Research map / Probabilistic and scientific modeling / Molecules and proteins
EnzyControl: Adding Functional and Substrate-Specific Control for Enzyme Backbone Generation
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 EnzyBind (11,100 PDBbind enzyme-substrate pairs) and EnzyControl, which adds a substrate cross-attention adapter (EnzyAdapter) and MSA-derived functional-site conditioning to the FrameFlow motif-scaffolding model for substrate-aware enzyme backbone generation.
Problem
Generating enzyme backbones that preserve catalytic functional sites and are tailored to a specific substrate, which existing motif-scaffolding and enzyme generation models do not handle, and which lack experimentally grounded benchmarks.
Approach
A lightweight EnzyAdapter with separate cross-attention over Uni-Mol substrate embeddings is inserted into each layer of a pretrained FrameFlow motif-scaffolding network, trained in two stages (adapter-only, then LoRA plus adapter with generation loss), with MSA-identified conserved residues as functional-site motifs.
Key results
- EnzyControl reaches 0.7160 designability on EnzyBind, a 13% relative improvement over the second-best model (Proteina at 0.6328).
- kcat of 2.9168 and EC match rate of 0.5041, outperforming the second-best model by about 13% and 10% respectively.
- Ablation shows removing EnzyAdapter lowers EC match rate by 5.55% and kcat by 11.43% relative to the full model.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add EnzyAdapter for substrate-aware conditioning (ablation) | EnzyControl on EnzyBind · EnzyBind | EC Match Rate: 0.4761 -> 0.5041 | Helped |
| Add EnzyAdapter for substrate-aware conditioning (ablation) | EnzyControl on EnzyBind · EnzyBind | kcat: 2.5833 -> 2.9168 | Helped |
| Use MSA functional-site annotations (ablation) | EnzyControl on EnzyBind · EnzyBind | kcat: 2.4615 -> 2.9168 | Helped |
| Use unperturbed rather than 50%-perturbed motif residues (ablation) | EnzyControl on EnzyBind · EnzyBind | EC Match Rate: 0.4918 -> 0.5041 | Helped |
| Use unperturbed rather than 50%-perturbed motif residues (ablation) | EnzyControl on EnzyBind · EnzyBind | Binding affinity: -6.61 -> -6.93 | Helped |
| Add EnzyControl rather than use EnzyGen | enzyme backbone generation · EnzyBind | Designability: 0.6328 -> 0.7160 | Helped |
| Add EnzyControl rather than use EnzyGen | enzyme backbone generation · EnzyBench | Binding Affinity: -9.44 -> -9.76 | Helped |
| Add EnzyControl rather than use EnzyGen | enzyme backbone generation · EnzyBench | ESP score: 0.65 -> 0.63 | Hurt |
Surprising findings
- Simply concatenating substrate features into IPA layers was found to be insufficiently effective, motivating the separate cross-attention design.
- Diversity and novelty lag behind RFDiffusion and Chroma, which the authors attribute to larger, more heterogeneous training sets.
Limitations
- Struggles to balance diversity with designability.
- Models only the enzyme backbone, not the substrate-bound conformation; substrate is represented as a chemical graph rather than a 3D pose.
- Restricted to single-chain enzymes and single substrates in the main method.
Experimental setup
Datasets: EnzyBind PDBbind
Benchmarks: EnzyBind EnzyBench
Metrics: scTM designability (scRMSD<2Å) EC match rate kcat binding affinity ESP score diversity novelty success rate
Baselines: EnzyGen PROTSEED RFDiffusion Chroma FADiff RFDiffusionAA Proteus Proteina
Builds on: FrameFlow EnzyGen LoRA Uni-Mol
Closest papers by meaning
- Generative Enzyme Design Guided by Functionally Important Sites and Small-Molecule Substrates (ICML 2024 · similarity 0.91)
- Retrieval Augmented Zero-Shot Enzyme Generation for Specified Substrate (ICML 2025 · similarity 0.87)
- ReactZyme: A Benchmark for Enzyme-Reaction Prediction (NeurIPS 2024 · similarity 0.85)
- ProteinBench: A Holistic Evaluation of Protein Foundation Models (ICLR 2025 · similarity 0.84)
- CLIPZyme: Reaction-Conditioned Virtual Screening of Enzymes (ICML 2024 · similarity 0.84)
- SurfPro: Functional Protein Design Based on Continuous Surface (ICML 2024 · 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.