AIRI-Institute/genatator-gene-finding-dataset

GENATATOR gene-finding dataset

Overview

genatator-gene-finding-dataset is a nucleotide-level dataset for training and evaluating sequence models on ab initio gene finding in mammalian genomes. Each sample contains a contiguous DNA block and 12 aligned target channels describing transcript starts, transcript ends, and transcript bodies on both genomic strands.

The dataset covers 39 mammalian training genomes, held-out human validation chromosomes, and an independent complete T2T human test genome. Across all splits it contains 100,738,122,317 nucleotides and 2,330,435 retained transcript isoforms:

• 1,906,421 mRNA/protein-coding transcript isoforms
• 424,014 lncRNA transcript isoforms

The dataset retains all annotated isoforms belonging to the retained transcript classes. It does not reduce each gene to one representative transcript. Consequently, alternative transcript starts, alternative transcript ends, and overlapping isoforms all contribute to the targets.

Each sample has three columns:

dna_sequence
targets
metadata

Intended use

The dataset is intended for:

• nucleotide-level gene finding from raw DNA
• TSS and transcript-end prediction
• strand-specific transcript-body prediction
• mRNA/protein-coding-only gene finding
• joint mRNA/protein-coding and lncRNA gene finding
• long-context DNA language-model training
• multispecies transfer and cross-reference evaluation
• isoform-aware analysis of genomic loci

Splits

Split	Content
train	hg38 human training chromosomes together with 38 additional mammalian genomes
validation	held-out hg38 chromosomes 8, 20, and 21
test	complete T2T-CHM13v2.0 human genome

The human hg38 training split excludes chromosomes 8, 20, and 21. The T2T genome is an independent reference assembly and is used only for testing.

What is one sample?

Targets are first assembled for an entire chromosome. The chromosome is then divided into consecutive, non-overlapping blocks of at most 10,000,000 nucleotides.

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

A block contains a DNA interval and all target values aligned to that interval. A full chromosome can be reconstructed by collecting blocks with the same genome and chrom, sorting them by start, and concatenating them.

The repository groups Parquet files into storage subfolders to satisfy Hugging Face repository file-count limits. The original dataset directory is not modified; this is only the remote repository layout:

data/
  train/
    hf_auto_shards/
      00000/*.parquet
      00001/*.parquet
      ...
      00210/*.parquet
  validation/
    hf_auto_shards/
      00000/*.parquet
  test/
    hf_auto_shards/
      00000/*.parquet
      ...

These subfolders are only a storage layout; they do not alter sample order, coordinates, or target values.

Retained transcripts

The targets are built from two transcript classes:

• mRNA/protein-coding transcripts
• lncRNA transcripts

Every retained transcript isoform contributes independently. If a gene has several retained isoforms, their distinct starts, ends, and transcript spans are represented in the same chromosome-level targets.

Other transcript classes are not part of the released target set.

How targets are constructed

For every retained transcript isoform:

1. The transcript span is defined by the outer boundaries of its annotated exons.
2. TSS is placed at the transcription-start boundary.
3. polyA/end is placed at the opposite transcript boundary.
4. The transcript body is marked across the full outer transcript span.
5. + and - strand targets are stored separately.

The TSS and polyA/end channels are smooth boundary targets:

• the target value is 1.0 at the annotated boundary
• values decay exponentially with nucleotide distance from that boundary
• the decay width is 10 nucleotides

When several isoforms contribute to the same target channel, the maximum boundary signal is retained at each nucleotide. Transcript-body channels represent the union of retained transcript spans.

Terminology

Chromosome

A complete reference sequence within a genome assembly. Targets are assembled chromosome-by-chromosome before block creation.

Block

A consecutive interval from a chromosome. Blocks do not overlap and contain at most 10,000,000 nucleotides.

Transcript isoform

One annotated transcript structure for a gene. Multiple isoforms of the same gene may have different starts, ends, or exon structures. All retained isoforms contribute to this dataset.

TSS

The transcription start site. On the + strand it is the left transcript boundary; on the - strand it is the right transcript boundary.

polyA/end

The transcript-end boundary. On the + strand it is the right transcript boundary; on the - strand it is the left transcript boundary.

Transcript body

The genomic span from the first transcript boundary to the last transcript boundary. It is stored as a strand-specific binary target.

Combined targets

The first six channels contain the union of mRNA/protein-coding and lncRNA isoforms.

mRNA-only targets

The final six channels contain only mRNA/protein-coding isoforms.

Data schema

dna_sequence

A DNA string for one block.

• Type: string
• Length: at most 10,000,000
• Alphabet: A, T, C, G, and N

targets

A nucleotide-aligned float array.

shape = sequence_length × 12

If L = len(row["dna_sequence"]), then:

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

Target order:

Index	Target	Description
0	primary_tss_+	Combined mRNA/protein-coding + lncRNA TSS signal on +
1	primary_tss_-	Combined mRNA/protein-coding + lncRNA TSS signal on -
2	primary_polya_+	Combined mRNA/protein-coding + lncRNA end signal on +
3	primary_polya_-	Combined mRNA/protein-coding + lncRNA end signal on -
4	intragenic_regions_+	Combined transcript-body target on +
5	intragenic_regions_-	Combined transcript-body target on -
6	mrna_tss_+	mRNA/protein-coding-only TSS signal on +
7	mrna_tss_-	mRNA/protein-coding-only TSS signal on -
8	mrna_polya_+	mRNA/protein-coding-only end signal on +
9	mrna_polya_-	mRNA/protein-coding-only end signal on -
10	mrna_intragenic_regions_+	mRNA/protein-coding-only transcript-body target on +
11	mrna_intragenic_regions_-	mRNA/protein-coding-only transcript-body target on -

The primary_* channels are the combined transcript channels. They contain all retained mRNA/protein-coding and lncRNA isoforms and do not indicate one representative transcript per gene.

metadata

A JSON string describing the block. Main fields include:

Field	Meaning
genome	genome or assembly identifier
split	train, validation, or test
chrom	chromosome or reference-sequence identifier
start	zero-based block start within the chromosome
end	zero-based exclusive block end
chrom_length	complete chromosome length
sequence_length	number of nucleotides in the block
target_shape	target-array shape
target_names	ordered target-channel names
block_size_bp	maximum block size used during construction

Dataset composition

Split-level summary

Split	Genomes	Chromosomes	Blocks	Nucleotides	Retained isoforms	mRNA/protein-coding	lncRNA	N nucleotides	N fraction
train	39	14,464	21,085	97,364,554,030	2,144,820	1,770,295	374,525	5,186,820,257	5.327216%
validation	1	3	27	256,292,786	24,359	6,388	17,971	7,491,774	2.923131%
test	1	24	327	3,117,275,501	161,256	129,738	31,518	0	0.000000%

Training genomes

The training split contains 39 mammalian genomes: hg38 human plus 38 additional mammalian assemblies.

Genome	Species	Chromosomes	Blocks	Genome length	Mean chromosome length	Retained isoforms	mRNA/protein-coding	lncRNA
GCF_000001635.26	Mus musculus	192	457	2.82 Gbp	14.67 Mbp	110,712	87,708	23,004
GCF_000001905.1	Loxodonta africana	379	624	3.16 Gbp	8.34 Mbp	43,109	40,878	2,231
GCF_000002285.3	Canis lupus familiaris	145	361	2.36 Gbp	16.29 Mbp	77,479	58,382	19,097
GCF_000002305.2	Equus caballus	157	377	2.41 Gbp	15.37 Mbp	39,755	35,618	4,137
GCF_000003025.6	Sus scrofa	146	380	2.48 Gbp	17.00 Mbp	72,350	63,184	9,166
GCF_000151735.1	Cavia porcellus	597	772	2.68 Gbp	4.48 Mbp	41,158	37,165	3,993
GCF_000151885.1	Dipodomys ordii	426	520	2.03 Gbp	4.76 Mbp	29,206	28,544	662
GCF_000165445.2	Microcebus murinus	50	276	2.46 Gbp	49.12 Mbp	70,202	58,882	11,320
GCF_000181275.1	Sorex araneus	341	491	2.38 Gbp	6.97 Mbp	22,905	22,901	4
GCF_000181295.1	Otolemur garnettii	435	560	2.50 Gbp	5.74 Mbp	33,203	32,354	849
GCF_000235385.1	Saimiri boliviensis boliviensis	284	438	2.60 Gbp	9.15 Mbp	37,815	36,114	1,701
GCF_000236235.1	Ictidomys tridecemlineatus	843	930	2.40 Gbp	2.85 Mbp	40,590	38,277	2,313
GCF_000243295.1	Trichechus manatus latirostris	505	668	3.07 Gbp	6.09 Mbp	37,371	36,282	1,089
GCF_000247695.1	Heterocephalus glaber	318	478	2.60 Gbp	8.18 Mbp	70,629	60,641	9,988
GCF_000260255.1	Octodon degus	570	717	2.96 Gbp	5.19 Mbp	43,826	42,228	1,598
GCF_000260355.1	Condylura cristata	108	253	1.76 Gbp	16.29 Mbp	29,869	29,017	852
GCF_000276665.1	Chinchilla lanigera	241	393	2.37 Gbp	9.85 Mbp	53,244	45,219	8,025
GCF_000280705.1	Jaculus jaculus	298	475	2.79 Gbp	9.37 Mbp	25,357	25,346	11
GCF_000283155.1	Ceratotherium simum simum	232	403	2.45 Gbp	10.57 Mbp	36,163	33,465	2,698
GCF_000292845.1	Ochotona princeps	277	427	2.19 Gbp	7.89 Mbp	25,702	25,388	314
GCF_000308155.1	Eptesicus fuscus	489	587	1.99 Gbp	4.07 Mbp	54,033	49,383	4,650
GCF_000317375.1	Microtus ochrogaster	264	431	2.26 Gbp	8.55 Mbp	38,400	37,563	837
GCF_000321225.1	Odobenus rosmarus divergens	1,744	1,750	2.37 Gbp	1.36 Mbp	26,207	26,207	0
GCF_000767855.1	Camelus bactrianus	507	576	1.96 Gbp	3.87 Mbp	48,144	41,016	7,128
GCF_000952055.2	Aotus nancymaae	675	767	2.68 Gbp	3.97 Mbp	50,806	46,772	4,034
GCF_000956105.1	Propithecus coquereli	983	1,032	2.60 Gbp	2.64 Mbp	28,321	27,931	390
GCF_001458135.1	Marmota marmota	369	546	2.45 Gbp	6.63 Mbp	33,318	30,644	2,674
GCF_001604975.1	Cebus imitator	1,015	1,068	2.69 Gbp	2.65 Mbp	60,894	55,556	5,338
GCF_002201575.1	Neomonachus schauinslandi	206	373	2.37 Gbp	11.52 Mbp	28,615	27,845	770
GCF_002263795.3	Bos taurus	143	399	2.72 Gbp	19.04 Mbp	72,650	64,787	7,863
GCF_002288905.1	Enhydra lutris kenyon	212	397	2.42 Gbp	11.44 Mbp	37,044	36,345	699
GCF_002940915.1	Desmodus rotundus	139	290	2.04 Gbp	14.71 Mbp	47,688	44,227	3,461
GCF_003327715.1	Puma concolor	50	271	2.43 Gbp	48.55 Mbp	23,780	23,277	503
GCF_009806435.1	Oryctolagus cuniculus	852	1,069	2.73 Gbp	3.20 Mbp	74,055	62,581	11,474
GCF_016772045.2	Ovis aries	51	304	2.65 Gbp	51.99 Mbp	84,479	76,670	7,809
GCF_018350175.1	Felis catus	56	288	2.42 Gbp	43.30 Mbp	91,845	71,395	20,450
GCF_036323735.1	Rattus norvegicus	74	343	2.85 Gbp	38.51 Mbp	106,592	85,569	21,023
GCF_900095145.1	Mus pahari	70	299	2.43 Gbp	34.74 Mbp	42,684	41,492	1,192
hg38	Homo sapiens	21	295	2.83 Gbp	134.86 Mbp	254,620	83,442	171,178

Human validation and test references

Split	Genome	Species	Chromosomes	Blocks	Nucleotides	Mean chromosome length	Retained isoforms	mRNA/protein-coding	lncRNA
validation	hg38	Homo sapiens	3	27	256,292,786	85.43 Mbp	24,359	6,388	17,971
test	GCF_009914755.1_T2T-CHM13v2.0	Homo sapiens	24	327	3,117,275,501	129.89 Mbp	161,256	129,738	31,518

DNA alphabet

The DNA sequence is retained as provided by each reference assembly. Canonical bases are A, T, C, and G; unresolved assembly positions are represented as N.

Split	Total nucleotides	Canonical A/T/C/G	N nucleotides	N fraction
train	97,364,554,030	92,177,733,773	5,186,820,257	5.327216%
validation	256,292,786	248,801,012	7,491,774	2.923131%
test	3,117,275,501	3,117,275,501	0	0.000000%

Loading the dataset

from datasets import load_dataset

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

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

For sequential access without downloading and indexing the complete dataset first:

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

Inspect one row:

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"])

assert targets.shape == (len(dna), 12)

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

Selecting a target group

import numpy as np

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

# mRNA/protein-coding + lncRNA
combined_targets = targets[:, :6]

# mRNA/protein-coding only
mrna_only_targets = targets[:, 6:12]

Use combined_targets for joint mRNA/lncRNA gene finding. Use mrna_only_targets when lncRNA annotations should not contribute to the objective.

Filtering to one genome

import json

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

mouse = select_genome(train, "GCF_000001635.26")
human_hg38 = select_genome(train, "hg38")

Reconstructing a full chromosome

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

Example:

chr20_dna, chr20_targets = get_full_chromosome(
    validation,
    genome_id="hg38",
    chrom_name="chr20",
)

print(len(chr20_dna))
print(chr20_targets.shape)

For very large chromosomes, process blocks sequentially rather than concatenating all target arrays in memory.

Tokenization

The dataset stores raw DNA and nucleotide-level targets. It does not store model-specific token IDs.

For a BPE or another variable-length tokenizer, tokenize dna_sequence during training with offset mappings. A token-level target can be obtained by taking the maximum nucleotide-level value over the nucleotide interval covered by that token. Special and padding tokens should be masked from the loss.

Evaluation

Use:

• validation for model selection on held-out hg38 chromosomes
• test for final evaluation on the complete independent T2T human reference

The test split measures transfer from multispecies/hg38 training to a separate complete human genome assembly.