Scorpio-Gene-Taxa Dataset: Curated Gene Sequences for Model Training and Evaluation

Protein and DNA Sequences Available for Evaluating DNA and Protein Language Models for Gene and Taxonomy Prediction

This dataset was created using the Woltka pipeline to compile the Basic Genome Dataset, containing 4,634 genomes. Each genus is represented by a single genome, with a focus on bacteria and archaea. Viruses and fungi were excluded due to insufficient gene information.

Key steps in data preparation:

1. Gene names were used to filter and group protein-coding sequences; unnamed genes with hypothetical or unknown functions were excluded.
2. Only genes with more than 1,000 named instances (497 genes) were included to improve reliability for training.

Dataset Splits (dataset_type)

Train

• Description: Contains sequences used for model training.

Test

• Description: Contains sequences from the same genus and gene as the training set but uses different sequences.
• Purpose: Evaluates sequence-level generalization.

Taxa_out

• Description: Excludes sequences from 18 phyla present in the training set.
• Purpose: Evaluates gene-level generalization.

Gene_out

• Description: Excludes 60 genes present in the training set but keeps sequences within the same phyla.
• Purpose: Evaluates taxonomic-level generalization.

Citation

@article{refahi2024scorpio,
  title={Scorpio: Enhancing Embeddings to Improve Downstream Analysis of DNA sequences},
  author={Refahi, Mohammadsaleh and Sokhansanj, Bahrad A and Mell, Joshua Chang and Brown, James and Yoo, Hyunwoo and Hearne, Gavin and Rosen, Gail},
  journal={bioRxiv},
  pages={2024--07},
  year={2024},
  publisher={Cold Spring Harbor Laboratory}
}