metadata

tags:
  - biology
  - DNA
  - genomics

Pretraining dataset

For pretraining and validation, we curated a dataset comprising all 19,551 representative genomes sourced from the National Center for Biotechnology Information (NCBI) database by February 23, 2024, including 18,268 bacterial, 647 archaeal, 577 fungal, 40 viral, and 1 human reference genome. Then, to create samples, we created chunks of 3200 long nucleotide sequences with 100 bases overlapping. The resulting train set has 27,831,882 samples, the validation set has 3,478,985, and the test set has 3,478,986 samples.

Dataset Structure

Each dataset sample contains the following information:

assembly_accession: Genome assembly identifier
seq_id: Accession identifier
assembly_id: Line number in the file
start: Start position of the sequence read within the accession ID
sequence: DNA sequence data
sequence_length: Length of the DNA sequence sample
split: Dataset split category