Research map / Label-efficient and robust learning / Recognition and long-tail
CLIPZyme: Reaction-Conditioned Virtual Screening of Enzymes
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 CLIPZyme, a contrastive model aligning reaction encodings (pseudo-transition-state DMPNN over atom-mapped reactants/products) with AlphaFold-structure enzyme encodings (EGNN over ESM-2 features) for virtual enzyme screening, compared against CLEAN EC-based ranking.
Problem
Virtual screening of naturally occurring enzymes for a query chemical reaction, ranking enzymes by predicted catalytic activity from a large uncharacterized sequence set, where no standard computational baseline exists.
Approach
Align reaction and enzyme embeddings with a CLIP-style contrastive objective, encoding a pseudo-transition state by summing bond features of reactant and product graphs, and encoding enzymes as 3D graphs with EGNN over ESM-2 residue features, enabling retrieval of enzymes for unseen reactions.
Key results
- CLIPZyme achieves BEDROC85 of 44.69% on EnzymeMap test screening, versus 25.86% for CLEAN given three EC levels.
- Combining CLEAN with CLIPZyme raises BEDROC85 to 57.03% (level 1), 75.57% (level 2), and 82.69% (level 3) of EC levels assumed known.
- On terpene synthase reactions CLIPZyme reaches BEDROC85 of 72.46%; excluding Foldseek-similar enzymes (30%) drops BEDROC85 to 21.44%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CLIPZyme instead of CLEAN for enzyme ranking when the reaction EC is known to Level 1 | Virtual enzyme screening · EnzymeMap test set, Level 1 EC known | BEDROC85 (%): 0.96 -> 57.03 | Helped |
| Combine CLEAN EC grouping with CLIPZyme reranking | CLIPZyme · EnzymeMap test set, Level 1 EC known | BEDROC85 (%): 44.69 to 57.03 | Helped |
| Combine CLEAN EC grouping with CLIPZyme reranking | CLIPZyme · EnzymeMap test set, Level 2 EC known | BEDROC85 (%): 75.57 | Helped |
| Combine CLEAN EC grouping with CLIPZyme reranking | CLIPZyme · EnzymeMap test set, Level 3 EC known | BEDROC85 (%): 25.86 -> 82.69 | Helped |
| Combine CLEAN EC grouping with CLIPZyme reranking | CLIPZyme · EnzymeMap test set, Level 4 EC known | BEDROC85 (%): 89.74 -> 89.57 | No clear effect |
| Replace Reaction SMILES with the structural reaction encoder (ablation) | CLIPZyme with frozen ESM + EGNN · EnzymeMap test set | BEDROC85 (%): 29.94% compared to 44.69% | Helped |
| Replace WLDN with the structural reaction encoder (ablation) | CLIPZyme with frozen ESM + EGNN · EnzymeMap test set | BEDROC85 (%): 29.84 -> 44.69 | Helped |
| Add EGNN structure encoding to sequence-based ESM (ablation) | CLIPZyme with our reaction encoder · EnzymeMap test set | BEDROC85 (%): 36.91% compared to 44.69% | Helped |
| Initialize EGNN node features with MSA-Transformer embeddings instead of ESM embeddings (ablation) | CLIPZyme protein encoder · EnzymeMap test set | BEDROC85 (%): 28.76 versus 44.69 | Hurt |
| Exclude exact training-sequence matches from the screening set | CLIPZyme · EnzymeMap test set | BEDROC85 (%): 39.13 | Hurt |
| Exclude enzymes with 30% or greater Foldseek fold similarity to training enzymes | CLIPZyme · EnzymeMap test set | BEDROC85 (%): 21.44 | Hurt |
Surprising findings
- With all four EC levels known, CLEAN alone (89.74% BEDROC85) outperforms CLIPZyme alone (44.69%).
- Using MSA-Transformer embeddings for the EGNN did not improve over ESM-2 features despite capturing evolutionary information.
Limitations
- Does not model physical interactions between reactants and enzymes or catalytic mechanisms.
- Limited chemical and enzyme sequence coverage (e.g., EC class 7 unrepresented).
- Random negative sampling may produce false negatives due to enzyme promiscuity.
Experimental setup
Datasets: EnzymeMap BRENDA UniProt AlphaFold Protein Structure Database Terpene synthase dataset
Benchmarks: EnzymeMap test set EnzymeMap unannotated reactions Terpene synthase reactions
Metrics: BEDROC85 BEDROC20 EF0.05 EF0.1
Baselines: CLEAN CLEAN + CLIPZyme Condensed graph reaction encoder Reaction SMILES transformer WLDN ESM-2 sequence-only encoder MSA-Transformer + EGNN
Builds on: CLIP DMPNN EGNN CLEAN
Closest papers by meaning
- ReactZyme: A Benchmark for Enzyme-Reaction Prediction (NeurIPS 2024 · similarity 0.88)
- CARE: a Benchmark Suite for the Classification and Retrieval of Enzymes (NeurIPS 2024 · similarity 0.87)
- Retrieval Augmented Zero-Shot Enzyme Generation for Specified Substrate (ICML 2025 · similarity 0.85)
- Generative Enzyme Design Guided by Functionally Important Sites and Small-Molecule Substrates (ICML 2024 · similarity 0.85)
- EnzyControl: Adding Functional and Substrate-Specific Control for Enzyme Backbone Generation (NeurIPS 2025 · similarity 0.84)
- UniZyme: A Unified Protein Cleavage Site Predictor Enhanced with Enzyme Active-Site Knowledge (NeurIPS 2025 · similarity 0.84)
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.