Loading…
Enzyme function assignment by embedding retrieval, fusing three frozen protein language models into a 256-D EC-aware search space.
No providers recorded yet. Browse all providers
Below about 30% pairwise sequence identity, an alignment can no longer tell a true homolog from a chance match, and enzyme annotation inherits the failure: on the CARE benchmark's remote-homology split, alignment-free and alignment-based methods alike sit near 51–55% accuracy at full four-digit Enzyme Commission (EC) specificity. Protein language models are the usual answer, but each is only as informative as its pretraining task — masked language modelling recovers co-evolutionary constraint, sequence-to-structure translation recovers fold topology, and Gene Ontology co-training recovers functional vocabulary. No single encoder holds all three, and enzyme function depends on all three at once.
FuncSeek is a supervised contrastive projection head that fuses them. It takes fixed, precomputed embeddings from three already-released backbones — ESM-2 for evolutionary co-variation, ProstT5 for 3Di fold topology, and ProteinBERT for GO-derived functional semantics — concatenates them into a 2,816-dimensional vector, and learns a projection into a 256-dimensional space where cosine distance tracks EC identity. None of the three backbones is fine-tuned; each acts as a frozen feature extractor, and the only trained component is the head. The paper shows why the learned step is necessary rather than cosmetic: naively concatenating the three raw embeddings performs worse at retrieval depth than the best single encoder, because the highest-magnitude component dominates the distance and drowns the rest.
The model was developed by Leendert Cloete and Hugh Patterton at the Centre for Bioinformatics and Computational Biology at Stellenbosch University, with Cloete also affiliated with Computational Sciences and Informatics at Roche in Cape Town, and posted to bioRxiv in August 2026. It sits alongside retrieval-first approaches such as PLMSearch, but targets catalytic function rather than structural homology.
The projection head is a three-layer feedforward network (2816 → 512 → 512 → 256) with
LayerNorm before each activation and dropout of 0.1, ending in an L2 normalisation that
places every embedding on the unit hypersphere — roughly 1.8 million trainable parameters.
It is trained with supervised contrastive loss (temperature 0.1) and hard negative mining:
each batch draws 64 EC4 classes, 8 positives per class and 30 negatives sampled from the 10
nearest EC4 centroids, with centroids refreshed every 50 epochs. Training data comes from
the CARE benchmark's SwissProt split — 153,888 training and 18,293 validation proteins, all
CARE test proteins excluded from training, validation and the reference index — and
converges in 50–100 epochs on a single GPU. On CARE's four test sets, nearest-neighbour
retrieval reaches 64.6% top-1 EC4 accuracy on the remote-homology ood30 split (CLEAN
55.1%, Foldseek 54.9%, Diamond BLASTp 51.4%), 83.8% on ood30_50, 93.7% on the promiscuous
multi-functional set, and 41.9% on the Price activity-cliff set. Applied unchanged to
8,031 BRENDA-cross-referenced TrEMBL enzymes never seen in training, it reaches 97.3%
Precision@1 and holds 96.1% at Precision@10.
The intended use is functional annotation at database scale: embed a collection once, store the 256-dimensional projections, and answer EC queries by vector similarity instead of running alignment and rule-based annotation pipelines over tens of millions of uncharacterised sequences. That suits metagenomic surveys, where most sequences have no detectable similarity to any characterised organism, and enzyme discovery workflows needing candidate catalytic assignments in the twilight zone. Because the fusion recipe is agnostic to the label, the same head can be retrained against Gene Ontology terms, substrate classes or thermostability bins without changing the retrieval infrastructure.
FuncSeek's contribution is a clean demonstration that complementary PLMs carry genuinely non-redundant functional signal, and that extracting it needs a learned metric rather than concatenation: contrastive training on ESM-2 alone lifts Precision@1000 from 51.6% to 87.1%, yet still trails the three-encoder fusion at 96.5%. The honest boundary is substrate specificity — accuracy falls from 95.9% at the top EC level to 41.9% on activity cliffs, because none of the three inputs encodes binding-pocket geometry at residue resolution. Training also uses only the first EC annotation per protein, so multi-functionality is handled at inference rather than learned. The work is an unreviewed preprint reporting its own benchmark figures, and no code, weights or projected index has been released, so the retrieval paradigm it proposes is not yet reproducible outside the authors' lab.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.