AIRI-Institute/genatator-gene-finding-dataset

Overview

genatator-gene-finding-dataset is a nucleotide-level dataset for training and evaluating DNA language models on ab initio gene finding. The dataset is designed for long-context genomic modeling: each example is a contiguous genomic block, and each nucleotide has aligned labels describing transcript boundary and intragenic signal.

The dataset contains annotations for both mRNA/protein-coding transcript isoforms and lncRNA transcript isoforms. Importantly, it retains information about all selected transcript isoforms, not only a single representative transcript per gene. This makes it suitable for isoform-aware gene finding, transcript-boundary prediction, and evaluation in genomic regions where multiple transcript isoforms overlap.

Each sample contains exactly three fields:

dna_sequence
targets
metadata

Intended use

This dataset is intended for:

• training DNA language models for nucleotide-level gene finding
• predicting TSS and polyA regions directly from genomic sequence
• predicting intragenic regions on both genomic strands
• comparing models on human held-out chromosomes and the complete T2T human genome
• studying long-range gene-finding behavior in multispecies mammalian genomes
• developing models that account for all selected transcript isoforms rather than a single canonical isoform

Dataset organization

The dataset contains three splits.

Split	Description
train	Multispecies training data. It contains human training chromosomes and additional mammalian genomes.
validation	Human held-out validation chromosomes.
test	Complete T2T human genome test split.

Human chromosomes are partitioned so that a subset of hg38 chromosomes is held out for validation. The T2T human genome is used as an independent test genome.

For repository storage, Parquet files are grouped into subfolders within each split. This layout only changes the file organization; it does not change the dataset rows, target values, or statistics.

data/
  train/
    part-00000/*.parquet
    part-00001/*.parquet
    part-00002/*.parquet
  validation/
    part-00000/*.parquet
  test/
    part-00000/*.parquet

What is one sample?

One sample is not a full chromosome. The dataset is built chromosome-first and saved block-by-block.

The construction process is:

1. Process one full chromosome.
2. Build nucleotide-level targets for the full chromosome.
3. Slice the chromosome into non-overlapping blocks of at most 10,000,000 nucleotides.
4. Save each block as one Parquet row.

Therefore:

one row = one genomic block
one Parquet file = one genomic block
one block <= 10,000,000 nucleotides

A full chromosome can be reconstructed by selecting all blocks with the same genome and chrom metadata fields and concatenating them in ascending start order.

Terminology

Chromosome

A chromosome is a reference sequence from a genome assembly. In this dataset, a chromosome may be represented by one or more blocks depending on its length.

Block

A block is a non-overlapping genomic interval sliced from a chromosome after full-chromosome target construction. Blocks are used to keep each sample manageable while preserving nucleotide-level alignment between DNA and targets.

Row / sample

A row, or sample, is one block. It contains the DNA sequence for that block, the aligned target matrix, and metadata describing the block coordinates.

Target channel

A target channel is one label track across the DNA sequence. Every nucleotide has 12 target values: six channels for all selected transcript isoforms and six additional channels for mRNA/protein-coding isoforms only.

TSS

TSS means transcription start site. The TSS channels contain smooth nucleotide-level signal around transcript starts, separated by genomic strand.

polyA

polyA means transcript polyadenylation/end signal. The polyA channels contain smooth nucleotide-level signal around transcript ends, separated by genomic strand.

Intragenic region

An intragenic region is the genomic span covered by a selected transcript isoform. The intragenic channels are strand-specific binary tracks.

Combined targets

The first six target channels describe all selected transcript isoforms together: mRNA/protein-coding plus lncRNA.

mRNA-only targets

The last six target channels describe mRNA/protein-coding transcript isoforms only. These channels are useful when training or evaluating a model specifically for protein-coding gene finding.

Data schema

Each row has exactly three columns.

dna_sequence

A string containing the DNA sequence for the block.

• Type: string
• Length: <= 10,000,000
• Alphabet: DNA characters such as A, T, C, G, and N

targets

A nested numeric array aligned to dna_sequence.

• Shape: sequence_length × 12
• Interpretation: 12 nucleotide-level target channels per nucleotide

If:

L = len(row["dna_sequence"])

then:

len(row["targets"]) == L
len(row["targets"][0]) == 12

The target channel order is:

Index	Target name	Meaning
0	primary_tss_+	TSS signal on + strand for all selected transcript isoforms.
1	primary_tss_-	TSS signal on - strand for all selected transcript isoforms.
2	primary_polya_+	polyA/end signal on + strand for all selected transcript isoforms.
3	primary_polya_-	polyA/end signal on - strand for all selected transcript isoforms.
4	intragenic_regions_+	Intragenic region on + strand for all selected transcript isoforms.
5	intragenic_regions_-	Intragenic region on - strand for all selected transcript isoforms.
6	mrna_tss_+	TSS signal on + strand for mRNA/protein-coding isoforms only.
7	mrna_tss_-	TSS signal on - strand for mRNA/protein-coding isoforms only.
8	mrna_polya_+	polyA/end signal on + strand for mRNA/protein-coding isoforms only.
9	mrna_polya_-	polyA/end signal on - strand for mRNA/protein-coding isoforms only.
10	mrna_intragenic_regions_+	Intragenic region on + strand for mRNA/protein-coding isoforms only.
11	mrna_intragenic_regions_-	Intragenic region on - strand for mRNA/protein-coding isoforms only.

In this dataset, the primary_* prefix denotes the combined mRNA/protein-coding + lncRNA target channels. These first six channels aggregate all selected transcript isoforms into one gene-finding target group.

metadata

A JSON string with block-level coordinates and provenance.

Typical fields include:

{
  "genome": "GCF_000001635.26",
  "split": "train",
  "chrom": "...",
  "start": 0,
  "end": 10000000,
  "chrom_length": 195471971,
  "sequence_length": 10000000,
  "target_shape": [10000000, 12],
  "target_names": ["primary_tss_+", "..."],
  "block_size_bp": 10000000
}

The start and end fields define the block interval within the chromosome. Blocks from the same chromosome can be ordered by start.

Dataset statistics

Split-level overview

split	species	chroms	blocks	rows	parquet files	nt	mean chr	std chr	min chr	max chr	selected transcripts	lncRNA transcripts	mRNA/protein-coding transcripts
train	39	14,464	21,085	21,085	21,085	97.36 Gbp	6.73 Mbp	18.65 Mbp	100.01 kbp	278.62 Mbp	2,144,820	374,525	1,770,295
validation	1	3	27	27	27	256.29 Mbp	85.43 Mbp	42.84 Mbp	46.71 Mbp	145.14 Mbp	24,359	17,971	6,388
test	1	24	327	327	327	3.12 Gbp	129.89 Mbp	56.62 Mbp	45.09 Mbp	248.39 Mbp	161,256	31,518	129,738

DNA alphabet summary

split	files	rows	sequence chars	canonical	noncanonical	noncanon%	top noncanonical
train	21,085	21,085	97,364,554,030	92,177,733,773	5,186,820,257	5.327216%	'N':5,186,820,257
validation	27	27	256,292,786	248,801,012	7,491,774	2.923131%	'N':7,491,774
test	327	327	3,117,275,501	3,117,275,501	0	0.000000%	-

Canonical nucleotides are A, T, C, and G. Non-canonical characters are retained in the DNA sequence rather than removed.

Target sums and densities

split	target	sum	max	sum / nt
train	primary_tss_+	7,818,964.643	1.000	0.008031%
train	primary_tss_-	7,789,998.340	1.000	0.008001%
train	primary_polya_+	6,595,925.705	1.000	0.006774%
train	primary_polya_-	6,591,562.233	1.000	0.006770%
train	intragenic_regions_+	21,010,914,925.000	1.000	21.579635%
train	intragenic_regions_-	20,674,425,368.000	1.000	21.234037%
train	mrna_tss_+	6,918,800.875	1.000	0.007106%
train	mrna_tss_-	6,906,097.508	1.000	0.007093%
train	mrna_polya_+	5,718,064.030	1.000	0.005873%
train	mrna_polya_-	5,727,850.108	1.000	0.005883%
train	mrna_intragenic_regions_+	19,241,295,394.000	1.000	19.762115%
train	mrna_intragenic_regions_-	18,926,075,324.000	1.000	19.438363%
validation	primary_tss_+	38,791.431	1.000	0.015136%
validation	primary_tss_-	37,230.242	1.000	0.014526%
validation	primary_polya_+	38,605.706	1.000	0.015063%
validation	primary_polya_-	36,584.631	1.000	0.014275%
validation	intragenic_regions_+	97,415,885.000	1.000	38.009609%
validation	intragenic_regions_-	91,566,938.000	1.000	35.727474%
validation	mrna_tss_+	10,656.457	1.000	0.004158%
validation	mrna_tss_-	9,541.151	1.000	0.003723%
validation	mrna_polya_+	9,681.428	1.000	0.003777%
validation	mrna_polya_-	9,034.210	1.000	0.003525%
validation	mrna_intragenic_regions_+	52,358,303.000	1.000	20.429097%
validation	mrna_intragenic_regions_-	49,236,176.000	1.000	19.210910%
test	primary_tss_+	132,879.039	1.000	0.004263%
test	primary_tss_-	128,112.914	1.000	0.004110%
test	primary_polya_+	107,023.249	1.000	0.003433%
test	primary_polya_-	101,235.996	1.000	0.003248%
test	intragenic_regions_+	883,996,214.000	1.000	28.357975%
test	intragenic_regions_-	846,656,221.000	1.000	27.160135%
test	mrna_tss_+	93,473.404	1.000	0.002999%
test	mrna_tss_-	89,777.048	1.000	0.002880%
test	mrna_polya_+	65,976.521	1.000	0.002116%
test	mrna_polya_-	62,182.009	1.000	0.001995%
test	mrna_intragenic_regions_+	651,533,008.000	1.000	20.900719%
test	mrna_intragenic_regions_-	619,807,185.000	1.000	19.882977%

Multispecies coverage

The training data includes human plus additional mammalian assemblies. The table below lists the 39 assemblies from the multispecies collection.

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 the dataset with the Hugging Face datasets library. Replace AIRI-Institute with the repository name.

from datasets import load_dataset

ds = load_dataset("AIRI-Institute/genatator-gene-finding-dataset")

train = ds["train"]
validation = ds["validation"]
test = ds["test"]

If loading from a local folder, use the same split-level file pattern:

from datasets import load_dataset

data_files = {
    "train": "data/train/part-*/*.parquet",
    "validation": "data/validation/part-*/*.parquet",
    "test": "data/test/part-*/*.parquet",
}

ds = load_dataset("parquet", data_files=data_files)

Inspect one sample:

import json
import numpy as np

row = train[0]

dna = row["dna_sequence"]
targets = np.asarray(row["targets"], dtype=np.float32)
metadata = json.loads(row["metadata"])

print(len(dna))
print(targets.shape)
print(metadata)

Selecting target groups

The first six channels are combined mRNA/protein-coding + lncRNA targets. The primary_* prefix denotes this combined selected-transcript target group. The last six channels are mRNA/protein-coding-only targets.

import numpy as np

TARGET_NAMES = [
    "primary_tss_+",
    "primary_tss_-",
    "primary_polya_+",
    "primary_polya_-",
    "intragenic_regions_+",
    "intragenic_regions_-",
    "mrna_tss_+",
    "mrna_tss_-",
    "mrna_polya_+",
    "mrna_polya_-",
    "mrna_intragenic_regions_+",
    "mrna_intragenic_regions_-",
]

COMBINED_TARGETS = list(range(0, 6))
MRNA_ONLY_TARGETS = list(range(6, 12))

row = train[0]
y = np.asarray(row["targets"], dtype=np.float32)

# Use mRNA/protein-coding + lncRNA labels.
y_combined = y[:, COMBINED_TARGETS]

# Use mRNA/protein-coding-only labels.
y_mrna_only = y[:, MRNA_ONLY_TARGETS]

For example, a model trained on y_combined learns the combined gene-finding task across mRNA/protein-coding and lncRNA isoforms. A model trained on y_mrna_only learns the mRNA/protein-coding-only task.

Filtering to one genome or species

The metadata field stores the assembly identifier in genome. You can filter a split to one genome.

import json

def genome_is(genome_id):
    def _fn(row):
        return json.loads(row["metadata"])["genome"] == genome_id
    return _fn

mouse_train = train.filter(genome_is("GCF_000001635.26"))

You can list genomes present in a split:

import json

genomes = sorted({
    json.loads(row["metadata"])["genome"]
    for row in train
})

print(genomes)

Reconstructing one full chromosome

Because one sample is one block, a full chromosome is reconstructed by collecting all blocks for the same genome and chromosome and sorting by start.

import json
import numpy as np

def get_full_chromosome(split_ds, genome_id, chrom_name):
    rows = split_ds.filter(
        lambda row: (
            json.loads(row["metadata"])["genome"] == genome_id
            and json.loads(row["metadata"])["chrom"] == chrom_name
        )
    )

    rows = sorted(rows, key=lambda row: json.loads(row["metadata"])["start"])

    dna = "".join(row["dna_sequence"] for row in rows)
    targets = np.concatenate(
        [np.asarray(row["targets"], dtype=np.float32) for row in rows],
        axis=0,
    )

    return dna, targets

chr1_dna, chr1_targets = get_full_chromosome(train, genome_id="hg38", chrom_name="chr1")

print(len(chr1_dna))
print(chr1_targets.shape)

Getting all DNA blocks for one species

To work with all blocks from one genome, filter by metadata["genome"].

import json

def select_genome(split_ds, genome_id):
    return split_ds.filter(
        lambda row: json.loads(row["metadata"])["genome"] == genome_id
    )

mouse_blocks = select_genome(train, "GCF_000001635.26")

for row in mouse_blocks:
    meta = json.loads(row["metadata"])
    dna = row["dna_sequence"]
    print(meta["chrom"], meta["start"], meta["end"], len(dna))

For large genomes, iterating over a whole species may read many blocks. Streamed or batched processing is recommended for memory-efficient training.

Notes for model training

The dataset stores nucleotide-level labels, not tokenized model inputs. Models that use BPE or another tokenizer should tokenize dna_sequence during training and map nucleotide-level labels to token-level labels inside the data pipeline.

A common strategy for BPE token-level training is to use tokenizer offset mappings and assign each token the maximum target value over the nucleotides covered by that token. This preserves positive boundary and intragenic signals when a token spans multiple nucleotides.

Recommended evaluation usage

Use the validation split for model selection and the test split for final evaluation. The test split contains a complete T2T human genome and is intended to evaluate cross-reference generalization after training on the multispecies training data and held-out human validation chromosomes.