gtf_pyparser

A lightweight, dependency-free Python library for parsing GTF (Gene Transfer Format) files into a structured object model.

Why gtf_pyparser?

Most GTF parsing libraries are either too heavy (requiring pandas, SQLite, or large C extensions) or return flat data structures that don't reflect the natural hierarchy of genomic annotation. gtf_pyparser gives you a clean Gene → Transcript → features object model with no dependencies beyond the Python standard library.

new V0.3+

added intervall tree support. you can now query genes over a position.

Roadmap

Currently working on direct access / indexing by name and position.

Installation

pip install gtf_pyparser

Position lookups (Gtf.build_intervals / Gtf.get_genes_at_position) require the optional intervaltree dependency:

pip install gtf_pyparser[intervals]

Quick start

import gtf_pyparser

genes = gtf_pyparser.parse_gtf("Homo_sapiens.GRCh38.gtf")

gene = genes["ENSG00000139618"]
print(gene.symbol, gene.start, gene.end)

for transcript_id, transcript in gene.transcripts.items():
    exons = transcript.features.get("exon", [])
    introns = gtf_pyparser.get_intron(exons)
    print(f"  {transcript_id}: {len(exons)} exons, {len(introns)} introns")

Coordinate convention

All coordinates are 0-based half-open (start inclusive, end exclusive), consistent with BED and BAM format. GTF files are 1-based inclusive; the conversion is applied automatically during parsing.

API

Parsing

gtf_pyparser.parse_gtf(gtf_file, primary_key="gene_id")

Parse a GTF file into a dict-like container of Gene objects, keyed by primary_key.

genes = gtf_pyparser.parse_gtf("annotation.gtf")
genes = gtf_pyparser.parse_gtf("annotation.gtf", primary_key="gene_name")

• gtf_file — path to the GTF file
• primary_key — attribute used to key the returned container (default: "gene_id")
• Returns — a Gtf object supporting indexing, in, len, iteration, and .keys()/.values()/.items(), exactly like dict[str, Gene]

gtf_pyparser.get_attr(string)

Parse a raw GTF attribute string into a key-value dictionary. Useful when processing GTF lines outside of the full parser.

attrs = gtf_pyparser.get_attr('gene_id "ENSG00000139618"; gene_name "BRCA2";')
# {"gene_id": "ENSG00000139618", "gene_name": "BRCA2"}

Repeated attribute keys (e.g. multiple tag "..." entries) are collected into a list instead of overwriting each other.

Position lookups

Requires the intervaltree package — see Installation.

Gtf.get_genes_at_position(position, flanking=0) returns the gene_id of every gene overlapping a genomic position. The underlying IntervalTree index is built lazily on first use (or explicitly via Gtf.build_intervals()).

genes = gtf_pyparser.parse_gtf("annotation.gtf")

genes.get_genes_at_position({"chr": "chr1", "start": 100_000, "end": 100_050})
# ["ENSG00000139618"]

genes.get_genes_at_position({"chr": "chr1", "start": 100_000, "end": 100_050}, flanking=500)

• position — dict-like with chr, start, end keys, start < end
• flanking — bases to extend the query on both sides (default 0)
• Returns — list[str] of overlapping gene_ids, or None if the chromosome isn't in the GTF or nothing overlaps

Derived features

gtf_pyparser.get_intron(exons)

Derive intron intervals from a list of exon Interval objects belonging to a single transcript. Introns are numbered biologically: from 1 upward on the + strand, and from n down to 1 on the - strand.

exons = transcript.features.get("exon", [])
introns = gtf_pyparser.get_intron(exons)

• exons — list of Interval, need not be pre-sorted
• Returns — list[Interval], empty if fewer than two exons are provided

Position classification

Transcript.classify_position(position, strand, strand_aware=True) and Gene.classify_position(position, strand, strand_aware=True) classify a 0-based genomic position relative to a transcript's (or every transcript of a gene's) exon/intron structure.

gene.classify_position(150, "+")
# {"ENST00000001": "exon", "ENST00000002": "intron"}

Returns one of "geneStart", "geneEnd", "exon", "intron", "junctionDonnor", "junctionAcceptor", or None if the position falls outside the feature's span. Gene.classify_position returns a dict[transcript_id, result] covering every transcript on the gene.

Data classes

Interval

Immutable (frozen) genomic coordinate record — replaced rather than mutated when coordinates need updating.

Other properties: length, position (dict with just chr/start/end/strand).

Methods:

• overlaps(other, strand_aware=True, closed=False) / contains(position, strand, strand_aware=True, closed=False) — half-open by default (touching intervals don't overlap; position == end isn't contained); pass closed=True to treat both endpoints as inclusive.
• eq_pos(other) — compare genomic position and strand only, ignoring phase/attribute.
• clone() — return a copy with its own independent attribute dict.
• to_dict() / Interval.from_dict(dict_) — round-trip through a plain dict.
• Interval.from_position_str(string, is_one_based=False) — parse a "chr:start-end(strand)" string; pass is_one_based=True to convert from 1-based inclusive (GTF/samtools region notation) to this class's 0-based half-open convention.

Transcript

Groups all GTF records sharing a transcript_id. The genomic span always reflects the union of all features seen so far.

Convenience properties: start, end, chr, phase, length, attribute (of the transcript's own interval), exons, exons_positions, intron (derived via get_intron), cds, mrna, utr_5p, utr_3p.

Gene

A gene locus containing one or more transcript isoforms.

Convenience properties: start, end, chr, phase, length, attribute, biotype (checks biotype, gene_biotype, then gene_type), transcript_names, transcript_length, has_transcript, has_exon, exon / intron (list of (gene_id, transcript_id, ...) tuples across all transcripts).

Logging

Progress and error messages are emitted under the "gtf_pyparser" logger. To suppress informational output:

import logging
logging.getLogger("gtf_pyparser").setLevel(logging.WARNING)

Versioning

gtf_pyparser.__version__ reflects the installed package's version, derived from git tags via setuptools_scm. In an unbuilt source checkout (no install step run yet) it falls back to "unknown".

Development

pip install -e ".[dev]"
pytest

License

MIT

Citation / acknowledgement: Romain Lannes 2026