shmelev/genatator-segmentation-dataset

Ab initio gene segmentation benchmark (GENATATORs)

Overview

genatator-segmentation-dataset is a nucleotide-level gene segmentation dataset designed for training and evaluating DNA language models and related sequence models on transcript structure prediction. The dataset targets biologically detailed reconstruction of transcript architecture and supports benchmarking in the context of ab initio gene annotation. Each example represents one annotated transcript and provides nucleotide-resolution labels describing transcript organization, including 5' untranslated region (5' UTR), exon, intron, 3' untranslated region (3' UTR), and coding sequence (CDS).

The repository contains three configurations:

• train-human
• train-multi-specie
• val-human

In the current version of the dataset, all configurations retain all annotated transcripts for all genes. No transcript-level truncation is applied. This design supports transcript-level segmentation studies while also enabling gene-level evaluation in multi-isoform settings; however, in many practical training and benchmarking scenarios, researchers may restrict the dataset to a single representative transcript per gene using the status field.

Each sample contains exactly four fields:

• dna_sequence
• labels
• metadata
• status

Intended use

This dataset is intended for:

• training DNA language models for transcript segmentation
• fine-tuning pretrained genomic foundation models
• benchmarking nucleotide-level and gene-structure-aware segmentation methods
• evaluating generalization from human-only to multispecies training
• studying segmentation of both protein-coding and long non-coding transcripts
• evaluating models in settings where multiple transcript isoforms per gene are retained

It is particularly suitable for methods that operate on long genomic or transcript-derived sequences and produce per-nucleotide predictions.

Dataset configurations

The repository contains three configurations.

Config	Split	Description
train-human	train	Human training dataset containing all annotated transcripts for all genes.
train-multi-specie	train	Multispecies training dataset containing all annotated transcripts for all genes across human and additional mammalian assemblies.
val-human	validation	Human validation dataset containing all annotated transcripts for all genes. This configuration also provides chromosome 20 examples used for final model evaluation and gene-level metric calculation.

Dataset organization and evaluation protocol

All three configurations retain the full set of annotated transcripts for each gene. Consequently, genes with multiple transcript isoforms are represented by multiple rows in the dataset.

Human chromosomes are partitioned so that chromosomes 8, 20, and 21 are excluded from human training. In this setup:

• train-human excludes human chromosomes 8, 20, and 21
• val-human contains held-out human chromosomes, including chromosome 20
• chromosome 20 from val-human is used for final model evaluation and for calculation of the gene-level metric
• gene-level evaluation on chromosome 20 uses all available transcripts per gene, rather than a single representative isoform

For the multispecies dataset:

• human examples follow the same held-out chromosome policy
• non-human species are included according to the multispecies construction protocol used in dataset generation

Sequence length policy

All transcripts are stored in full length.

• No transcript is truncated.
• No maximum transcript-length cap is applied in the released dataset.
• Each dna_sequence and its aligned labels correspond to the complete sequence of the represented transcript.

This is important for long-context modeling, complete transcript reconstruction, and biologically rigorous evaluation of exon-intron architecture.

Data schema

Each row has exactly four columns.

dna_sequence

A string containing the DNA sequence for the transcript.

• Type: string
• Alphabet: uppercase DNA characters (A, T, C, G)

labels

A nested array of nucleotide-level target annotations aligned to dna_sequence.

• Type: nested numeric array
• Shape: sequence-length by class-dimension
• Interpretation: per-nucleotide segmentation targets used for transcript structure prediction

The target class order is:

["5UTR", "exon", "intron", "3UTR", "CDS"]

metadata

A compact string encoding transcript-level annotation in the following format:

<transcript_id>|<gene_id>|<transcript_type>|<strand>|<genome>|<chrom>|<start>:<end>

This schema is identical across all dataset configurations.

status

A binary indicator identifying the representative transcript within a gene.

• Type: integer
• Typical values: 0 or 1

Interpretation:

• status = 1 marks the transcript selected as the representative isoform for its gene
• for protein-coding transcripts, this corresponds to the transcript with the longest coding region
• for lncRNA transcripts, this corresponds to the transcript with the longest cumulative exon length
• status = 0 denotes all other transcripts of the same gene

This field is useful for researchers who wish to run training with a single transcript per gene while retaining access to the complete multi-isoform dataset. In such cases, one should restrict the data to rows with status == 1.

Metadata fields

The metadata column contains biologically interpretable attributes packed into a single string.

1. transcript_id

Transcript identifier.

Typical values may correspond to transcript accessions or annotation-specific transcript names.

This field identifies the specific transcript isoform represented by the row.

2. gene_id

Gene identifier associated with the transcript.

Typical values may correspond to reference gene identifiers or annotation-derived gene names.

This field identifies the parent gene of the transcript.

3. transcript_type

Transcript class.

Typical values include:

• mRNA
• lnc_RNA

This field indicates whether the transcript is protein-coding or long non-coding.

4. strand

Genomic strand orientation.

Typical values are:

• +
• -

This field indicates whether the transcript is encoded on the forward or reverse strand relative to the reference assembly.

5. genome

Genome or assembly identifier associated with the example.

Typical values include assembly accessions such as:

• GCF_009914755.1
• GCF_000001635.26

This field is particularly important in the multispecies dataset.

6. chrom

Chromosome or reference sequence identifier on which the transcript is located.

This field specifies the genomic contig or chromosome associated with the transcript.

7. start:end

Genomic coordinate interval associated with the transcript.

Example:

23090370:23092686

This field stores the genomic span as:

• start: 1-based genomic start coordinate
• end: genomic end coordinate

Representative-transcript filtering with status

Although the dataset retains all transcript isoforms, some training or evaluation protocols may require one transcript per gene. The status column was introduced precisely for this purpose.

Recommended use:

• use the full dataset when studying transcript diversity, isoform-aware evaluation, or gene-level metrics
• filter to status == 1 when a single representative transcript per gene is required

Selection rule for status == 1:

• protein-coding genes: transcript with the longest coding region
• lncRNA genes: transcript with the longest cumulative exon length

Multispecies training dataset

The train-multi-specie configuration includes data from 39 mammalian assemblies.

Assembly	Species
GCF_000952055.2	Aotus nancymaae
GCF_002263795.3	Bos taurus
GCF_000767855.1	Camelus bactrianus
GCF_000002285.3	Canis lupus familiaris
GCF_000151735.1	Cavia porcellus
GCF_001604975.1	Cebus imitator
GCF_000283155.1	Ceratotherium simum simum
GCF_000276665.1	Chinchilla lanigera
GCF_000260355.1	Condylura cristata
GCF_002940915.1	Desmodus rotundus
GCF_000151885.1	Dipodomys ordii
GCF_002288905.1	Enhydra lutris kenyon
GCF_000308155.1	Eptesicus fuscus
GCF_000002305.2	Equus caballus
GCF_018350175.1	Felis catus
GCF_000247695.1	Heterocephalus glaber
GCF_009914755.1	Homo sapiens
GCF_000236235.1	Ictidomys tridecemlineatus
GCF_000280705.1	Jaculus jaculus
GCF_000001905.1	Loxodonta africana
GCF_001458135.1	Marmota marmota
GCF_000165445.2	Microcebus murinus
GCF_000317375.1	Microtus ochrogaster
GCF_000001635.26	Mus musculus
GCF_900095145.1	Mus pahari
GCF_002201575.1	Neomonachus schauinslandi
GCF_000292845.1	Ochotona princeps
GCF_000260255.1	Octodon degus
GCF_000321225.1	Odobenus rosmarus divergens
GCF_009806435.1	Oryctolagus cuniculus
GCF_000181295.1	Otolemur garnettii
GCF_016772045.2	Ovis aries
GCF_000956105.1	Propithecus coquereli
GCF_003327715.1	Puma concolor
GCF_036323735.1	Rattus norvegicus
GCF_000235385.1	Saimiri boliviensis boliviensis
GCF_000181275.1	Sorex araneus
GCF_000003025.6	Sus scrofa
GCF_000243295.1	Trichechus manatus latirostris

Loading examples

Load a configuration from the Hugging Face datasets library:

from datasets import load_dataset

train_human = load_dataset("shmelev/genatator-segmentation-dataset", "train-human")["train"]
train_multi = load_dataset("shmelev/genatator-segmentation-dataset", "train-multi-specie")["train"]
val_human = load_dataset("shmelev/genatator-segmentation-dataset", "val-human")["validation"]

Access one example:

sample = train_human[0]
print(sample["dna_sequence"])
print(sample["labels"])
print(sample["metadata"])
print(sample["status"])

Filter to one representative transcript per gene:

representative_only = train_human.filter(lambda x: x["status"] == 1)

Recommended evaluation usage

For final model evaluation and for calculation of the gene-level metric:

• use chromosome 20 from val-human
• retain all available transcripts per gene
• compare predictions against the complete isoform set for each gene