance requirements.
Picture a real handover: A plant annotation includes Note describing a predicted protein family. The repository could provide an approved Spanish display note alongside its original.
An unsafe shortcut: Changing ‘predicted’ into ‘confirmed’ or removing a negation alters scientific certainty, not merely wording.
For a trustworthy release: Translate qualifiers faithfully, keep a source mapping, and validate encoding before any authorized rewrite of Note.
24. Dbxref and cross-references are machine keys
Dbxref attributes connect annotations to external database identifiers. Their syntax and accession values support precise cross-referencing and are not ordinary vocabulary.
To make it concrete: A French tooltip describes an external database but keeps the accession string exactly as supplied.
A tempting confusion: Translating a database prefix or inserting spaces into an accession may create a link that points nowhere.
Before publishing: Test cross-reference syntax and representative links, and identify database version or retirement where relevant.
25. Ontology_term values denote controlled concepts
Ontology references point to structured terms with identifiers and definitions. The local-language label can change, but the stable ontology identifier is the semantic anchor.
For example: An interface displays a translated term label with the original Sequence Ontology accession and its source.
The main failure mode: Replacing a controlled term ID with a free-form phrase destroys the ability to combine annotations across languages.
Check it this way: Use ontology-backed labels and preserve identifiers, allowing experts to check definitions rather than relying on surface similarity.
26. Reserved punctuation needs percent encoding
GFF3 attribute syntax reserves separators and special characters. Percent encoding according to the specification allows values to contain characters that would otherwise be misinterpreted.
In practice: A localized Note contains a literal semicolon. It must be encoded appropriately inside an attribute rather than mistaken for the next key-value pair.
What goes wrong: A translator may paste raw punctuation into a value, causing the parser to read one intended note as multiple attributes.
The acceptance test: Use a GFF3-aware serializer and test that decoding restores the intended characters without changing attribute structure.
27. Commas can divide multiple values
In list-valued attributes, commas have structural meaning. A comma intended as text must be handled according to attribute encoding rules.
Picture a real handover: A list of two Parent IDs is represented with a delimiter, while prose containing a comma inside Note must not be parsed as multiple parents.
An unsafe shortcut: Blindly replacing commas with localized list punctuation can collapse values or change how software counts relationships.
For a trustworthy release: Define each attribute’s multiplicity and round-trip its parsed values before and after a localization change.
28. Tabs and line breaks are not decorative
GFF3 depends on tab-separated columns and line-oriented records. Literal tab, newline or carriage-return characters in attribute text require appropriate encoding and cannot break a record.
To make it concrete: A multilingual note is stored as a display-layer paragraph while the GFF3 remains a single parseable line.
A tempting confusion: Exporting a richly formatted translated sentence with raw line breaks into column nine may create false new feature records.
Before publishing: Compare record counts, column counts and encoding after export; avoid ordinary text editors for production rewrites.
29. Target and Gap may encode alignments
Target and Gap attributes can describe relationships between a feature and another sequence or alignment using structured notation. Their values are not simply notes about scientific targets.
For example: A transcript-alignment annotation retains its target accession and Gap operators while documentation explains the relationship.
The main failure mode: Rewriting Gap punctuation or translating operator codes can change the correspondence between two sequences.
Check it this way: Validate target coordinates and operations with a format-aware parser and supporting alignment evidence.
30. Derives_from does not mean Parent
The Derives_from attribute can express derivation relationships distinct from part-of hierarchy encoded by Parent. Different graph edges answer different biological questions.
In practice: An annotation documents a processed feature derived from another sequence feature without claiming the source is its structural parent.
What goes wrong: Treating Derives_from and Parent as interchangeable in a translation may construct the wrong feature hierarchy.
The acceptance test: Preserve attribute names and verify both derivation and part-of relationships in a graph-aware check.
31. Sequence-region directives set context
Directive lines beginning with ## can identify version, sequence regions or other metadata. They give context for interpreting features and may have specification-defined syntax.
Picture a real handover: A translated tutorial explains ##sequence-region while preserving the real seqid and coordinate bounds in the source file.
An unsafe shortcut: Changing a declared region to match a localized display label can make valid features appear out of bounds.
For a trustworthy release: Validate sequence-region declarations against features and reference lengths.
32. Version directives control expectations
A GFF3 file should declare its format version, typically ##gff-version 3, so readers know which dialect to parse. This is a protocol marker, not a book edition caption.
To make it concrete: A training guide describes version three in Tamil while leaving the exact directive intact in the example data.
A tempting confusion: Replacing the directive with a translated sentence can make the file unrecognizable to strict parsers.
Before publishing: Keep version declarations byte-compatible with the expected tools and audit any claimed migration separately.
33. The embedded FASTA section is different data
A GFF3 file can include a ##FASTA directive followed by sequence records. Those base strings represent reference data rather than annotation prose.
For example: A repository adds local-language documentation about the embedded sequence but preserves its FASTA headers and nucleotide strings.
The main failure mode: Running text translation over everything after ##FASTA can modify genomic sequence and invalidate annotation coordinates.
Check it this way: Detect the FASTA transition explicitly and verify sequence identities, lengths and checksums.
34. GFF2, GTF and GFF3 are not synonyms
Related annotation formats share some visual conventions but differ in attribute syntax, feature semantics and hierarchy handling. Converting between them can require assumptions.
In practice: A pipeline receives a GTF export and a GFF3 file. The documentation states that conversion is a separate technical step.
What goes wrong: Renaming a .gtf file to .gff3 or translating its attribute keys does not create a valid GFF3 model.
The acceptance test: Use a tested converter, retain original files and compare gene, transcript, exon and CDS relationships.
35. Gene names follow nomenclature authorities
Official gene symbols, accessions, synonyms and informal protein names may differ. Translation should distinguish a label intended for display from a stable recognized symbol.
Picture a real handover: A plant genomics site offers a Spanish description but keeps an established Latin-letter gene symbol and its accession.
An unsafe shortcut: Translating a gene symbol into a common word can make literature searches fail and may accidentally refer to another locus.
For a trustworthy release: Consult the appropriate organism nomenclature authority and preserve approved identifiers across all language views.
36. Partial and uncertain models need qualifiers
Genome annotation often includes predicted, incomplete or computationally inferred features. These qualifiers matter because they describe evidence status, not stylistic preference.
To make it concrete: A translated Note preserves ‘putative’ or ‘predicted’ for an uncertain gene model rather than making the statement sound more definite.
A tempting confusion: Removing uncertain wording because it seems hesitant transforms the scientific claim and may mislead downstream users.
Before publishing: Keep inference qualifiers attached to each source claim and seek domain review for disputed terminology.
37. Protein translation is not language translation
A coding sequence may be biologically translated into amino acids, but that genetic-code operation is not the same as rendering an explanatory sentence in another spoken language.
For example: A lesson explicitly distinguishes protein translation using codons from human-language localization of a CDS description.
The main failure mode: A search-intent article can accidentally imply that translating GFF3 means generating a protein sequence or converting languages inside DNA.
Check it this way: Define the intended output at the start of each project and keep biological sequence computation separate from prose translation.
38. Isoform counts depend on graph structure
Counting transcripts or alternative splice patterns requires accurate parent-child links and feature types. Human-friendly naming conventions cannot substitute for the relationships in the data.
In practice: A browser lists three isoforms for one gene based on three transcript nodes and their exon/CDS structures.
What goes wrong: Renaming or merging transcript identifiers for readability may reduce the apparent count even when no biological evidence changed.
The acceptance test: Compare node and edge counts, and inspect representative multi-isoform genes using a genome browser.
39. Validate parent cycles and orphan children
Hierarchical annotation should be checked for missing Parent targets, inappropriate cycles and inconsistent containment relationships, with exceptions guided by the specification.
Picture a real handover: A synthetic example includes a transcript pointing to a gene ID absent from the file, prompting an explicit repair request rather than a translation guess.
An unsafe shortcut: A translated report may gloss over a broken reference because the gene names appear coherent, leaving software unable to reconstruct the model.
For a trustworthy release: Run a graph-aware GFF3 validator and escalate source errors to the annotation owner.
40. Coordinate containment and strand should agree
A child feature is typically located consistently with its parent on the reference and strand, subject to biological and format-specific rules. Simple text proofreading cannot establish this.
To make it concrete: An exon assigned to the wrong transcript appears visually plausible but falls outside that transcript’s stated interval.
A tempting confusion: Replacing an ID while translating a note can create a false containment or strand relationship that ordinary spell-check cannot detect.
Before publishing: Check child-parent coordinate relationships and visualize sampled gene models with a trusted reference browser.
41. Visualizations are interpretations of data
A genome browser draws boxes, arrows and labels from GFF3. It can localize visible labels without changing feature positions or graph edges.
For example: A Spanish browser translates tooltips and category names but renders the same exon and CDS boundaries as the English view.
The main failure mode: An attractive translated diagram can conceal that a parser skipped records because of changed attribute punctuation.
Check it this way: Compare screenshots with parsed feature counts and boundary tests, not appearance alone.
42. Original source and revised annotation need separate identities
If scientists deliberately correct a coordinate, ID, Parent link or phase, they are making an annotation update, not a language translation. That change deserves its own version and provenance.
In practice: An annotation team discovers a phase error during translation review and routes a proposed corrected file to a scientific owner.
What goes wrong: Silently fixing an apparent source error in the local-language copy creates different biological models under language labels.
The acceptance test: Keep a canonical source, a separate issue log and a documented review process for all scientific edits.
43. Multilingual glossary entries must include context
Useful entries distinguish GFF3 tokens, biological concepts and target-language display wording. A short glossary should explain Parent, ID, phase, transcript, exon and CDS with examples.
Picture a real handover: A bilingual editor attaches the same ontology accession to translated feature labels and flags machine-readable keys as locked.
An unsafe shortcut: A one-word dictionary mapping for ‘phase’ or ‘source’ can be grammatical but wrong in this technical context.
For a trustworthy release: Check each glossary entry against a real GFF3 record and ask a domain reviewer to test the implied relationship.
44. A synthetic gene model is the best classroom start
Build a miniature GFF3 with one gene, one mRNA, two exons and two CDS segments. Verify IDs, Parent references, strand, boundaries and CDS phase with software.
To make it concrete: A student in Arabic and another in English can reconstruct the same hierarchy from the unchanged source line values.
A tempting confusion: If their translated explanations produce different parent-child diagrams, at least one has altered the scientific meaning.
Before publishing: Assess parsed structure and language clarity together, using a known-good synthetic file as the answer key.
45. Metadata privacy and consent still matter
Some annotation projects relate to restricted genomes or unpublished research. A format that mainly describes features may still contain identifying sample names or private database references.
For example: A consortium publishes a generic multilingual GFF3 tutorial but grants access to the real restricted annotations only under its agreements.
The main failure mode: A translator may expose unpublished genome IDs or sample names by pasting raw rows into a public translation website.
Check it this way: Review identifiers and provenance before sharing, use synthetic training data and follow approved access and retention rules.
46. A safe release is an evidence-preserving handover
The release should contain the original GFF3 and assembly context, separately versioned translations, a glossary of protected identifiers, parsing checks, and sign-off by a scientific owner.
In practice: An archive offers several language editions of a user guide pointing to the same feature-file checksum and reference accession.
What goes wrong: A release without proof of graph and coordinate fidelity may distribute different gene models disguised as different languages.
The acceptance test: Validate all nine columns, directives, attribute decoding, Parent graphs, CDS phase and embedded FASTA before publication.
47. Quality assurance must test edge cases
A useful test corpus includes negative-strand models, multiple isoforms, shared exons, discontinuous features, encoded punctuation, missing values and optional FASTA.
Picture a real handover: A validator passes a small simple gene, then exposes an escaping error in a more realistic translated Note containing reserved punctuation.
An unsafe shortcut: If QA examines only an English-only gene name and a single exon, it cannot establish safe behavior on complex annotations.
For a trustworthy release: Test representative records for the actual species and pipeline, and document all exceptions with the specification version.
48. Learning the enduring principle
A scientific identifier is not an English word merely because it resembles one. In GFF3, that principle applies from seqid and ontology type through Parent, phase and attribute separators.
To make it concrete: Readers who understand this distinction can localize an interface confidently while respecting computational structure across languages.
A tempting confusion: The false shortcut is to assume that every piece of text in a text file is a valid translation unit.
Before publishing: Ask a reader to sort a synthetic annotation into protected tokens, translatable prose and scientific changes requiring approval.
Six worked examples
A gene with two transcripts
Imagine one gene ID=geneA, two mRNA children and several exons, some shared. A French browser can label ‘transcript’ and ‘exon’ in French while retaining each ID and Parent association. The test is whether the graph still has exactly the intended nodes and edges after importing the source file.
A minus-strand CDS
A transcript on the minus strand has CDS segments whose phases depend on coding-sequence continuity in transcription orientation. An ordinary-language report should not say the lowest genomic coordinate is always the transcript’s 5-prime end. Reconstruct the CDS and translated protein with a domain tool before approving descriptions of boundaries.
A semicolon inside a Note
A human-readable Note contains the text ‘predicted; low confidence’. In raw GFF3, a literal semicolon within an attribute value requires proper encoding. A translator who pastes the punctuation directly can create an apparent new attribute. Use the specification’s encoding rules and verify the parsed Note exactly.
A shared exon
Two isoforms refer to the same exon by a multi-value Parent list. A spreadsheet conversion keeps only the first parent, silently changing the model. A translated explanation must identify the shared component, and a data engineer must restore the graph from the approved source rather than inventing a new exon ID.
Reference build mismatch
A gene appears in two annotation releases with changed genomic coordinates. A multilingual educational article should identify the specific assembly and release. It must not imply that one set of coordinates can be copied to another reference merely by translating the gene name.
A mistaken GTF import
A file visually resembles GFF3 but uses GTF-style attributes. Renaming its extension does not convert syntax or relationships. Establish the true source format, use a suitable converter, validate Parent and phase fields, and log the scientific conversion independently from the human-language documentation.
Frequently asked questions
Can I translate gene feature IDs?
Do not alter stable IDs in a production GFF3. Translate descriptions and user-facing labels in a separate presentation layer.
Is GFF3 phase the same as a reading-frame label?
No. In CDS records it describes the skip to the next codon boundary from the segment’s relevant beginning.
Why is Parent important?
Parent creates part-of relationships between annotation features, allowing software to reconstruct gene and transcript hierarchies.
Can a GFF3 feature have two parents?
Yes, in permitted contexts such as a shared exon. Preserve the complete relationship list.
Why is the Note attribute risky to edit?
It can be genuinely linguistic text, but it still sits inside structured syntax and its scientific qualifiers and escaping must be preserved.
Is a GTF file the same as GFF3?
No. Their syntactic and modelling conventions differ; conversion is a separate data-processing operation.
Can I translate nucleotide strings after ##FASTA?
No. Those are reference sequence data, not prose.
What is the most important acceptance test?
Parse the localized release’s actual data artifacts and compare coordinates, IDs, feature types, Parent graphs and CDS structure with the canonical source.
Practical release checklist
Lock the source assembly, GFF3 file and specification conventions; inventory seqids, feature types, IDs and Parent links; identify approved display text and scientific qualifiers; avoid editing coordinates, CDS phases or ontology IDs during language work; percent-encode authorized attribute text through a proper serializer; validate all nine columns and graph links; check reference regions and optional FASTA; test shared and discontinuous features; compare counts and sample gene models; and obtain linguistic and biological sign-off independently.
Authoritative sources and related reading
Sequence Ontology: GFF3 specification
NCBI Datasets: GFF3 format and NCBI-specific conventions