README.md

metadata

language:
  - en
license: bsd-3-clause
size_categories:
  - 10K<n<100K
task_categories:
  - graph-ml
pretty_name: TEDBench-CATH
tags:
  - protein
  - structure
  - fold-classification
  - tedbench
  - cath
  - experimental

TEDBench-CATH (CATH 4.4 external test set)

External test set derived from CATH 4.4 experimental structures (40 % non-redundant set), used to evaluate TEDBench models on crystallographic data.

This dataset is part of the TEDBench benchmark, introduced in the paper Protein Fold Classification at Scale: Benchmarking and Pretraining.

Links:

• Code: https://github.com/BorgwardtLab/TEDBench
• Paper: https://huggingface.co/papers/2605.18552

Dataset statistics

Split	Structures
Test	28,010

965 CATH topology (T-level) classes — same label space as TEDBench/ted.

Schema

Column	Type	Description
name	string	PDB chain identifier in <pdbid>.<chain> format (e.g. 1abc.A)
sequence	string	Amino-acid sequence (single-letter code)
coords	[L, 3, 3] float32	Backbone N/Cα/C coordinates (Å) from experimental structure
plddt	[L] float32	Per-residue confidence proxy (set to 100 for experimental structures)
residue_index	[L] int64	Author residue number from the PDB file
seq_ids	[L] int64	ESM-tokenised sequence IDs
label	ClassLabel	CATH topology class index (same label space as TEDBench)

Protein identifiers follow the <pdbid>.<chain> convention (e.g. 1abc.A), matching the CATH domain annotation files. The label ClassLabel uses the same CATH topology strings and integer indices as the TEDBench training set.

Usage

Load from HuggingFace

from datasets import load_dataset
import torch

cath = load_dataset("TEDBench/cath", split="test")
sample = cath[0]

coords = torch.tensor(sample["coords"])          # [L, 3, 3]
label  = sample["label"]                         # int index
cath_code = cath.features["label"].int2str(label)

Use in TEDBench training scripts

python main_test_ted.py datamodule=hf_cath_test pretrained_model_path=<ckpt>

Source data

CATH 4.4 40 % non-redundant representative set from CATHDB. Structures fetched from PDB-REDO / RCSB, quality-filtered, and converted to single-chain PDB files. The plddt column is set to 100.0 for all residues (these are experimental structures, not AlphaFold predictions).

Citation

@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