Research map / Probabilistic and scientific modeling / Molecules and proteins
UniZyme: A Unified Protein Cleavage Site Predictor Enhanced with Enzyme Active-Site Knowledge
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes UniZyme, a unified protein cleavage-site predictor that uses a biochemically-informed (energy frustration + distance) graph transformer enzyme encoder, active-site pretraining, and active-site-aware pooling to generalize across seen and unseen proteolytic enzymes.
Problem
Predicting protease cleavage sites in substrate proteins generalizes poorly: existing models are enzyme-specific and cannot handle enzymes without labeled cleavage data, while MEROPS covers few enzymes.
Approach
Encode enzymes with a transformer whose attention is biased by residue-pair energetic frustration and C-alpha distances, pretrain the enzyme encoder on active-site prediction over a large UniProt/MEROPS protease set, and pool enzyme residues with weights from predicted active-site probabilities, then jointly train cleavage and active-site losses.
Key results
- Supervised setting over 69 enzyme families: UniZyme average PR-AUC 79.3% vs ClipZyme 74.7% and ReactZyme 70.0%; deep baselines ProsperousPlus, DeepCleave, DeepDigest under 7.2%.
- Zero-shot setting over 23 unseen enzyme families: UniZyme average PR-AUC 71.1% vs ReactZyme 64.7% and ClipZyme 61.7%.
- Ablation: removing energy/structure (SE) drops supervised PR-AUC to 75.3% and zero-shot to 65.3%; removing active-site pooling or pretraining also reduces performance.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use UniZyme instead of ClipZyme | cleavage-site prediction · Supervised, 69 enzyme families | Average PR-AUC (%): 74.7 -> 79.3 | Helped |
| use UniZyme instead of ReactZyme | cleavage-site prediction · Supervised, 69 enzyme families | Average PR-AUC (%): 70.0 -> 79.3 | Helped |
| use UniZyme instead of ReactZyme | cleavage-site prediction · Zero-shot, 23 enzyme families | Average PR-AUC (%): 64.7 -> 71.1 | Helped |
| use UniZyme instead of ClipZyme | cleavage-site prediction · Zero-shot, 23 enzyme families | Average PR-AUC (%): 61.7 -> 71.1 | Helped |
| add energy frustration and 3D structure to the enzyme encoder (ablation) | UniZyme · Supervised, 69 enzyme families | PR-AUC (%): 75.3±1.9 -> 79.3±1.2 | Helped |
| add energy frustration and 3D structure to the enzyme encoder (ablation) | UniZyme · Zero-shot, 23 enzyme families | PR-AUC (%): 65.3±1.4 -> 71.1±2.3 | Helped |
| use active site-aware pooling instead of average pooling (ablation) | UniZyme · Supervised, 69 enzyme families | PR-AUC (%): 72.5±1.5 -> 79.3±1.2 | Helped |
| use active site-aware pooling instead of average pooling (ablation) | UniZyme · Zero-shot, 23 enzyme families | PR-AUC (%): 60.6±1.8 -> 71.1±2.3 | Helped |
| pretrain the enzyme encoder on the supplemented active-site dataset (ablation) | UniZyme · Supervised, 69 enzyme families | PR-AUC (%): 73.0±2.2 -> 79.3±1.2 | Helped |
| pretrain the enzyme encoder on the supplemented active-site dataset (ablation) | UniZyme · Zero-shot, 23 enzyme families | PR-AUC (%): 62.1±2.7 -> 71.1±2.3 | Helped |
| increase the active-site prediction loss weight to λ=100 (ablation) | UniZyme · Hyperparameter analysis on enzyme families | PR-AUC | Mixed |
| reduce the active-site prediction loss weight to 0.1 or 0.01 (ablation) | UniZyme · Hyperparameter analysis on enzyme families | ROC-AUC | Hurt |
Surprising findings
- Individual-enzyme deep baselines (ProsperousPlus, DeepCleave, DeepDigest) perform very poorly (under 7.2% average PR-AUC) when retrained per enzyme family.
Limitations
- Relies on predicted (AlphaFold/OmegaFold) structures for many enzymes and substrates; zero-shot performance drops with generated structures.
- Restricted to proteolytic enzymes; extension to other enzyme categories is future work.
- Lower accuracy on metalloprotease families such as M10.003, attributed to higher substrate diversity and metal-dependent catalysis.
Experimental setup
Datasets: MEROPS UniProt PDB AlphaFoldDB
Benchmarks: MEROPS supervised split (69 enzyme families) MEROPS zero-shot split (23 enzyme families) HIV-1 protease case study
Metrics: PR-AUC ROC-AUC rank-1 rate average rank
Baselines: ClipZyme ReactZyme ProsperousPlus DeepCleave DeepDigest CAT3 ScreenCap3
Builds on: Transformer-M ESM-2 AWSEM / Frustratometer ClipZyme ReactZyme
Closest papers by meaning
- UniSite: The First Cross-Structure Dataset and Learning Framework for End-to-End Ligand Binding Site Detection (NeurIPS 2025 · similarity 0.85)
- ReactZyme: A Benchmark for Enzyme-Reaction Prediction (NeurIPS 2024 · similarity 0.84)
- CLIPZyme: Reaction-Conditioned Virtual Screening of Enzymes (ICML 2024 · similarity 0.84)
- CARE: a Benchmark Suite for the Classification and Retrieval of Enzymes (NeurIPS 2024 · similarity 0.83)
- Retrieval Augmented Zero-Shot Enzyme Generation for Specified Substrate (ICML 2025 · similarity 0.83)
- Generative Enzyme Design Guided by Functionally Important Sites and Small-Molecule Substrates (ICML 2024 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.