CATH Domain Classification

This repository contains CATH v4.4.0 protein domain classification data packaged for the Hugging Face Dataset Viewer and the datasets API.

The original raw CATH files are preserved in this dataset repository. The added Parquet files normalize the domain list, domain FASTA sequences, hierarchy names, and nonredundant subset membership into a single tabular dataset with deterministic train/test splits.

Splits

Split	Rows
train	541,123
test	60,205
total	601,328

The split is deterministic and S35-cluster-aware: all domains sharing the same CATH homologous superfamily and S35 cluster key are kept in the same split. This avoids putting close S35-cluster relatives in both train and test.

Dataset Statistics

Metric	Value
Domains	601,328
Unique S35 cluster keys	37,350
Unique homologous superfamilies	6,630
Unique topologies	1,472
Unique architectures	43
Train/test S35 cluster overlap	0
Unknown structure-resolution sentinel rows	9,726

Class distribution:

CATH class	Rows
Alpha Beta	305,361
Mainly Beta	158,943
Mainly Alpha	126,178
Few Secondary Structures	6,034
Special	4,812

Sequence length ranges from 9 to 1,275 amino acids, with a median length of 140.

Load With datasets

from datasets import load_dataset

ds = load_dataset("LiteFold/CATH")
print(ds)

train = ds["train"]
test = ds["test"]

print(train[0])

Load one split directly:

from datasets import load_dataset

train = load_dataset("LiteFold/CATH", split="train")
test = load_dataset("LiteFold/CATH", split="test")

Stream rows without downloading the full dataset first:

from datasets import load_dataset

streamed = load_dataset("LiteFold/CATH", split="train", streaming=True)
first_row = next(iter(streamed))

Filter to one of the CATH nonredundant representative subsets:

from datasets import load_dataset

ds = load_dataset("LiteFold/CATH", split="train")
s35_train = ds.filter(lambda row: row["in_s35_nonredundant_subset"])

Main Columns

Column	Description
domain_id	CATH domain identifier, for example 1oaiA00.
pdb_id, chain_id, pdb_chain_id	Parsed PDB and chain identifiers.
cath_version	CATH release version from the FASTA headers.
cath_code	Full CATH classification code at homologous superfamily level.
class_*, architecture_*, topology_*, homologous_superfamily_*	Numeric codes, full hierarchy codes, names, and example domains.
s35_cluster_id, s60_cluster_id, s95_cluster_id, s100_cluster_id	CATH sequence cluster identifiers from the domain list.
s35_cluster_key	Composite key used for leakage-aware splitting.
domain_length	Domain length from the CATH domain list.
raw_structure_resolution_angstrom	Original structure resolution value.
structure_resolution_angstrom	Resolution with CATH's 999.000 unknown sentinel converted to null.
sequence	Amino acid sequence from cath-domain-seqs.fa.
sequence_length	Length of sequence.
sequence_range	Source FASTA residue range, including discontinuous ranges such as 2-78_187-208.
sequence_range_start, sequence_range_end	Parsed min start and max end residue positions when available.
sequence_segment_count	Number of underscore-separated residue segments in sequence_range.
in_s35_nonredundant_subset, in_s60_nonredundant_subset, in_s95_nonredundant_subset, in_s100_nonredundant_subset	Whether the domain appears in the corresponding CATH nonredundant subset list.

Source Files Used

• cath-domain-list.txt
• cath-domain-list-S35.txt
• cath-domain-list-S60.txt
• cath-domain-list-S95.txt
• cath-domain-list-S100.txt
• cath-domain-seqs.fa
• cath-names.txt

The raw PDB tarball remains in the repository, but it is not embedded in the Parquet table.

Reproducibility

The processed files were generated from the raw files already present in this repository. The preparation script parses CATH v4.4.0 source rows, joins sequences and hierarchy names, adds subset flags, and writes data/train-00000-of-00001.parquet and data/test-00000-of-00001.parquet.