borgwardtlab/proteinshake_release

This repository contains the code for releasing proteinshake datasets. It also computes the random, sequence, and structure splits.

To release, install proteinshake, foldseek, and cd-hit:

• https://github.com/BorgwardtLab/proteinshake
• https://github.com/steineggerlab/foldseek
• https://github.com/weizhongli/cdhit

Adjust the TAG, SCRATCH, DESTINATION, NJOBS variables in release.py according to your system.

Then run: python release.py

Note that this repository is licensed under GPLv3, due to dependency licenses.

proteinshake_release's Issues

split sizes differ between similarity thresholds

cd-hit drops sequences

only a few (25 out of 30k), but we should look into this. Temporary fix is to assign those sequences to a -1 cluster, we might also filter them out

file versioning for efficient TMalign re-calculation

catch http errors in the download

retry the download after some minutes to not crash the release

dynamic release

automatically fetch all available datasets
automatically upload to zenodo

add class_list to release

to avoid computing the token map on the user end

parallelize foldseek over thresholds

not per task, overhead is too big

multiple CVs for the splits

Could be implemented like:

protein_dict['protein'] = {
...
'CV1': 'test',
'CV2': 'train',
'CV3': 'train',
...
}