Research map / Probabilistic and scientific modeling / Molecules and proteins
ReactZyme: A Benchmark for Enzyme-Reaction Prediction
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.
Introduces ReactZyme, a large enzyme-reaction dataset (SwissProt + Rhea, 178k pairs) and retrieval benchmark with three splits, evaluating graph/structure encoders with ESM and SaProt embeddings against BLAST.
Problem
Enzyme annotation via EC/GO/KO categories is coarse and ambiguous, so the paper targets predicting which catalyzed reactions an enzyme performs (and which enzymes catalyze a given reaction) as a retrieval problem.
Approach
Frames enzyme-reaction prediction as ranking enzymes for a reaction (and reactions for an enzyme) using a large SwissProt/Rhea dataset, combining multi-view reaction encoders (2D/3D molecular graphs, cross-attention between substrates and products) with enzyme encoders that use ESM and structure-aware SaProt embeddings plus SE(3)-invariant frame-averaging GNNs, followed by an MLP pair scorer.
Key results
- On the time-based split, MAT-2D + ESM reaches 32.46% top-1 accuracy and 0.455 MRR when ranking reactions per enzyme; UniMol-3D + ESM with GNN encoding is among the top models.
- On the enzyme-similarity split, MAT-2D + SaProt reaches 66.91% top-1 accuracy for ranking reactions per enzyme, versus 21.75% top-1 for ranking enzymes per reaction on the time split with MAT-2D + ESM.
- On the reaction-similarity split, performance drops sharply (best top-1 around 9-14%), and BLAST is competitive on the time split for ranking enzymes per reaction only at higher ranks, underscoring the difficulty of unseen reactions.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace the MLP decoder with a Bi-RNN decoder (ablation) | MAT-2D + ESM · Time split, enzyme-to-reaction | Top1: 0.3246 -> 0.3911 | Helped |
| replace the MLP decoder with a Bi-RNN decoder (ablation) | MAT-2D + ESM · Time split, reaction-to-enzyme | Top1: 0.2175 -> 0.2650 | Helped |
| replace the MLP decoder with a Bi-RNN decoder (ablation) | MAT-2D + ESM · Enzyme-similarity split, enzyme-to-reaction | Top1: 0.5987 -> 0.8151 | Helped |
| replace the MLP decoder with a Bi-RNN decoder (ablation) | MAT-2D + ESM · Reaction-similarity split, enzyme-to-reaction | MRR: 0.2005 -> 0.2399 | Helped |
| add 3D molecular conformation and enzyme GNN encoding (ablation) | MAT-2D + ESM · Time split, enzyme-to-reaction | Top1: 0.3246 -> 0.3588 | Helped |
| add 3D molecular conformation and enzyme GNN encoding (ablation) | MAT-2D + ESM · Reaction-similarity split, enzyme-to-reaction | Top1: 0.0914 -> 0.1104 | Helped |
| use SaProt instead of ESM | MAT-2D enzyme-reaction retrieval · Enzyme-similarity split, enzyme-to-reaction | Top1: 0.5987 -> 0.6691 | Helped |
| use SaProt instead of ESM | MAT-2D enzyme-reaction retrieval · Time split, enzyme-to-reaction | Top1: 0.3246 -> 0.2073 | Hurt |
Surprising findings
- Structure-aware SaProt does not consistently beat vanilla ESM; in several settings ESM-based models outperform SaProt-based ones.
- BLAST sequence search performs strongly in some settings (e.g., top-1 of 35.81% for ranking reactions on the time split), comparable to neural models.
Limitations
- Lacks atom-mapping data, and reaction count is smaller than EnzymeMap because some reactions are represented by functional groups.
- Reaction-similarity split remains very challenging; neither Transformer nor Bi-RNN decoders give substantial gains there.
- Negative sample construction via top-k similarity is an open question.
Experimental setup
Datasets: SwissProt Rhea ReactZyme ESP EnzymeMap
Benchmarks: ReactZyme time-based split ReactZyme enzyme-similarity split ReactZyme reaction-similarity split
Metrics: top-k accuracy top-k accuracy-N mean rank MRR
Baselines: BLAST MAT-2D + ESM UniMol-3D + ESM ESP CLIPZyme
Builds on: CLIPZyme CLEAN SaProt Molecule Attention Transformer Uni-Mol
Closest papers by meaning
- CARE: a Benchmark Suite for the Classification and Retrieval of Enzymes (NeurIPS 2024 · similarity 0.90)
- CLIPZyme: Reaction-Conditioned Virtual Screening of Enzymes (ICML 2024 · similarity 0.88)
- Generative Enzyme Design Guided by Functionally Important Sites and Small-Molecule Substrates (ICML 2024 · similarity 0.87)
- EnzyControl: Adding Functional and Substrate-Specific Control for Enzyme Backbone Generation (NeurIPS 2025 · similarity 0.85)
- Retrieval Augmented Zero-Shot Enzyme Generation for Specified Substrate (ICML 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.