Blame for README.md

---
language:
  - en
pretty_name: "NTv3 annotation dataset - Genomes & reference gene annotations"
tags:
  - 🧬 genomics
  - 🏷️ gene annotation
  - 🎯 benchmark
  - ⚡ fine-tuning
  - 🔬 bioinformatics
task_categories:
  - other
size_categories:
  - n<1K
license: cc-by-nc-4.0
---

# NTv3 Annotation Dataset

Genomes and reference gene annotations behind the NTv3 gene-annotation models
([`NTv3_annot_mammals`](https://huggingface.co/InstaDeepAI/NTv3_annot_mammals),
[`NTv3_annot_plants`](https://huggingface.co/InstaDeepAI/NTv3_annot_plants),
[`NTv3_annot_insects`](https://huggingface.co/InstaDeepAI/NTv3_annot_insects)). It holds two
things:

- **The benchmark**: the 17 test species held out from training (2 mammals, 5 plants,
  10 insects). Each species comes with its complete soft-masked genome, the reference
  annotation the benchmark scores against, and the reference's every annotated isoform.
  The models' own predictions on these genomes are in each model repository's
  `annot_results/` folder.
- **The plant tutorial data**: training and validation genomes with per-base labels, used by
  the genome-annotation notebooks of the
  [`ntv3_revised`](https://huggingface.co/spaces/InstaDeepAI/ntv3_revised) Space
  (03a training, 03b decoding, 03c evaluation).

## Dataset organization

```
InstaDeepAI/NTv3_annot_dataset/
├── <clade>/test_species/<assembly>/     # clade = mammals, plants or insects
│   ├── genome.fa.gz                     # soft-masked genome, bgzipped
│   ├── genome.fa.gz.fai                 # samtools faidx index
│   ├── genome.fa.gz.gzi                 # bgzip index
│   ├── reference_longest_isoform.gff3.gz
│   ├── reference_all_isoforms.gff3.gz
│   └── pseudogenes.bed
└── plants/
    ├── metadata.tsv                     # one row per tutorial species
    ├── class_counts.json
    ├── class_weights.json
    ├── train_species/<assembly>/        # genome.fa.gz(.fai, .gzi), labels.bed.gz
    ├── val_species/<assembly>/          # same as test species, plus labels.bed.gz,
    │                                    #   without reference_all_isoforms.gff3.gz
    └── test_species/<assembly>/         # Medicago and Mimulus also carry labels.bed.gz;
                                         #   Medicago carries demo_log_probs/
```

Folders are named by assembly, as in the model repositories' `annot_results/` files.

## Files

- **`genome.fa.gz`**: the soft-masked genome (repeats in lowercase), bgzipped, one record
  per sequence, record names cut to the sequence ID. For test species it is the complete
  assembly, every contig, with the sequence names of the reference annotation. Tutorial
  training and validation genomes keep only the contigs of at least 131,072 bp, the
  tutorial's training window.
- **`reference_longest_isoform.gff3.gz`**: the reference the benchmark scores against.
  Protein-coding genes only, one transcript per gene (the one with the longest CDS),
  `gene`, `mRNA` and `CDS` rows, `ID` and `Parent` attributes only. 1-based closed
  coordinates (GFF3).
- **`reference_all_isoforms.gff3.gz`** (test species): the same genes, with every
  transcript the source annotation release gives them, in the same format. A prediction
  that matches any transcript of a gene counts under the any-isoform rule. When the source
  release does not carry a reference transcript's exact CDS, that transcript is kept as
  well, so the file is a strict superset of the longest-isoform reference: 171 genes whose
  IDs were assigned when the longest isoform was extracted (`agat-gene-*`: 151 immune
  receptor gene segments in cow and whale, 15 chloroplast genes, 5 insect genes) and one
  transcript in cow (*PEG10*, ID suffixed `.longest_isoform`).
- **`pseudogenes.bed`**: pseudogene regions of the source annotation that do not overlap a
  reference CDS. Predicted genes there are ignored when scoring. BED, 0-based half-open.
- **`labels.bed.gz`** (tutorial species): the per-base training labels, one of 15 classes
  (`intergenic`, `intron_p0-2`, `CDS_p0-2`, `splice_donor_p0-2`, `splice_acceptor_p0-2`,
  `start_codon`, `stop_codon`), as BED runs per strand (`chrom`, `start`, `end`, `class`,
  `.`, `strand`), 0-based half-open.
- **`plants/metadata.tsv`**: the tutorial species, their role (`train`, `validation`,
  `test`), tier (the tutorial's presets train on tiers 1-2), source, accession and the
  chromosome the tutorial evaluates on (`eval_chromosomes`).
- **`plants/class_counts.json`**: per-species base counts of each class. Notebook 03a
  derives its loss weights from them, for the genomes it trains on.
- **`plants/class_weights.json`**: those weights precomputed over all the training genomes
  (mean inverse class frequency).
- **`demo_log_probs/`** (Medicago): precomputed per-base log-probabilities of
  `NTv3_annot_plants` over `NC_053042.1:11,010,048-12,582,912`, decoded in notebook 03b;
  `meta.json` records the model, code revision, windows and hardware.

Scores are gene- and CDS-level F1: a predicted gene counts only if its whole CDS structure
matches exactly, predicted genes with less than 201 nt of CDS are ignored, and so are those
in `pseudogenes.bed`. Notebook 03c implements the evaluation.

## Test species

| Clade | Folder | Species | Contigs | Genes | Transcripts | Genome | Reference annotation |
|---|---|---|---:|---:|---:|---|---|
| mammals | `GCF_002263795.3_ARS-UCD2.0` | *Bos taurus* (cow) | 1,958 | 21,577 | 49,639 | NCBI RefSeq | NCBI RefSeq GCF_002263795.3-RS_2024_12 |
| mammals | `GCF_002288925.2_ASM228892v3` | *Delphinapterus leucas* (beluga whale) | 5,905 | 18,389 | 40,864 | NCBI RefSeq | NCBI RefSeq Annotation Release 101 |
| plants | `GCF_000005505.3_Brachypodium_distachyon_v3.0` | *Brachypodium distachyon* | 11 | 25,477 | 32,269 | NCBI RefSeq | NCBI RefSeq Annotation Release 103 |
| plants | `EcalifornicavarOrangeKingHAP1v1.1` | *Eschscholzia californica* 'Orange King' HAP1 | 6 | 29,562 | 42,970 | Phytozome 828 | Phytozome v1.1 |
| plants | `FmultifloravarABG20040013HAP1v1.1` | *Freycinetia multiflora* ABG20040013 HAP1 | 30 | 28,264 | 38,805 | Phytozome 868 | Phytozome v1.1 |
| plants | `GCF_003473485.1_MtrunA17r5.0-ANR` | *Medicago truncatula* | 42 | 31,794 | 37,660 | NCBI RefSeq | NCBI RefSeq GCF_003473485.1-RS_2025_04 |
| plants | `MguttatusvarIM62v3.1` | *Mimulus guttatus* IM62 | 23 | 25,096 | 31,078 | Phytozome 797 | Phytozome v3.1 |
| insects | `GCF_030269925.1_ASM3026992v2` | *Bombyx mori* | 30 | 13,251 | 22,941 | NCBI RefSeq | NCBI RefSeq GCF_030269925.1-RS_2024_01 |
| insects | `GCF_021464435.1_ULB_Chis1_1.0` | *Cataglyphis hispanica* | 453 | 10,380 | 17,463 | NCBI RefSeq | NCBI RefSeq Annotation Release 100 |
| insects | `GCF_905220415.1_ilColCroc2.1` | *Colias croceus* | 32 | 13,364 | 18,775 | NCBI RefSeq | NCBI RefSeq Annotation Release 100 |
| insects | `GCF_018135715.1_MEX_DaPlex` | *Danaus plexippus* | 66 | 12,669 | 19,058 | NCBI RefSeq | NCBI RefSeq GCF_018135715.1-RS_2023_11 |
| insects | `GCF_905404315.1_ilLepSina1.1` | *Leptidea sinapis* | 49 | 12,276 | 19,563 | NCBI RefSeq | NCBI RefSeq Annotation Release 100 |
| insects | `GCF_905147045.1_ilAglIoxx1.1` | *Nymphalis io* (syn. *Aglais io*) | 41 | 12,024 | 18,680 | NCBI RefSeq | NCBI RefSeq Annotation Release 100 |
| insects | `GCF_907164935.1_iOsmBic2.1` | *Osmia bicornis* | 441 | 11,262 | 18,822 | NCBI RefSeq | NCBI RefSeq Annotation Release 101 |
| insects | `GCF_031307605.1_icTriCast1.1` | *Tribolium castaneum* | 149 | 11,953 | 19,088 | NCBI RefSeq | NCBI RefSeq GCF_031307605.1-RS_2024_04 |
| insects | `GCF_905220365.1_ilVanCard2.1` | *Vanessa cardui* | 36 | 13,070 | 17,975 | NCBI RefSeq | NCBI RefSeq Annotation Release 100 |
| insects | `GCF_012273895.1_Zerene_cesonia_1.1` | *Zerene cesonia* | 275 | 11,791 | 15,989 | NCBI RefSeq | NCBI RefSeq Annotation Release 100 |

*Genes* is the number of reference genes; *Transcripts*, the coding transcripts in
`reference_all_isoforms.gff3.gz`. NCBI genomes are the `_genomic.fna.gz` of the assembly;
the Phytozome genomes are its `softmasked` FASTA (Freycinetia: the 30 scaffolds of
Phytozome 868 concatenated).

## Plant tutorial species

| Role | Folder | Species | Source |
|---|---|---|---|
| train (tier 1) | `TAIR10` | *Arabidopsis thaliana* | Ensembl Plants (Araport11 genes) |
| train (tier 1) | `GCF_000004075.3_Cucumber_9930_V3` | *Cucumis sativus* | NCBI RefSeq |
| train (tier 1) | `GCF_022201045.2_DVS_A1.0` | *Citrus sinensis* | NCBI RefSeq |
| train (tier 1) | `GCF_019578655.1_ASM1957865v1` | *Ricinus communis* | NCBI RefSeq |
| train (tier 1) | `GCF_020921035.1_ASM2092103v1` | *Rhodamnia argentea* | NCBI RefSeq |
| train (tier 1) | `GCF_005286985.2_Setaria_viridis_v4.0` | *Setaria viridis* | NCBI RefSeq |
| train (tier 1) | `GCF_016808095.1_ASM1680809v1` | *Vigna angularis* | NCBI RefSeq |
| train (tier 1) | `GCF_030704535.1_ASM3070453v1` | *Vitis vinifera* | NCBI RefSeq |
| train (tier 2) | `GCF_031755915.1_ASM3175591v1` | *Ziziphus jujuba* | NCBI RefSeq |
| train (tier 2) | `GCF_018687715.1_C.illinoinensisPawnee_v1` | *Carya illinoinensis* | NCBI RefSeq |
| train (tier 2) | `GCF_036884655.1_Cavendish_Baxijiao_AAA` | *Musa acuminata* | NCBI RefSeq |
| train (tier 2) | `GCF_958298935.1_lpPhrAust1.1` | *Phragmites australis* | NCBI RefSeq |
| train (tier 2) | `IRGSP-1.0` | *Oryza sativa* | Ensembl Plants genome, NCBI RefSeq genes |
| validation | `GCF_000002775.5_P.trichocarpa_v4.1` | *Populus trichocarpa* | NCBI RefSeq |
| validation | `Tlatifoliav1.1` | *Typha latifolia* | JGI (GenBank copy, GCA_059706395.1) |
| test | `GCF_003473485.1_MtrunA17r5.0-ANR` | *Medicago truncatula* | see the test species table |
| test | `MguttatusvarIM62v3.1` | *Mimulus guttatus* IM62 | see the test species table |

`plants/metadata.tsv` also lists tier 3-4 species; their files are not in this dataset.

## Download

```python
from huggingface_hub import hf_hub_download, snapshot_download

# One file
path = hf_hub_download(
    "InstaDeepAI/NTv3_annot_dataset",
    "mammals/test_species/GCF_002288925.2_ASM228892v3/reference_longest_isoform.gff3.gz",
    repo_type="dataset",
)

# One species
local_dir = snapshot_download(
    "InstaDeepAI/NTv3_annot_dataset",
    repo_type="dataset",
    allow_patterns="plants/test_species/GCF_003473485.1_MtrunA17r5.0-ANR/*",
)
```

## Sources and licence

The files we derived (labels, class statistics, log-probabilities) are
released under CC BY-NC 4.0. Genomes and reference annotations are redistributed from NCBI
RefSeq, Ensembl Plants and JGI Phytozome under their own terms; please cite the original
assemblies and annotations. The three Phytozome genomes are listed as unrestricted:

- *Mimulus guttatus* IM62 (assembly v3.0, annotation v3.1): Lovell et al., *Molecular
  Ecology Resources* (2025), https://doi.org/10.1111/1755-0998.70012
- *Eschscholzia californica* 'Orange King' HAP1 v1.1: Rössner et al., *The Plant Cell*
  (2026), https://doi.org/10.1093/plcell/koag039
- *Freycinetia multiflora* ABG20040013 HAP1 (assembly v1.0, annotation v1.1): JGI Open
  Green Genomes, https://doi.org/10.25585/1488189

## Citation

Citation to be added on publication.