| --- |
| license: cc-by-4.0 |
| viewer: false |
| tags: |
| - biology |
| - protein-structure |
| - inverse-folding |
| - binding-site-prediction |
| size_categories: |
| - 10K<n<100K |
| --- |
| |
| # E3former Datasets |
|
|
| Processed datasets for *Energy-Adaptive Equivariant State Space Models for |
| Noise-Robust Protein Structure Representation*. |
|
|
| ## Datasets |
|
|
| | Dataset | Structure | Task | Train | Validation | Test | |
| |---|---|---|---:|---:|---:| |
| | [CATH](datasets/cath/README.md) | Experimental | Inverse folding | 18,009 | 607 | 1,119 | |
| | [CATH-AF](datasets/cath/README.md) | AlphaFold-derived | Inverse folding | 16,467 | 559 | 1,015 | |
| | [PPBS](datasets/ppbs/README.md) | Experimental | Binding-site prediction | 12,577 | 3,178 | 3,984 | |
| | [PPBS-AF](datasets/ppbs/README.md) | AlphaFold-derived | Binding-site prediction | 11,345 | 2,876 | 3,636 | |
|
|
| The archives contain processed PyTorch/PyG protein objects and fixed |
| train/validation/test splits. PPBS includes the 70, homology, topology, and none |
| test subsets. Family-specific manifests record split counts and archive checksums. |
|
|
| ## Download |
|
|
| ```bash |
| hf download GENTEL-Lab/E3former_data --repo-type dataset --local-dir e3former_data |
| ``` |
|
|
| Or, from the E3former code directory: |
|
|
|
|
| ```bash |
| python scripts/download.py --namespace GENTEL-Lab |
| e3former check-data |
| ``` |
|
|
| To install one dataset, add `--only cath`, `--only cath-af`, `--only ppbs`, or |
| `--only ppbs-af` to the download command. The script verifies SHA256 and extracts |
| the archives into the matching `data/` directories. |
|
|
| Use Python 3.10 and the dependencies pinned in the E3former installation. |
| These are processed inputs for the provided data loader, not raw PDB files. |
|
|
| ## Sources |
|
|
| CATH uses [CATH](https://www.cathdb.info/) protein classifications and benchmark |
| splits. PPBS labels and evaluation subsets originate from |
| [ScanNet](https://github.com/jertubiana/ScanNet). |
| The AF variants provide AlphaFold-derived structures for the corresponding tasks. |
| This release uses CC BY 4.0 for our contributions. Source materials retain their original licenses; see |
| [THIRD_PARTY_DATA.md](THIRD_PARTY_DATA.md). |
|
|
| Please cite E3former and the relevant source datasets when using this release. |
|
|
