README.md

---
language:
- en
license: bsd-3-clause
size_categories:
- 100K<n<1M
task_categories:
- graph-ml
task_ids:
- multi-class-classification
pretty_name: TEDBench
tags:
- protein
- structure
- fold-classification
- tedbench
- alphafold
- cath
---

# TEDBench

[**Paper**](https://huggingface.co/papers/2605.18552) | [**GitHub**](https://github.com/BorgwardtLab/TEDBench)

Large-scale, non-redundant benchmark for **protein fold classification** built from
[Encyclopedia of Domains (TED)](https://zenodo.org/records/13908086) annotations
projected onto the Foldseek-clustered AlphaFold Database.

This dataset was presented in the paper [Protein Fold Classification at Scale: Benchmarking and Pretraining](https://huggingface.co/papers/2605.18552).

## Dataset statistics

| Split | Structures |
|-------|----------:|
| Train | 369,740 |
| Val   | 46,217 |
| Test  | 46,218 |

**965 CATH topology (T-level) classes** — rare topologies with fewer
than 10 samples are merged into architecture-level "x" classes.

## Schema

| Column | Type | Description |
|--------|------|-------------|
| `name` | `string` | AlphaFold domain identifier (e.g. `AF-Q8IYB3-F1`) |
| `sequence` | `string` | Amino-acid sequence (single-letter code) |
| `coords` | `[L, 3, 3]` float32 | Backbone N/Cα/C coordinates (Å) |
| `plddt` | `[L]` float32 | Per-residue AlphaFold pLDDT confidence score |
| `residue_index` | `[L]` int64 | Residue index in the original AlphaFold model |
| `seq_ids` | `[L]` int64 | ESM-tokenised sequence IDs |
| `label` | `ClassLabel` | CATH topology class index (names are CATH T-level code strings, e.g. `"3.40.50.300"`) |

`label` is a `datasets.ClassLabel` whose `.names` list contains the CATH
topology strings (e.g. `"3.40.50.300"`), so the dataset is fully self-contained.

## Usage

### Load from HuggingFace

```python
from datasets import load_dataset
import torch

ted = load_dataset("TEDBench/ted")
sample = ted["train"][0]

coords = torch.tensor(sample["coords"])          # [L, 3, 3]
plddt  = torch.tensor(sample["plddt"])           # [L]
label  = sample["label"]                         # int index
# Decode label → CATH code string:
cath_code = ted["train"].features["label"].int2str(label)
```

### Use in TEDBench training scripts

```bash
# Fine-tune pretrained MiAE-B on TEDBench
python main_finetune_ted.py datamodule=hf_ted \
    pretrained_model_path=TEDBench/miae-b \
    experiment=finetune_ted_base_n1g8

# Linear probing with pretrained MiAE-B
python main_linprobe_ted.py datamodule=hf_ted \
    pretrained_model_path=TEDBench/miae-b

# Evaluate a fine-tuned model on the CATH 4.4 external test set
python main_test_ted.py datamodule=hf_cath_test \
    pretrained_model_path=TEDBench/miae-b-ft
```

## Source data

- AlphaFold DB clustered by
  [Foldseek](https://afdb-cluster.steineggerlab.workers.dev) (pLDDT > 80)
- CATH topology annotations from
  [TED v365M](https://zenodo.org/records/13908086)

## Citation

```bibtex
@inproceedings{chen2026tedbench,
  title={Protein Fold Classification at Scale: Benchmarking and Pretraining},
  author={Chen, Dexiong and Manolache, Andrei and Niepert, Mathias and Borgwardt, Karsten},
  booktitle={Proceedings of the 43rd International Conference on Machine Learning (ICML)},
  year={2026}
}
```

## License

BSD-3-Clause