In this repository, we collect benchmarks for classification of genomic sequences. It is shipped as a Python package, together with functions helping to download & manipulate datasets and train NN models. Current SOTA model on genomic benchmarks is HyenaDNA, see metrics in the experiments folder.
Genomic Benchmarks can be installed as follows:
pip install genomic-benchmarksTo use it with papermill, TF or pytorch, install the corresponding dependencies:
# if you want to use jupyter and papermill
pip install jupyter>=1.0.0
pip install papermill>=2.3.0
# if you want to train NN with TF
pip install tensorflow>=2.6.0
pip install tensorflow-addons
pip install typing-extensions --upgrade # fixing TF installation issue
# if you want to train NN with torch
pip install torch>=1.10.0
pip install torchtext
For the package development, use Python 3.8 (ideally 3.8.9) and the installation described here.
Get the list of all datasets with the list_datasets function
>>> from genomic_benchmarks.data_check import list_datasets
>>>
>>> list_datasets()
['demo_coding_vs_intergenomic_seqs', 'demo_human_or_worm', 'dummy_mouse_enhancers_ensembl', 'human_enhancers_cohn', 'human_enhancers_ensembl', 'human_ensembl_regulatory', 'human_nontata_promoters', 'human_ocr_ensembl']You can get basic information about the benchmark with info function:
>>> from genomic_benchmarks.data_check import info
>>>
>>> info("human_nontata_promoters", version=0)
Dataset `human_nontata_promoters` has 2 classes: negative, positive.
All lenghts of genomic intervals equals 251.
Totally 36131 sequences have been found, 27097 for training and 9034 for testing.
train test
negative 12355 4119
positive 14742 4915The function download_dataset downloads the full-sequence form of the required benchmark (splitted into train and test sets, one folder for each class). If not specified otherwise, the data will be stored in .genomic_benchmarks subfolder of your home directory. By default, the dataset is obtained from our cloud cache (use_cloud_cache=True).
>>> from genomic_benchmarks.loc2seq import download_dataset
>>>
>>> download_dataset("human_nontata_promoters", version=0)
Downloading 1VdUg0Zu8yfLS6QesBXwGz1PIQrTW3Ze4 into /home/petr/.genomic_benchmarks/human_nontata_promoters.zip... Done.
Unzipping...Done.
PosixPath('/home/petr/.genomic_benchmarks/human_nontata_promoters')Getting TensorFlow Dataset for the benchmark and displaying samples is straightforward:
>>> from pathlib import Path
>>> import tensorflow as tf
>>>
>>> BATCH_SIZE = 64
>>> SEQ_TRAIN_PATH = Path.home() / '.genomic_benchmarks' / 'human_nontata_promoters' / 'train'
>>> CLASSES = ['negative', 'positive']
>>>
>>> train_dset = tf.keras.preprocessing.text_dataset_from_directory(
... directory=SEQ_TRAIN_PATH,
... batch_size=BATCH_SIZE,
... class_names=CLASSES)
Found 27097 files belonging to 2 classes.
>>>
>>> list(train_dset)[0][0][0]
<tf.Tensor: shape=(), dtype=string, numpy=b'TCCTGCCTTTCCACTTGCACCAGTTTTCCCACCCCAGCCTCAGGGCGGGGCTGCCTCGTCACTTGTCTCGGGGCAGATCTGCCCTACACACGTTAGCGCCGCGCGCAAAGCAGCCCCGCAGCACCCAGGCGCCTCCTGGCGGCGCCGCGAAGGGGCGGGGCTGTCGGCTGCGCGTTGTGCGCTGTCCCAGGTTGGAAACCAGTGCCCCAGGCGGCGAGGAGAGCGGTGCCTTGCAGGGATGCTGCGGGCGG'>See How_To_Train_CNN_Classifier_With_TF.ipynb for more detailed description how to train CNN classifier with TensorFlow.
Getting Pytorch Dataset and displaying samples is also easy:
>>> from genomic_benchmarks.dataset_getters.pytorch_datasets import HumanNontataPromoters
>>>
>>> dset = HumanNontataPromoters(split='train', version=0)
>>> dset[0]
('CAATCTCACAGGCTCCTGGTTGTCTACCCATGGACCCAGAGGTTCTTTGACAGCTTTGGCAACCTGTCCTCTGCCTCTGCCATCATGGGCAACCCCAAAGTCAAGGCACATGGCAAGAAGGTGCTGACTTCCTTGGGAGATGCCATAAAGCACCTGGATGATCTCAAGGGCACCTTTGCCCAGCTGAGTGAACTGCACTGTGACAAGCTGCATGTGGATCCTGAGAACTTCAAGGTGAGTCCAGGAGATGT', 0)See How_To_Train_CNN_Classifier_With_Pytorch.ipynb for more detailed description how to train CNN classifier with Pytorch.
We also provide these benchmarks through HuggingFace Hub: https://huggingface.co/katarinagresova
If you are used to using Hugging Face dataset, you can use this option to access Genomic Benchmarks. See How_To_Use_Datasets_From_HF.ipynb for a guide.
- datasets: Each folder is one benchmark dataset (or a set of bechmarks in subfolders), see README.md for the format specification
- docs: Each folder contains a Python notebook that has been used for the dataset creation
- experiments: Training a simple neural network model(s) for each benchmark dataset, can be used as a baseline
- notebooks: Main use-cases demonstrated in a form of Jupyter notebooks
- src/genomic_benchmarks: Python module for datasets manipulation (downlading, checking, etc.)
- tests: Unit tests for
pytestandpytest-cov
If you beat our current best model on any dataset or just came with an interesting new idea, let us know about it: Make you code publicly available (GitHub repo, Colab...) and fill in the form at
https://forms.gle/pvkkrgHNCNmAAC1TA
If you have an interesting genomic dataset, send us an issue with the description and possibly link to the data (e.g. BED file and FASTQ reference). In the future, we will provide functions to make the import easy.
If you are a hero, read the specification of our dataset format and send us a pull request with new datasets/[YOUR_DATASET_NAME] and docs/[YOUR_DATASET_NAME] folders.
We welcome new code contributors. If you see a bug, send us an issue with a minimal reproducible example. Or even better, fix the bug and send us a pull request.
If you use Genomic Benchmarks in your research, please cite it as follows.
Grešová, Katarína, et al. "Genomic benchmarks: a collection of datasets for genomic sequence classification." BMC Genomic Data 24.1 (2023): 25.
@article{grevsova2023genomic,
title={Genomic benchmarks: a collection of datasets for genomic sequence classification},
author={Gre{\v{s}}ov{\'a}, Katar{\'\i}na and Martinek, Vlastimil and {\v{C}}ech{\'a}k, David and {\v{S}}ime{\v{c}}ek, Petr and Alexiou, Panagiotis},
journal={BMC Genomic Data},
volume={24},
number={1},
pages={25},
year={2023},
publisher={Springer}
}