README.md

metadata

license: mit
tags:
  - biology
pretty_name: Magneton
size_categories:
  - 100K<n<1M
configs:
  - config_name: full
    data_files:
      - split: full
        path: interpro_103.0/swissprot_subset/swissprot.with_ss.*.jsonl.gz

Magneton

Datasets for training and evaluating substructure-aware protein representation learning models. Use this dataset in conjunction with our environment for training and evaluation on GitHub.

Substructure data

The core of the dataset is 530,601 SwissProt proteins with associated substructure annotations from DSSP and the 103.0 release of InterPro. These proteins are stored as sharded JSONL files with 10,000 proteins per file. The dataset has the following structure:

• interpro_103.0/
  • swissprot_subset/ - JSONL files of proteins with substructure annotations
  • debug_subset/ - a small subset of 10,000 proteins for debugging and testing
  • labels/ - a standardized ordering of InterPro element types for mapping to integer labels. selected_subset contains the subset of labels used in our publication
  • dataset_splits/ - dataset splits (train/val/test) for the SwissProt and debug datasets. Splits are available for both sequenced-based clustering (AFDB50) and structure-based clustering (FoldSeek cluster). See FoldSeek cluster for more information.
• sequences/ - FASTA file of SwissProt protein sequences and files associated with dataset selection.
• afdb_structures/ - Placeholder default directory for downloading protein structures in PDB format from AlphaFold DB. This is the default search location for protein structures, but can be configured per-run using the data.struct_template parameter if you already have a AFDB download elsewhere.

To load the datasets, please refer to our GitHub repo, or if you would prefer to directly load the JSONL files into your own environment, please refer to this implementation of loading Protein objects from JSONL.

Model-specific data

Currently, Magneton incorporates two models that utilize protein structure as an input: SaProt and ProSST. These models both use structure tokens to represent the protein structure, and we precompute these structure tokens for the Magneton SwissProt dataset to speed up the training process. The generated_data/ directory contains these files.

Evaluation datasets

The evaluations directory contains the data files for the various evaluation tasks outlined here. The exact format of the files varies by the source dataset. Please see Appendix A.2 in our paper for additional details about each dataset. For some of the larger datasets, the evaluations directory also contains precomputed structure tokens for ProSST and SaProt.