Genomic Corpus · tattabio
OMG
A 3.1-trillion base pair mixed-modality metagenomic corpus for pretraining genomic language models.
What OMG is
Open Mixed Genomes (OMG) is a large-scale metagenomic corpus containing 3.1 trillion base pairs of sequence data. Developed by Tatta Bio, it is designed for pretraining mixed-modality genomic foundation models, such as the gLM2 model family.
The dataset integrates sequence records from two major public repositories: EMBL's MGnify database and the Joint Genome Institute's (JGI) Integrated Microbial Genomes (IMG) system. By merging these sources into a single corpus, it provides extensive representation of diverse uncultured and environmental microbial genomes.
Rather than treating genomes purely as raw DNA sequences, OMG structures genomic scaffolds into a mixed modality. Protein coding regions are provided as translated amino acid sequences, while intergenic regions are retained as nucleotide sequences, preserving both functional protein sequence information and regulatory or non-coding context.
How it was made
OMG standardises metagenomic scaffolds by splitting each scaffold into an ordered progression of coding sequences (CDS) and intergenic sequences (IGS). Protein-coding segments are translated into amino acid sequences, whereas the intervening non-coding regions are recorded as nucleotide sequences.
Each entry represents a single scaffold containing sequence lists alongside positional indices (`CDS_position_ids` and `IGS_position_ids`) to preserve their relative physical order. Metadata identifiers detail the sample accession, contig ID, feature type, gene ID, strand, and start-to-end coordinates, supplemented by boolean flags indicating CDS strand orientation.
Key facts
- Total sequence length
- 3.1 trillion base pairs
- Underlying sources
- EMBL MGnify and JGI IMG
- Data format
- Parquet
- Target models
- gLM2 genomic foundation models
- Primary modalities
- Amino acid sequences (CDS) and nucleotide sequences (IGS)
- Scaffold features
- CDS_seqs, IGS_seqs, CDS_position_ids, IGS_position_ids, CDS_ids, IGS_ids, CDS_orientations
- Release year
- 2024
What you can use it for
- Genomic Foundation Model Pretraining
- Train large language models across interleaved protein-coding and intergenic regions across diverse metagenomic assemblies.
- Metagenomic Gene Context Modelling
- Investigate gene synteny and contextual organisation using explicitly ordered amino acid coding sequences and intergenic spacers.
- Comparative Microbial Genomics
- Analyse wide-scale microbial diversity derived jointly from MGnify and IMG metagenomic data collections.
Explore the data
The schema, real rows, full-text search and a SQL filter — read in your browser, nothing downloaded.
Explore
Read the real rows without downloading anything
Reading rows…
LiveRead from tattabio/OMG at the moment you asked. Nothing is cached or stored — every row above came from that request.
How the columns are distributed
Over 846,636 rows of default/train
CDS_ids
list- min
- 4
- median
- 6
- mean
- 12.33
- max
- 616
CDS_orientations
list- min
- 4
- median
- 6
- mean
- 12.33
- max
- 616
CDS_position_ids
list- min
- 4
- median
- 6
- mean
- 12.33
- max
- 616
CDS_seqs
list- min
- 4
- median
- 6
- mean
- 12.33
- max
- 616
IGS_ids
list- min
- 0
- median
- 6
- mean
- 10.56
- max
- 486
IGS_position_ids
list- min
- 0
- median
- 6
- mean
- 10.56
- max
- 486
IGS_seqs
list- min
- 0
- median
- 6
- mean
- 10.56
- max
- 486
Strengths and limits
Good at
- Combines two major metagenomic resources, EMBL MGnify and JGI IMG, into a cohesive corpus.
- Employs a mixed-modality representation that preserves translated protein sequences alongside intervening nucleotide regions.
- Preserves positional order and strand orientation across entire genomic scaffolds.
Watch out for
- Intergenic sequences are restricted to raw nucleotide strings and lack functional annotations in the primary schema.
- Processing the full 3.1-trillion base pair dataset requires significant computational and storage capacity.
Questions
Can I look inside OMG without downloading it?
Yes. This page reads its schema, real rows and per-column statistics in the browser, with full-text search and a SQL filter. Nothing is downloaded to your machine.
What licence is OMG under?
Its authors declare cc-by-sa-4.0. Read the terms on the original page before reusing the data.
How are genomic sequences represented in the dataset?
Each row corresponds to a genomic scaffold where protein-coding sequences (CDS) are translated into amino acid strings and intergenic regions (IGS) are kept as nucleotide strings, accompanied by relative positions and strand orientation booleans.
What is the identifier format for CDS and IGS entries?
Identifiers follow the pipe-delimited structure: sample_accession|contig_id|feature_type|gene_id|strand|start:end.
Are there derived or smaller subsets of OMG available?
Yes. Tatta Bio provides OG, which consists of high-quality genomes with taxonomic information, and OMG_prot50, a protein-only corpus of 207 million sequences clustered at 50% sequence identity.
How to cite it
Cornman, Andre and West-Roberts, Jacob and Camargo, Antonio Pedro and Roux, Simon and Beracochea, Martin and Mirdita, Milot and Ovchinnikov, Sergey and Hwang, Yunha. The OMG dataset: An Open MetaGenomic corpus for mixed-modality genomic language modeling. bioRxiv, 2024. DOI: 10.1101/2024.08.14.607850
Written from OMG's own dataset card. Statements citing a figure that does not appear there are removed before publication. Licence and running details come from our registry.