TEDBench-AFDB (pretraining corpus)

Paper | GitHub

Representative proteins from Foldseek-clustered AlphaFold Database (pLDDT > 80), used to pretrain MiAE (Masked Invariant Autoencoders) in the TEDBench benchmark.

TEDBench is a large-scale, non-redundant benchmark for protein fold classification constructed from the Encyclopedia of Domains (TED) and Foldseek-clustered AlphaFold structures.

Dataset statistics

Split	Structures
Train	742,183
Val	7,496
Total	749,679

One representative structure per Foldseek sequence-similarity cluster.

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

No label column — this dataset is for unsupervised pretraining only.

Usage

Load from HuggingFace

from datasets import load_dataset
import torch

afdb = load_dataset("TEDBench/afdb", split="train")
sample = afdb[0]

coords = torch.tensor(sample["coords"])   # [L, 3, 3]
plddt  = torch.tensor(sample["plddt"])    # [L]

Pretrain MiAE using this dataset

python main_pretrain.py datamodule=hf_afdbfs

# Multi-GPU (effective batch size 4096)
python main_pretrain.py \
    experiment=tedbench_base_n4g8 \
    datamodule=hf_afdbfs

Source data

• Foldseek cluster representatives from AlphaFold Database v4 (pLDDT > 80)

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