EZpred

Deep learning transformed protein structure prediction in large part by exploiting sequence homologs: AlphaFold 2 builds its input features from a multiple sequence alignment, and single-sequence predictors such as ESMFold — released alongside ESM-2 — are measurably less accurate as a result. Protein function prediction never inherited that lesson. Template-based annotation transfer uses only homologs that already carry function labels; deep learning predictors overwhelmingly featurize the target sequence alone. Unannotated homologs — the overwhelming majority of sequence space — are discarded.

EZpred closes that gap for enzymes. Developed by Chengxin Zhang, Quancheng Liu, and Lydia Freddolino at the University of Michigan and the Shenzhen Institutes of Advanced Technology of the Chinese Academy of Sciences, and posted as a bioRxiv preprint in July 2025, it is the first deep learning model to draw input features from sequence homologs that lack function labels. A query is searched with MMseqs2 against a database of training enzymes; query and top hits are each embedded with the ESM Cambrian protein language model, and the pooled homolog representation is concatenated with the query's own before a classifier predicts Enzyme Commission (EC) numbers.

That branch is then fused with a classical template branch searching both sequence space (MMseqs2) and structure space (Foldseek over AlphaFold DB models), weighting each hit by how well it aligns to the template's catalytic pocket.

#Key Features

• Features from unlabeled homologs: Embeddings of the top four MMseqs2 hits are combined by a bit-score- and identity-weighted average, regardless of whether those homologs have any EC annotation. Adding them raises 4-digit F1 by 2% over an otherwise identical single-sequence model.
• Multi-layer PLM readout: Rather than taking only the final hidden state, EZpred concatenates mean-pooled embeddings from ESM Cambrian's last three layers — worth 2% over the last layer alone and 0.8% over using the last ten.
• Rare EC numbers retained: Every EC number in the training set is kept as a label, even those with a single supporting protein. Filtering to labels with at least 5 or 10 training enzymes costs 5% and 16% F1 respectively.
• Template fusion with catalytic-site weighting: Templates are scored by bit-score, identity, TM-score, and the fraction of the template's active- and substrate-binding-site pocket that aligns to the query.
• Explicit non-enzyme class: A parallel model head carries a 0.-.-.- label trained against 23,354 experimentally annotated non-enzymes, so the tool can decline to call an enzyme at all.

#Technical Details

EZpred embeds sequences with the 600M-parameter ESM Cambrian checkpoint, taking layers 34–36 of 36; each yields a 1152-dimensional mean-pooled vector, concatenated to 3456 dimensions per sequence and to 6912 after the homolog vector is appended. Sequences longer than 2048 residues are truncated to ESM Cambrian's context limit. A feed-forward network with GELU activations, 30% dropout, and batch size 512 maps this to 928 EC labels (7 first-digit, 64 two-digit, 188 three-digit, 665 four-digit), trained with a composite loss multiplying the ZLPR ranking loss by information-content-weighted protein-centric and EC-centric F1 terms; hierarchy-consistent scores are enforced by iterative post-processing. Training used 30,592 enzymes with experimental EC evidence from UniProt release 2023_02, clustered at 60% identity by CD-HIT and split into five folds.

On an independent set of 753 enzymes that gained EC annotations only in UniProt 2024_02, EZpred reaches F1 0.889 for the first three EC digits and 0.788 for all four — 27% above CLEAN and 32% above GraphEC, the strongest of eight benchmarked predictors. For enzyme-versus- non-enzyme discrimination across 753 enzymes and 625 non-enzymes it scores F1 0.911, against 0.780 for ECPred and 0.568 for DeepECtransformer. On the hardest third of targets (under 30% identity to training data) the template branch (0.638) edges out the deep learning branch (0.598).

#Applications

EZpred targets the routine bottleneck of annotating enzymes in newly sequenced genomes and metagenomes, where a 4-digit EC assignment determines which reaction a protein is credited with in a pathway or metabolic model. Its explicit non-enzyme class makes it usable on unfiltered proteomes rather than curated enzyme lists, and the template branch supplies interpretable evidence — a named homolog with an aligned catalytic pocket — alongside the network score, which matters when a prediction is going to be tested at the bench.

#Impact

EZpred's central claim is methodological rather than architectural: the classifier is a modest MLP over frozen embeddings, and the gains come from what is fed into it. By showing that unlabeled homologs carry usable signal for function, it imports into function prediction the assumption that has underpinned structure prediction since AlphaFold 2. The work remains a preprint awaiting peer review, and its authors note two open limitations — the deep learning branch uses no tertiary structure, and mean pooling weights active-site residues no more heavily than any other. Code is MIT-licensed, with trained weights and processed data on Zenodo.