ContrasTED

A CATH homologous superfamily may be represented by thousands of solved domains or by exactly one. Pair- and triplet-based contrastive objectives learn from a superfamily only when two of its members land in the same mini-batch, so the singletons and near-singletons that make up most of CATH — precisely where remote homology is hardest to detect — barely train at all. ContrasTED removes that constraint with a bank of learnable proxy centers, one per superfamily, comparing every domain against all of them at once.

The model is a projection head from Christine Orengo's group at University College London, with collaborators at Helmholtz Munich, the Technical University of Munich and the University of Padova, and posted to bioRxiv in September 2026. It consumes frozen embeddings from ProstT5 — a mean-pooled amino-acid vector concatenated with a mean-pooled 3Di structural-alphabet vector, 2,048 dimensions in total — and maps them into a 128-dimensional space where cosine distance is the operative metric. Reference centroids are then the L2-normalized means of projected CATH domains, so superfamily assignment is a centroid lookup.

That split matters at scale: the ProstT5 embedding is the expensive step, at 9.8–10.4 ms per domain on an NVIDIA L40S, and everything after it is arithmetic on 128-dimensional vectors. The design sits alongside CATHe2, a supervised softmax head over the same inputs, and structural search tools such as Foldseek and Foldclass.

#Key Features

• Proxy centers instead of positive pairs: Center-contrastive loss keeps a learnable vector per CATH superfamily and scores every domain against the full bank, so singleton superfamilies generate gradient updates in every step.
• Sequence and structure in one vector: Each domain is encoded twice by ProstT5 — from its amino-acid sequence and from the Foldseek 3Di token string derived from its coordinates — and the two 1,024-dimensional embeddings are concatenated before projection.
• Calibrated rejection threshold: A cosine cutoff chosen on validation at 99% empirical precision is applied unchanged at test time, retaining 64.0% of S20 queries at 98.5% precision and rejecting 1,382 of 1,401 domains from 50 withheld superfamilies.
• Centroid search at database scale: Collapsing each superfamily to one prototype reduces the CATH reference to 5,659 centroids — 1.4 MB as an fp16 matrix, the figure the paper reports — and supports over 100,000 queries per second on a single CPU core. The centroid file distributed in the repository is stored at higher precision and is correspondingly larger.
• Structural ordering preserved: Cosine distance tracks TM-score at ρ = −0.82 overall and −0.61 below 20% sequence identity, so the metric sharpens superfamily boundaries without discarding continuous structural similarity.

#Technical Details

The trainable component is a two-layer perceptron — 2,048 to 512 with batch normalization, GELU and dropout 0.1, then 512 to 128 with L2 normalization — roughly 1.1 million parameters. Center-contrastive loss uses scale 32, additive cosine margin 0.3, center regularization weight 1.0 and label smoothing 0.1, with batches of 1,024 domains sampled uniformly across superfamilies. Training data comes from CATH v4.4 classes 1–3, split so no training domain shares ≥20% sequence identity with the evaluation sets: 116,301 training, 514 validation and 1,028 test domains across 5,659 superfamilies, augmented with 11,131,930 filtered domain representatives from The Encyclopedia of Domains carrying consensus labels. On the S20 benchmark, superfamily accuracy reaches 92.9% by 1-nearest-neighbour and 91.4% by nearest centroid, against 88.2% for Foldseek in sensitive mode, 84.2% for CATHe2, 81.9% for the frozen ProstT5 embeddings it projects and 75.0% for ProtTucker. The margin is widest in the sparsest quartile, superfamilies with one to six training domains, at 90.7% against 79.2% for Foldseek.

#Applications

The headline use is bulk annotation of predicted structures. Applied to 20,833,783 domains in The Encyclopedia of Domains that carried a topology label but no superfamily assignment, ContrasTED produced 4,666,451 high-confidence assignments spanning 3,796 superfamilies, 92.2% of which agree with the existing topology annotation, and 696 superfamilies gained at least 1,000 new members. One case shows what the threshold admits: the uncharacterized domain AF-A0A074XWQ2-F1-model_v4_TED01 joined superfamily 2.115.10.20 at 2.8% sequence identity to the CATH representative, with TM-align support at TM-score 0.656 over 251 residues. Database curators and groups annotating proteomes from predicted structures are the natural users.

#Impact

ContrasTED shows that the bottleneck in embedding-based homology classification is the training objective rather than the backbone: the same frozen ProstT5 vectors gain eleven accuracy points once projected through a proxy-based metric, the gain concentrated where alignment methods and softmax classifiers run out of examples. The limits are stated by the authors: domain boundaries must be supplied in advance, coverage is restricted to CATH classes 1–3, centroid lookup returns a label but no residue-level alignment, and the TED labels used for augmentation carry whatever residual noise their automated pipelines introduced. The work is a preprint awaiting peer review. The training and inference package is MIT-licensed and the CATH centroid index ships in the repository, but the canonical AA‖3Di projection head is not yet distributed and the versioned data archive promised in the preprint has not been posted.