Domain Classification · LiteFold
CATH
Hierarchical protein domain classification dataset structured into train and test splits partitioned by S35 clusters.
What CATH is
CATH (LiteFold/CATH) provides a tabular compilation of 601,328 protein domain structures categorised according to the CATH hierarchical classification scheme: class, architecture, topology, and homologous superfamily. It integrates sequence strings, domain boundaries, resolution metadata, and hierarchy identifiers into machine-learning-ready tabular tables.
The dataset addresses redundancy and data leakage in computational biology benchmarks by offering a pre-computed, deterministic split. By partitioning records using composite S35 sequence identity cluster keys, all domains sharing an S35 cluster and homologous superfamily are placed exclusively into either the 541,123-domain training split or the 60,205-domain test split.
How it was made
Data records are generated by assembling information across multiple CATH flat files, including domain lists, nonredundant subset lists (S35, S60, S95, and S100), domain FASTA sequences, and taxonomy names. Tabular rows capture numeric and textual classification paths, residue segment boundaries, and experimental structure resolutions. Structural resolution values marked with the CATH unknown sentinel value (999.000) are converted to null in the cleaned resolution column.
Partitioning enforces zero S35 cluster overlap between splits. Each record includes the full sequence from the source FASTA, parsed PDB and chain codes, segment counts for discontinuous sequence ranges, and flags identifying whether the domain belongs to any of the nonredundant representative subsets.
Key facts
- Publisher
- LiteFold
- Total domains
- 601,328
- Train split size
- 541,123 domains
- Test split size
- 60,205 domains
- Unique S35 cluster keys
- 37,350
- Unique homologous superfamilies
- 6,630
- Unique topologies
- 1,472
- Unique architectures
- 43
- Sequence length range
- 9 to 1,275 amino acids (median 140)
- Data format
- Parquet
What you can use it for
- Protein fold classification
- Train and evaluate models to predict CATH class, architecture, topology, or homologous superfamily directly from primary amino acid sequences.
- Leakage-free sequence benchmarking
- Evaluate generalization on unseen sequence families using the strict zero-overlap S35 cluster division between training and test sets.
- Nonredundant domain sampling
- Filter domains by S35, S60, S95, or S100 nonredundant flags to construct balanced sub-datasets at controlled sequence similarity thresholds.
- Discontinuous domain analysis
- Study multi-segment and split domains using parsed residue ranges and sequence segment counts.
Explore the data
The schema, real rows, full-text search and a SQL filter — read in your browser, nothing downloaded.
Explore
Read the real rows without downloading anything
Reading rows…
LiveRead from LiteFold/CATH at the moment you asked. Nothing is cached or stored — every row above came from that request.
How the columns are distributed
Over 541,123 rows of default/train
architecture_code
string_label- 3.4020%
- 3.3014%
- 2.6014%
architecture_example_domain_id
string_label- 1byiA0020%
- 6n3dA0114%
- 4unuA0014%
architecture_name
string_label- 3-Layer(aba) Sandwich20%
- 2-Layer Sandwich14%
- Sandwich14%
architecture_number
int- min
- 10
- median
- 40
- mean
- 39.16
- max
- 180
cath_code
string_text- min
- 9
- median
- 11
- mean
- 10.96
- max
- 13
cath_version
string_label- 4.4.0100%
chain_id
string_label- A37%
- B21%
- C9%
class_code
string_label- 351%
- 226%
- 121%
class_example_domain_id
string_label- 1byiA0051%
- 4unuA0026%
- 6e1fA0021%
class_name
string_label- Alpha Beta51%
- Mainly Beta26%
- Mainly Alpha21%
class_number
int- min
- 1
- median
- 3
- mean
- 2.35
- max
- 6
domain_id
string_text- min
- 7
- median
- 7
- mean
- 7
- max
- 7
domain_index
int- min
- 0
- median
- 1
- mean
- 1.14
- max
- 20
domain_length
int- min
- 9
- median
- 136
- mean
- 160.90
- max
- 1,221
domain_suffix
string_label- 0034%
- 0132%
- 0224%
homologous_superfamily_code
string_text- min
- 9
- median
- 11
- mean
- 10.96
- max
- 13
homologous_superfamily_example_domain_id
string_text- min
- 7
- median
- 7
- mean
- 7
- max
- 7
homologous_superfamily_name
string_text- min
- 0
- median
- 28
- mean
- 27.68
- max
- 107
homologous_superfamily_number
int- min
- 1
- median
- 10
- mean
- 362.31
- max
- 12,820
in_s100_nonredundant_subset
bool- False80%
- True20%
in_s35_nonredundant_subset
bool- False94%
- True6%
in_s60_nonredundant_subset
bool- False91%
- True9%
in_s95_nonredundant_subset
bool- False88%
- True12%
pdb_chain_id
string_text- min
- 5
- median
- 5
- mean
- 5
- max
- 5
pdb_id
string_text- min
- 4
- median
- 4
- mean
- 4
- max
- 4
raw_structure_resolution_angstrom
float- min
- 0.48
- median
- 2.39
- mean
- 66.38
- max
- 1,000
s100_cluster_id
int- min
- 1
- median
- 1
- mean
- 2.34
- max
- 214
s100_sequence_count
int- min
- 1
- median
- 5
- mean
- 37.08
- max
- 2,120
s35_cluster_id
int- min
- 1
- median
- 8
- mean
- 46.93
- max
- 1,117
s35_cluster_key
string_text- min
- 15
- median
- 18
- mean
- 17.56
- max
- 20
s60_cluster_id
int- min
- 1
- median
- 1
- mean
- 1.47
- max
- 21
s95_cluster_id
int- min
- 1
- median
- 1
- mean
- 2.26
- max
- 211
sequence
string_text- min
- 9
- median
- 141
- mean
- 165.50
- max
- 1,275
sequence_length
int- min
- 9
- median
- 141
- mean
- 165.50
- max
- 1,275
sequence_range
string_text- min
- 4
- median
- 6
- mean
- 7.14
- max
- 47
sequence_range_end
int- min
- -112
- median
- 246
- mean
- 320.14
- max
- 9,140
sequence_range_start
int- min
- -369
- median
- 26
- mean
- 137.45
- max
- 9,098
sequence_segment_count
int- min
- 1
- median
- 1
- mean
- 1.16
- max
- 6
structure_resolution_angstrom
float- min
- 0.48
- median
- 2.35
- mean
- 51.45
- max
- 1,000
1.6% empty
structure_resolution_is_unknown
bool- False98%
- True2%
topology_code
string_text- min
- 6
- median
- 7
- mean
- 7.58
- max
- 9
topology_example_domain_id
string_text- min
- 7
- median
- 7
- mean
- 7
- max
- 7
topology_name
string_text- min
- 3
- median
- 22
- mean
- 27.69
- max
- 109
topology_number
int- min
- 4
- median
- 70
- mean
- 302.68
- max
- 4,200
Strengths and limits
Good at
- Eliminates evaluation data leakage by ensuring zero overlap of S35 sequence clusters between training and testing partitions.
- Includes granular structural hierarchy labels spanning class, architecture, topology, and homologous superfamily.
- Integrates curated sequence identity subsets (S35, S60, S95, S100) alongside primary amino acid sequences.
- Cleans experimental resolution records by converting unknown sentinel values into null entries.
Watch out for
- Three-dimensional atomic coordinates from the underlying PDB files are not embedded directly in the Parquet tables.
- Structure resolution values are missing or recorded as unknown sentinels for 9,726 entries.
Questions
Can I look inside CATH without downloading it?
Yes. This page reads its schema, real rows and per-column statistics in the browser, with full-text search and a SQL filter. Nothing is downloaded to your machine.
What licence is CATH under?
Its authors declare cc-by-4.0. Read the terms on the original page before reusing the data.
How are train and test splits divided?
Splits are partitioned deterministically using homologous superfamily identifiers and S35 sequence identity clusters, ensuring that close S35 relatives never span both train and test splits.
Does the dataset contain 3D atomic coordinates?
No, atomic coordinates are kept in a separate PDB tarball and are not stored inside the Parquet dataset tables, which focus on sequence and hierarchical metadata.
What is the meaning of the nonredundant subset columns?
Columns like in_s35_nonredundant_subset and in_s95_nonredundant_subset denote whether a domain is designated as a representative structure at 35% or 95% sequence similarity thresholds.
How are discontinuous sequence ranges handled?
Discontinuous domains contain underscore-delimited residue ranges in the sequence_range column, alongside a sequence_segment_count field that quantifies the number of segments.
How to cite it
Ian Sillitoe, Nicola Bordin, Natalie Dawson, and others. CATH: increased structural coverage of functional space. Nucleic Acids Research, 49(D1):D266–D273, 2021. doi: 10.1093/nar/gkaa1079.
Written from CATH's own dataset card. Statements citing a figure that does not appear there are removed before publication. Licence and running details come from our registry.