Foldclass

The domain, not the chain, is the unit of protein evolution, and the flood of predicted structures has made that distinction expensive to ignore. A typical entry in the AlphaFold Database is a multi-domain chain whose globular units are strung together by long, low-confidence linkers corresponding to nothing in an experimental structure. Compare two such chains end to end and the score is diluted by residues belonging to no domain at all, so a real match on one domain vanishes into the noise of the rest.

Foldclass, from David Jones's group at University College London, embeds one protein domain, and only a domain, into a fixed 128-dimensional vector, so structural comparison becomes a cosine nearest-neighbour lookup against a precomputed library rather than an alignment. Merizo-search is the pipeline built around it: Merizo, the group's earlier invariant-point-attention segmentation model, cuts a query chain into domains, Foldclass embeds each one, the top k cosine neighbours are retrieved, and TM-align rescores that shortlist to confirm the hits. The embedding step is a prefilter that shrinks an alignment problem over hundreds of millions of structures to a few hundred candidate pairs.

That places Foldclass alongside embedding-based search methods such as FoldMatch and PLMSearch, with two departures: the unit is a domain rather than a chain, and the embedding is learned by plain fold classification rather than a contrastive objective. First posted as a preprint in March 2024, the work was published in Bioinformatics in 2025.

#Key Features

• Fixed-size domain vectors: Any domain, of any length, becomes one 128-dimensional vector, so a database is a flat matrix that inner-product search handles directly.
• Classifier pretraining: The network is trained as a multiclass CATH classifier rather than with contrastive pairs; the classification heads are discarded at inference and the pooled representation becomes the embedding.
• Segment-then-search: The easy-search mode chains Merizo segmentation into Foldclass retrieval, so an unannotated multi-domain chain is searched per domain in one command.
• Multi-domain architecture matching: Hits are graded into four categories, from a "bag-of-domains" match to an exact architecture match preserving domain order.
• Faiss backend for large libraries: Exhaustive inner-product search over normalised vectors scales past system memory, covering all 365 million domains in the Encyclopedia of Domains in roughly two minutes on NVMe storage with a GPU.

#Technical Details

Foldclass is a stack of two E(n)-equivariant graph neural network blocks with a hidden dimension of 128 and an output dimension of 256 in the node and edge update sub-networks. Cα coordinates supply the coordinate inputs and sinusoidal positional encodings the node features; the coordinate-update step of the original EGNN formulation is omitted, and sequence is deliberately excluded as an input because it caused rapid overfitting. Node features from the final block are averaged into the 128-dimensional embedding, which is what makes the output size independent of domain length. The released checkpoint is 5.5 MB.

Training used the 31,885 domains of the CATH 4.3 S30 non-redundant set, with three linear heads predicting the Class (5 labels), Architecture (43) and Topology (1,421) levels under class-weighted cross-entropy, the coarser levels acting as auxiliary losses. Optimisation ran up to 300 epochs with AdamW at a learning rate of 3e-4 and weight decay of 1e-2, with 1.5 Å Gaussian noise added to input coordinates. On a held-out set of 62 CASP13–15 domains with no detectable CATH homologue, the classifier is correct for 45 (73%), rising to 95% precision at a confidence threshold of 0.9. Across 21,005 CATH S40 queries against S20 targets, the top target shares the query superfamily for roughly 89% of queries at k=20. For whole-chain retrieval — 1,828 multi-domain queries against 312,544 CATH chains — mean precision reaches 0.865 at superfamily level and 0.910 at topology level for exact architecture matches.

#Applications

The practical use is annotation transfer at predicted-structure scale: take an AlphaFold or ESMFold model of unknown function, segment it, and ask which CATH superfamilies its domains belong to. Because matching is per domain, the method also answers questions whole-chain search cannot express — finding proteins with the same domain composition in a different order, or shared architectures between enzymes with no overall structural similarity. Prebuilt databases cover CATH 4.3 and all TED domains, and Merizo-search runs in easy-search mode on the PSIPRED web server for users who would rather not host a terabyte of embeddings locally.

#Impact

Foldclass complements alignment-based search rather than replacing it: it recovers several hundred correct matches that Foldseek misses even in exhaustive mode, while Foldseek's sensitive settings stay slightly ahead on overall superfamily recovery. The limits are that embedding retrieval is inexact, that precision falls as query chains gain domains, and that whole-chain results inherit any error in the Merizo segmentation — the authors advise checking the domain parse before trusting a multi-domain hit. Code is GPL-3.0 with the trained checkpoint in the repository, and the prebuilt domain databases are deposited under CC BY 4.0.