UniZyme

Enzyme-catalyzed protein cleavage underpins everything from zymogen activation to apoptosis, and knowing where a given protease cuts a given substrate matters for drug development, proteomics workflows, and enzyme engineering. Computational cleavage-site predictors have historically been built one enzyme at a time: a separate model for caspase-3, another for trypsin, each requiring enough experimental substrate data to train on. That design ignores what proteases share, and it fails outright for the vast majority of enzymes with too few annotated substrates to fit a model.

UniZyme, from the Hong Kong University of Science and Technology (Guangzhou), replaces that per-enzyme habit with a single predictor conditioned on the enzyme itself. Both enzyme and substrate are encoded from sequence via ESM-2, and the enzyme encoder is enriched with the property that actually determines specificity: the structure and chemistry of the catalytic active site. Because the model reads an enzyme rather than memorizing one, it can be handed a protease it has never seen and still predict where it cuts.

The insight that makes this practical is an asymmetry in data availability. Experimental cleavage-site assays are expensive, so MEROPS covers only a few hundred enzymes, but active-site annotation is comparatively cheap and UniProt lists thousands of proteolytic enzymes with labelled active sites. UniZyme uses the plentiful annotation to compensate for the scarce one. The work was published at NeurIPS 2025.

#Key Features

• One model across enzyme families: A single trained checkpoint covers hundreds of proteolytic enzymes, in place of the separate per-enzyme predictors that prior deep learning methods require.
• Active-site pretraining: An auxiliary active-site prediction task, trained on a much larger supplementary enzyme set, teaches the encoder general catalytic patterns before cleavage-site supervision begins.
• Active site-aware pooling: Residue representations are aggregated into an enzyme embedding with weights derived from predicted active-site probabilities, so catalytic residues dominate the representation instead of being averaged away.
• Biochemically informed encoder: A graph transformer incorporates energetic frustration scores and 3D positional information through Gaussian basis kernels, capturing the local environment that lowers the activation energy for peptide bond hydrolysis.
• Zero-shot inference: Test enzymes' active sites are not supplied at inference time; the model predicts them, so a novel protease needs only its sequence and structure.

#Technical Details

Residue features for both enzyme and substrate come from ESM-2 t12 35M, producing 480-dimensional embeddings. Cleavage-site data is drawn from MEROPS, which annotates 866 proteolytic enzymes across roughly 10,146 substrate proteins; propagating substrate-site annotations within enzyme families expands this to about 220k enzyme–substrate pairs. The supplementary active-site set combines MEROPS annotations with UniProt hydrolases (EC 3.4.*.*) homologous to those enzymes, yielding 11,530 enzymes with labelled active sites. Training and evaluation split MEROPS into 677 enzyme families for training, 69 families for the supervised test, and 23 held-out families for the zero-shot test, whose enzymes share under 60% sequence identity with anything seen during training or pretraining. Cleavage-site and active-site losses are optimized jointly.

In the supervised setting UniZyme averages 79.3% PR-AUC across 69 families and ranks first in 75.0% of them, against 74.7% for a ClipZyme-encoder variant, 70.0% for a ReactZyme variant, and single-digit scores for the per-enzyme baselines ProsperousPlus, DeepCleave, and DeepDigest. On the 23 zero-shot families it reaches 71.1% average PR-AUC with a 78.3% rank-1 rate, versus 64.7% and 61.7% for the ReactZyme and ClipZyme variants.

#Applications

UniZyme is useful wherever proteolysis needs to be reasoned about but experimental substrate data is thin: mapping candidate cleavage sites for a newly characterized protease, identifying which host proteins a pathogen enzyme might process, designing protease-resistant therapeutic proteins, or planning inhibitor campaigns. The paper's worked example applies it to HIV-1 proteases to enumerate cleavage sites across arbitrary substrate proteins, exactly the setting where per-enzyme predictors cannot be trained.

#Impact

UniZyme's contribution is showing that enzyme specificity can be learned as a function of active-site chemistry rather than memorized per enzyme, which converts cleavage-site prediction from a collection of narrow tools into one transferable model. The margin over encoders borrowed from enzyme–substrate reaction predictors isolates the active-site knowledge as the source of the gain. Practical caveats: the released repository ships code and evaluation scripts under the MIT license with datasets and weights distributed separately through Zenodo under CC-BY 4.0, inference requires enzyme and substrate structures plus precomputed frustration scores rather than sequence alone, and the model size is not reported.