Datasets
Each folder contains either one benchmark or a set of benchmarks. See docs/ for code used to create these benchmarks.
Naming conventions
- dummy_...: small datasets, used for testing purposes
- demo_...: middle size datasets, not necesarily biologically relevant or fully reproducible, used in demos
Versioning
We recommend to check the version number when working with the dataset (i.e. not using default None). The version should be set to 0 when the dataset is proposed, after inicial curration it should be changed to 1 and then increased after every modification.
Data format
Each benchmark should contain metadata.yaml file with its main folder with the specification in YAML format, namely
the version of the benchmark (0 = in development)
the classes of genomic sequences, for each class we further need to specify
- url with the reference
- type of the reference (currently, only fa.gz implemented)
- extra_processing, a parameter helping to overcome some know issues with identifiers matching
The main folder should also contain two folders, train and test. Both those folders should contain gzipped CSV files, one for each class (named class_name.csv.gz).
The format of gzipped CSV files closely resemble BED format, the column names must be the following:
- id: id of a sequence
- region: chromosome/transcript/... to be matched with the reference
- start, end: genomic interval specification (0-based, i.e. same as in Python)
- strand: either '+' or '-'
To contribute a new datasets
Create a new branch. Add the new subfolders to datasets and docs. The subfolder of docs should contain a description of the dataset in README.md. If the dataset comes with the paper, link the paper. If the dataset is not taken from the paper, make sure you have described and understand the biological process behind it.
If you have access to cloud_cache folder on GDrive, upload your file there and update CLOUD_CACHE in cloud_caching.py.
To review a new dataset
Make sure you can run and reproduce the code. Check you can download the actual sequences and/or create a data loader. Do you understand what is behind these data? (either from the paper or the description) Ask for clarification if needed.