Research map / Probabilistic and scientific modeling / Molecules and proteins
Generative Enzyme Design Guided by Functionally Important Sites and Small-Molecule Substrates
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 EnzyGen, a unified generative model that co-designs enzyme sequence and backbone coordinates conditioned on EC family tags, functionally important sites mined by MSA, and a substrate, trained on the new EnzyBench dataset from PDB/BRENDA.
Problem
Designing functional enzymes that bind specific small-molecule substrates across thousands of enzyme families, where prior methods are family-specific, lack substrate modeling, or depend on scarce fitness data.
Approach
Interleave ESM-2 global attention layers with neighborhood equivariant layers that update Cα coordinates and features from K-nearest neighbors, conditioned on EC tags and MSA-mined functional sites, with an added enzyme-substrate interaction prediction loss.
Key results
- EnzyGen achieves average ESP score 0.65 vs 0.61 for ESM2+EGNN and 0.59 for PROTSEED across 30 third-level categories (323 fourth-level families).
- EnzyGen reaches average Gnina binding affinity of -9.44 kcal/mol, improving on the best baseline (RFDiffusion+ProteinMPNN, -8.57) by about 0.87 kcal/mol (~10.79% per the abstract).
- EnzyGen averages pLDDT 87.45 on AlphaFold2, higher than all three baselines.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use EnzyGen instead of RFDiffusion+ProteinMPNN | enzyme design · 30 testing BRENDA EC categories | ESP score: 0.53 -> 0.65 | Helped |
| use EnzyGen instead of RFDiffusion+ProteinMPNN | enzyme design · 30 testing BRENDA EC categories | binding affinity (kcal/mol): -0.87 | Helped |
| use EnzyGen instead of ESM2+EGNN | enzyme design · 30 testing BRENDA EC categories | pLDDT: 85.13 -> 87.45 | Helped |
| add masked language model pretraining (ablation) | EnzyGen · 30 testing BRENDA EC categories | ESP score: 0.65 -> 0.67 | Helped |
| add masked language model pretraining (ablation) | EnzyGen · 30 testing BRENDA EC categories | binding affinity (kcal/mol): -9.44 -> -9.58 | Helped |
| add masked language model pretraining (ablation) | EnzyGen · 30 testing BRENDA EC categories | pLDDT: 87.45 -> 86.15 | Hurt |
| include the interleaving NAEL network rather than directly concatenating ESM2 and EGNN (ablation) | enzyme design · 323 testing families | ESP score | Helped |
| include the interleaving NAEL network rather than directly concatenating ESM2 and EGNN (ablation) | enzyme design · 323 testing families | pLDDT | Helped |
| add the enzyme-substrate interaction constraint to training (ablation) | EnzyGen · enzyme-substrate binding | binding affinity | Helped |
| increase model scale from 117M to 714M parameters (ablation) | EnzyGen · 323 testing categories | ESP score | Helped |
| finetune on each third-level category for 30 epochs (ablation) | EnzyGen · 30 testing BRENDA EC categories | ESP score: 15 categories exhibit improvements; 8 categories show degradation | Mixed |
Surprising findings
- Zero-shot generalization to unseen substrates (Mg2+, Zn2+) and unseen fourth-level families yields average ESP 0.83-0.89 without training on them.
- Finetuning per category improved ESP for 15 categories but degraded 8, mostly those with fewer than 1000 training samples.
Limitations
- Dataset covers only 3,157 of 8,422 fourth-level EC categories with experimentally confirmed structures.
- Substrate-binding constraint is a weak proxy for catalytic function; no wet-lab validation.
- Binding evaluated via ESP predictor and docking (Gnina) rather than experiments.
Experimental setup
Datasets: EnzyBench BRENDA PDB enzyme-substrate pairs (BRENDA-derived)
Benchmarks: EnzyBench
Metrics: ESP score substrate binding affinity (Gnina, kcal/mol) AlphaFold2 pLDDT RMSD self-consistency Uniprot blastp identity
Baselines: ESM2+EGNN PROTSEED RFDiffusion+ProteinMPNN
Builds on: ESM-2 EGNN Transformer ProteinMPNN
Closest papers by meaning
- EnzyControl: Adding Functional and Substrate-Specific Control for Enzyme Backbone Generation (NeurIPS 2025 · similarity 0.91)
- Retrieval Augmented Zero-Shot Enzyme Generation for Specified Substrate (ICML 2025 · similarity 0.90)
- ReactZyme: A Benchmark for Enzyme-Reaction Prediction (NeurIPS 2024 · similarity 0.87)
- SurfPro: Functional Protein Design Based on Continuous Surface (ICML 2024 · similarity 0.86)
- CLIPZyme: Reaction-Conditioned Virtual Screening of Enzymes (ICML 2024 · similarity 0.85)
- BindGPT: A Scalable Framework for 3D Molecular Design via Language Modeling and Reinforcement Learning (AAAI 2025 · similarity 0.85)
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.