README.md

metadata

license: apache-2.0
pretty_name: OpenGenome2
tags:
  - huggingscience
  - science
  - biology
  - genomics
  - dna
  - sequence-modeling
  - long-context
  - metagenomics
  - microbiology
  - bacteriophage
  - plasmid
  - mrna
  - promoter
  - ncrna
  - organelle-genomes
  - gtdb
  - imgvr
  - imgpr
  - ncbi
  - json
  - fasta

OpenGenome2

OpenGenome2 is a large-scale, diverse DNA corpus spanning prokaryotes (bacteria & archaea), eukaryotes, organelles, bacteriophages, plasmids, and metagenomes. It was assembled to train long-context genomic foundation models (e.g., Evo 2) and emphasizes functionally informative regions (e.g., genic windows, promoters, noncoding RNAs) while providing comprehensive whole-genome coverage across the tree of life.

• License: Apache-2.0
• Scope: trillions of nucleotides aggregated across multiple public sources (e.g., GTDB v220, IMG/VR, IMG/PR, NCBI eukaryotic references) with curation and quality filtering to support genome-scale modeling.
• Formats: JSONL (token-ready sequences) and FASTA (raw sequences).
• Splits: train, validation, test are provided for most JSONL subsets.

Contents & Source Mapping

The on-hub layout mirrors data provenance and training phase:

• json/pretraining_or_both_phases/ Functional and broad-coverage sources used during pretraining (or used in both pre- and mid-training):

  • eukaryotic_genic_windows/ – windows enriched for coding/genic content
  • gtdb_v220_imgpr/ – prokaryotes from GTDB v220 plus curated plasmids/phages from IMG/PR
  • imgvr_untagged/ – bacteriophage sequences from IMG/VR
  • metagenomes/ – metagenomic contigs
  • mrna/, mrna_splice_promoter/ – mRNA and splicing/promoter-focused windows
  • ncrna/ – noncoding RNAs
  • organelle/ – organellar genomes
  • promoters/ – promoter regions

• json/midtraining_specific/ Additional curated data emphasized during mid-training:

  • gtdb_v220_stitched/ – stitched windows / references from GTDB v220
  • imgpr/ – curated plasmids & phages from IMG/PR
  • ncbi_eukaryotic_genomes/ – curated eukaryotic references from NCBI

• fasta/ mirrors these sources for users preferring raw FASTA (e.g., fasta/gtdb_v220, fasta/mrna, fasta/organelles, fasta/promoters, etc.).

Recommended Uses

• Pretraining/continued-pretraining DNA language models (autoregressive or masked).
• Long-context modeling (100 kbp–1 Mbp) for regulatory architecture and genome context tasks.
• Zero-shot variant effect predictions on promoters/genic windows.
• Embedding-based functional clustering across genomes, plasmids, and phages.
• Generative modeling for protein–DNA/RNA system design when coupled with appropriate safety review.

Citation

If you use OpenGenome2 or models trained on it, please cite:

• Evo 2 (trained on OpenGenome2): Brixi et al., “Genome modeling and design across all domains of life with Evo 2,” bioRxiv (2025). https://doi.org/10.1101/2025.02.18.638918

• Data sources:

  • GTDB updated references (v220)
  • IMG/VR (bacteriophages) and IMG/PR (plasmids/phages)
  • NCBI eukaryotic genomes/references

Attribution

• Dataset maintained by the Arc Institute. See the Hub repo for updates and discussions.