GenerTeam/pretrain_data_eukaryote

GENERator-v2-Eukaryote Gene-Centric Pretraining Corpus

This repository provides the gene-centric pretraining corpus underlying GENERator-v2-Eukaryote, a large-scale DNA language model for eukaryotic genome understanding.

The dataset is constructed by leveraging RefSeq annotations to extract biologically meaningful functional genomic regions, which serve as the foundation for large-context DNA language model pretraining.

---

📌 Dataset Construction Overview

The core design philosophy of this dataset is gene-centric functional sequence modeling.

High-confidence reference annotations (e.g. RefSeq) are used as a scaffold to identify and extract contiguous functional regions from eukaryotic genomes, including protein-coding genes and diverse RNA genes.

---

🧬 Data Schema

Each row in the dataset corresponds to one functional genomic segment.

Column	Type	Description
record_id	string	RefSeq record identifier
taxonomy	string	Full taxonomic lineage (semicolon-separated)
species_type	string	High-level species category token
gene_type	string	Functional gene category token
strand	string	DNA strand in the reference genome (<+> or <->)
sequence	string	Extracted functional DNA sequence
start	int	Start coordinate of the functional region on the RefSeq record
end	int	End coordinate of the functional region on the RefSeq record

---

🌍 Species Type Tokens (species_type)

Each sample is annotated with a coarse-grained evolutionary category:

Token	Meaning
<prt>	Protozoa
<fng>	Fungi
<pln>	Plant
<inv>	Invertebrate
<vrt>	Vertebrate (non-mammalian)
<mam>	Vertebrate (mammalian)

---

🧠 Gene Type Tokens (gene_type)

Functional regions are categorized as follows:

Token	Description
<cds>	Protein-coding gene (gene-centric region, not limited to CDS only)
<pseudo>	Pseudogene
<tRNA>	Transfer RNA gene
<rRNA>	Ribosomal RNA gene
<ncRNA>	Non-coding RNA
<misc_RNA>	RNA genes not assigned to a specific class

---

🔁 Strand Orientation

• <+> denotes the positive strand
• <-> denotes the negative strand in the reference genome

---

🔬 Sequence Characteristics

• Raw DNA sequences (A/C/G/T/N)
• Uppercase encoding
• N denotes ambiguous nucleotides
• No tokenization, masking, or augmentation is applied at this stage

This representation preserves maximum flexibility for downstream preprocessing and modeling strategies.

---

🚀 Intended Use

This dataset is designed to support:

• Large-scale DNA language model pretraining
• Gene-centric functional sequence modeling
• Cross-species and cross-gene-type representation learning
• Research in comparative and functional genomics

---

🧪 Relationship to GENERator-v2-Eukaryote Training

This repository provides raw functional sequence data.

The actual pretraining pipeline of GENERator-v2-Eukaryote applies additional post-processing steps, including:

• Sequence length filtering (<100k)
• Sequence concatenation and segmentation
• Tokenization and phase augmentation

These steps are not applied in this dataset and are described in detail in the GENERator-v2 Technical Report.

---

🔮 Future Data Releases

The training corpus for GENERator-v2-Prokaryote is currently under active evaluation and optimization.
We plan to release the corresponding prokaryotic pretraining data after thorough validation of data quality and downstream performance.

In addition, the GENERanno series of genome annotation datasets, covering both eukaryotic and prokaryotic genomes at substantially larger scale, will be made publicly available in future releases.

Please stay tuned for updates.

---

🔗 Related Resources

For more information about the GENERator family of models and ongoing developments, please visit our GitHub repository:

👉 https://github.com/GenerTeam/

---

📝 Citation

@article {li2026generator2,
    author = {Li, Qiuyi and Zhan, Zhihao and Feng, Shikun and Zhu, Yiheng and He, Yuan and Wu, Wei and Shi, Zhenghang and Wang, Shengjie and Hu, Zongyong and Yang, Zhao and Li, Jiaoyang and Tang, Jian and Liu, Haiguang and Qin, Tao},
    title = {GENERator-v2: Reconciling Coarse Tokenization with Single-Nucleotide Resolution in Genomic Language Modeling},
    elocation-id = {2026.01.27.702015},
    year = {2026},
    doi = {10.64898/2026.01.27.702015},
    publisher = {Cold Spring Harbor Laboratory},
    URL = {https://www.biorxiv.org/content/early/2026/05/04/2026.01.27.702015},
    journal = {bioRxiv}
}

@article{wu2025generator,
  title={GENERator: a long-context generative genomic foundation model},
  author={Wu, Wei and Li, Qiuyi and Li, Mingyang and Fu, Kun and Feng, Fuli and Ye, Jieping and Xiong, Hui and Wang, Zheng},
  journal={arXiv preprint arXiv:2502.07272},
  year={2025}
}