TEDBench: Large-Scale Protein Fold Classification Benchmark with 965 Topology Classes

369,740 protein structures for training, with 46,217 for validation and 46,218 for testing, form this benchmark for protein fold classification. The dataset, named TEDBench, is built from TED annotations projected onto the Foldseek-clustered AlphaFold Database. It was presented in the paper 'Protein Fold Classification at Scale: Benchmarking and Pretraining'.

Use Cases

• Benchmarking protein fold classification models based on the 965 CATH topology classes.
• Pretraining models for structural bioinformatics using the large-scale, non-redundant set of protein structures.
• Analyzing the distribution and characteristics of rare protein topologies mentioned in the description.

Strengths

• Large scale with 369,740 training structures, 46,217 validation structures, and 46,218 test structures.
• Covers 965 distinct CATH topology (T-level) classes, including rare topologies.
• Designed as a non-redundant benchmark, which may reduce data bias.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• Last updated 2026-05-20 06:39:14; freshness should be verified.

Provenance

Source

TEDBench, built from Encyclopedia of Domains (TED) annotations and the AlphaFold Database.

Collection Method

Annotations were projected onto the Foldseek-clustered AlphaFold Database.

Freshness

Last updated 2026-05-20 06:39:14.