GrimSqueaker/cath43-eat

CATH v4.3 Annotation Transfer (ProtTucker / EAT splits)

Can a protein language model recognise that two proteins share a fold when their sequences look unrelated? That is the question this dataset asks.

You are given a large lookup set of protein domains with known CATH structural classifications, and a small set of query domains that have been filtered so no query has a detectable sequence-alignment relative in the lookup set. Predict each query's CATH class by finding its nearest neighbour in the lookup set. Sequence search does poorly here by construction; the interesting question is whether embeddings do better.

This is the "midnight zone" benchmark from Heinzinger et al. (2022). The splits are the authors' own, so results are directly comparable to their Table 1.

Quick start

from datasets import load_dataset

ds = load_dataset("GrimSqueaker/cath43-eat")
lookup, queries = ds["lookup"], ds["test_h"]

# 1-nearest-neighbour annotation transfer, the whole method:
#   embed both sets, then for each query take the label of the closest lookup
#   protein by Euclidean distance. Accuracy = how often that label is right.

Splits

Split	Rows	What it is
lookup	69,605	Labelled set. Annotations are transferred from here.
test_h	150	The benchmark. Queries answerable at the H level (see below).
test219	219	All queries, for per-level analysis at C/A/T.
train	66,052	Authors' training set, if you want to train a model rather than evaluate one.
validation	200	Authors' early-stopping set.
test300	300	Superset of test219 before the strictest redundancy filter.

test219 is non-redundant to lookup at HVAL ≤ 0 — roughly, no query has a lookup relative that sequence alignment can find. That is what makes this a midnight-zone test rather than a lookup exercise.

Why test_h has 150 rows and not 219. A query whose superfamily does not appear anywhere in the lookup set cannot be answered correctly by any method. Scoring those queries would charge every method for an impossible case and put a ceiling below 100% for reasons unrelated to the method. The original paper excludes them per level; test_h is that exclusion applied for the H level, pre-computed so you get comparable numbers by default.

Queries answerable at each level:

Level	Answerable / 219
C (class)	219
A (architecture)	219
T (topology)	208
H (homologous superfamily)	150

If you evaluate at C, A, or T, use test219 and apply the exclusion for that level yourself: drop queries whose label at that level is absent from lookup.

Fields

Field	Example	Notes
id	2vwaA00	CATH domain identifier
sequence	INKINLNKPII...	Amino acid sequence of the domain (median 134 aa)
cath_c	1	Class
cath_a	1.10	Architecture
cath_t	1.10.8	Topology
cath_h	1.10.8.10	Homologous superfamily — the hardest and headline level

The four levels nest, so cath_h implies cath_t implies cath_a implies cath_c. They are stored separately so you do not have to re-split strings.

Reference numbers

From Heinzinger et al. (2022), Table 1 — accuracy transferring from lookup to the queries, per CATH level:

Method	C	A	T	H
Random baseline	29	9	1	0
MMseqs2 (sequence search)	52	36	29	35
HMMER (profile search)	70	60	59	77
ProtT5 embeddings, raw	84	67	57	64
ProtTucker(ProtT5), contrastively tuned	89	75	64	76

Note the shape of it: sequence search collapses, profile search holds up, and raw embeddings beat both profile search at the coarse levels while losing at H. Contrastive tuning on CATH closes that gap.

Provenance

Assembled from two sources, both by the original authors:

• Splits and sequences — github.com/Rostlab/EAT, data/ProtTucker/*.fasta (repository licensed GPL-3.0).
• CATH labels — Zenodo 14675997, cath-domain-list.txt, CathDomainList v4.3.0 in CLF 2.0 format (CC-BY-4.0). Columns 2–5 give C, A, T and H.

Underlying structural classification is CATH v4.3.0. Every domain in every split matched a CATH entry, so nothing was dropped for missing labels.

Build script: scripts/build_cath_eat_dataset.py in ProtBench.

Citation

If you use this dataset, cite the paper it comes from:

@article{heinzinger2022contrastive,
  title   = {Contrastive learning on protein embeddings enlightens midnight zone},
  author  = {Heinzinger, Michael and Littmann, Maria and Sillitoe, Ian and
             Bordin, Nicola and Orengo, Christine and Rost, Burkhard},
  journal = {NAR Genomics and Bioinformatics},
  volume  = {4},
  number  = {2},
  pages   = {lqac043},
  year    = {2022},
  doi     = {10.1093/nargab/lqac043}
}

And CATH itself:

@article{sillitoe2021cath,
  title   = {CATH: increased structural coverage of functional space},
  author  = {Sillitoe, Ian and Bordin, Nicola and Dawson, Natalie and others},
  journal = {Nucleic Acids Research},
  volume  = {49},
  number  = {D1},
  pages   = {D266--D273},
  year    = {2021},
  doi     = {10.1093/nar/gkaa1079}
}