README.md

---
license: mit
task_categories:
- text-classification
tags:
- biology
- genomics
- long-context
papers: 2502.07272
configs:
- config_name: drug_resistence_prediction
  data_files:
  - split: train
    path: drug_resistence_prediction/train.parquet
  - split: test
    path: drug_resistence_prediction/test.parquet
- config_name: fitness_prediction_Ammonium-chloride_N
  data_files:
  - split: train
    path: fitness_prediction_Ammonium-chloride_N/train.parquet
  - split: test
    path: fitness_prediction_Ammonium-chloride_N/test.parquet
- config_name: fitness_prediction_Cisplatin_stress
  data_files:
  - split: train
    path: fitness_prediction_Cisplatin_stress/train.parquet
  - split: test
    path: fitness_prediction_Cisplatin_stress/test.parquet
- config_name: fitness_prediction_D-Alanine_N
  data_files:
  - split: train
    path: fitness_prediction_D-Alanine_N/train.parquet
  - split: test
    path: fitness_prediction_D-Alanine_N/test.parquet
- config_name: fitness_prediction_L-Arabinose_C
  data_files:
  - split: train
    path: fitness_prediction_L-Arabinose_C/train.parquet
  - split: test
    path: fitness_prediction_L-Arabinose_C/test.parquet
- config_name: fitness_prediction_L-Histidine_nutrient
  data_files:
  - split: train
    path: fitness_prediction_L-Histidine_nutrient/train.parquet
  - split: test
    path: fitness_prediction_L-Histidine_nutrient/test.parquet
- config_name: fitness_prediction_LB-10C
  data_files:
  - split: train
    path: fitness_prediction_LB-10C/train.parquet
  - split: test
    path: fitness_prediction_LB-10C/test.parquet
- config_name: fitness_prediction_LB-20C
  data_files:
  - split: train
    path: fitness_prediction_LB-20C/train.parquet
  - split: test
    path: fitness_prediction_LB-20C/test.parquet
- config_name: fitness_prediction_LB-30C
  data_files:
  - split: train
    path: fitness_prediction_LB-30C/train.parquet
  - split: test
    path: fitness_prediction_LB-30C/test.parquet
- config_name: fitness_prediction_LB-pH6
  data_files:
  - split: train
    path: fitness_prediction_LB-pH6/train.parquet
  - split: test
    path: fitness_prediction_LB-pH6/test.parquet
- config_name: fitness_prediction_LB-pH8
  data_files:
  - split: train
    path: fitness_prediction_LB-pH8/train.parquet
  - split: test
    path: fitness_prediction_LB-pH8/test.parquet
- config_name: fitness_prediction_Min-media-glucose_nutrient
  data_files:
  - split: train
    path: fitness_prediction_Min-media-glucose_nutrient/train.parquet
  - split: test
    path: fitness_prediction_Min-media-glucose_nutrient/test.parquet
- config_name: fitness_prediction_Pyruvate_C
  data_files:
  - split: train
    path: fitness_prediction_Pyruvate_C/train.parquet
  - split: test
    path: fitness_prediction_Pyruvate_C/test.parquet
- config_name: fitness_prediction_perchlorate_stress
  data_files:
  - split: train
    path: fitness_prediction_perchlorate_stress/train.parquet
  - split: test
    path: fitness_prediction_perchlorate_stress/test.parquet
- config_name: gene_classification_bacteria
  data_files:
  - split: train
    path: gene_classification_bacteria/train.parquet
  - split: test
    path: gene_classification_bacteria/test.parquet
- config_name: taxonomic_classification_mixed_marker
  data_files:
  - split: train
    path: taxonomic_classification_mixed_marker/train.parquet
  - split: test
    path: taxonomic_classification_mixed_marker/test.parquet
- config_name: taxonomic_classification_random_fragment
  data_files:
  - split: train
    path: taxonomic_classification_random_fragment/train.parquet
  - split: test
    path: taxonomic_classification_random_fragment/test.parquet
- config_name: taxonomic_classification_ssu
  data_files:
  - split: train
    path: taxonomic_classification_ssu/train.parquet
  - split: test
    path: taxonomic_classification_ssu/test.parquet
---

# Prokaryotic Gener Tasks

[**Paper**](https://arxiv.org/abs/2502.07272) | [**GitHub**](https://github.com/generteam/generator)

Prokaryotic Gener Tasks is a suite of biologically meaningful benchmark tasks in the prokaryotic domain, introduced as part of the **GENERator** project. GENERator is a generative genomic foundation model for long-context DNA modeling, pre-trained on expansive DNA datasets derived from the RefSeq database.

## Dataset Summary

This collection includes various downstream tasks designed to evaluate the sequence understanding (classification and regression) capabilities of genomic foundation models in the prokaryotic domain. The benchmark includes:

- **Drug Resistance Prediction**: Predicting the resistance profiles of prokaryotic sequences to various drugs.
- **Fitness Prediction**: Estimating the biological fitness of organisms under diverse environmental conditions, including temperature variations (LB-10C, LB-20C, LB-30C), pH levels (pH6, pH8), and nutrient stresses (Ammonium-chloride, Pyruvate, etc.).
- **Gene Classification**: Classification of genes within the bacterial domain.
- **Taxonomic Classification**: Classifying sequences into phylogenetic categories using various markers such as Small Subunit (SSU) sequences, random fragments, or mixed markers.

## Usage

To evaluate a model on these tasks using the official [GENERator implementation](https://github.com/generteam/generator), you can use the following command structure:

```shell
# Using single GPU for sequence understanding
python src/tasks/downstream/sequence_understanding.py \
    --model_name GenerTeam/GENERator-eukaryote-1.2b-base \
    --dataset_name GenerTeam/prokaryotic-gener-tasks \
    --subset_name gene_classification_bacteria \
    --batch_size 32 \
    --problem_type classification \
    --main_metrics accuracy
```

Replace `--subset_name` with one of the configurations (e.g., `taxonomic_classification_ssu` or `fitness_prediction_LB-10C`) and adjust `--problem_type` (classification or regression) as appropriate.

## Citation

```bibtex
@misc{wu2025generator,
      title={GENERator: A Long-Context Generative Genomic Foundation Model}, 
      author={Wei Wu and Qiuyi Li and Mingyang Li and Kun Fu and Fuli Feng and Jieping Ye and Hui Xiong and Zheng Wang},
      year={2025},
      eprint={2502.07272},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2502.07272}, 
}
```