CATHe2

The CATH database organises protein domains into a hierarchy of class, architecture, topology and homologous superfamily, and is one of the reference resources for reasoning about evolutionary and structural relationships between proteins. Assigning a new domain to its superfamily once involved expert curation, but the volume of predicted structures now flowing in from deep learning has made manual assignment infeasible. The CATH team's answer was CATHe, a feed-forward classifier over ProtT5 protein language model embeddings that detects remote homologues down to 20% sequence identity.

CATHe2 is a direct successor built by Orfeú Mouret and Jad Abbass at Kingston University London, with Mouret also affiliated with Télécom Physique Strasbourg at the University of Strasbourg. It keeps CATHe's dataset and problem formulation and changes three things: the upstream language model becomes ProstT5, a bilingual model that can embed both amino acid sequences and 3Di structural-alphabet sequences; 3D structural information enters the classifier through those 3Di embeddings; and the classifier head itself is re-tuned. The preprint was posted to bioRxiv in June 2025 and the peer-reviewed version appeared in Biology Methods and Protocols in November 2025.

The result is a substantial accuracy gain on an unchanged benchmark, which makes the paper a fairly clean ablation of what structural-alphabet input buys a superfamily classifier: swapping ProtT5 for ProstT5 alone accounts for 6.7 points of F1, and adding the 3Di channel carries the total to 9.9.

#Key Features

• Dual-modality input: The best configuration concatenates ProstT5 embeddings of the amino acid sequence with ProstT5 embeddings of the 3Di structural-alphabet sequence derived from the domain's PDB coordinates, so sequence and geometry are both available to the classifier.
• Sequence-only fallback: A simplified variant needing only amino acid sequences still improves on CATHe's F1 score by 6.7 ± 1.3% and accuracy by 6.6 ± 0.7%, which matters because most query domains have no experimental structure.
• Remote homology at 20% identity: Like its predecessor, the model is trained and tested under a stringent sequence identity threshold below 20%, so reported accuracy reflects genuinely remote homology detection rather than trivial sequence matching.
• Prepared inference pipeline: The repository ships a setup script that pulls trained weights from Zenodo, plus documented steps for supplying FASTA sequences and matching PDB files and running prediction end to end.
• Backward comparability: Both the original CATHe and CATHe2 models are packaged in the same repository behind separate environments, making head-to-head comparison on a user's own domains straightforward.

#Technical Details

The classifier is a feed-forward network with two blocks, each comprising a dense layer of 2,048 units, a LeakyReLU activation, batch normalisation and dropout at rate 0.3. Embeddings are produced by ProstT5, a fine-tuned derivative of ProtT5 trained to translate between amino acid and 3Di alphabets, and the 3Di sequences encode each residue's geometric relationship to its nearest spatial neighbour rather than backbone torsions, which the authors argue gives better information density and fewer false positives.

Training used the CATHe dataset filtered to domains for which 3Di sequences could be derived, leaving 901,437 training domains with validation and test sets of 6,863 domains each. On the largest split, covering roughly 1,700 superfamilies, the best CATHe2 configuration reaches 92.2 ± 0.7% accuracy and an F1 score of 82.3 ± 1.3%, against 85.6 ± 0.4% accuracy and 72.4 ± 0.7% F1 for CATHe — a 9.9-point F1 gain and a 6.6-point accuracy gain. Code is released under the MIT licence, weights are hosted on Zenodo, and the datasets carry their own Zenodo DOI.

#Applications

The immediate beneficiary is structural bioinformatics infrastructure: automated superfamily assignment is how a database like CATH keeps pace with the tens of millions of predicted structures now available, and a classifier that improves F1 by ten points reduces the annotation backlog needing human review. For individual researchers, the tool answers a common practical question — given a newly characterised or predicted domain, which known superfamily does it belong to, and therefore what function and fold can be hypothesised. The sequence-only variant serves metagenomics and other settings where structures are unavailable.

#Impact

CATHe2 is a narrow, single-task classifier rather than a general-purpose backbone, and its contribution is correspondingly focused: quantitative evidence that structural-alphabet embeddings add real signal to remote homology detection beyond what a sequence language model captures, on a benchmark where the sequence-only ceiling was already high. That finding generalises past this task — 3Di representations are cheap to compute from predicted structures and are becoming a standard extra input channel for protein classification problems. The main limitation is scope: performance is established on the CATH split alone and has not been shown to transfer to other classification hierarchies. The model also inherits whatever biases exist in CATH's coverage of fold space, and its full-strength configuration depends on having a structure for the query domain.