Research map / Probabilistic and scientific modeling / Molecules and proteins

Retrieval Augmented Zero-Shot Enzyme Generation for Specified Substrate

ICML 2025 · Probabilistic and scientific modeling › Molecules and proteins · approach: Structure-based molecule design with diffusion

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 SENZ, a retrieval-augmented discrete diffusion model that retrieves enzymes via substrate similarity and generates new enzymes for zero-shot target substrates, guided by an enzyme-substrate classifier, plus a new substrate-enzyme dataset from RHEA.

Problem

Generating enzymes for new substrate molecules with no known enzymes (zero-shot), where no positive training samples exist and generated enzymes must diverge from recorded ones.

Approach

Retrieve enzymes whose substrates are structurally similar to the target molecule (rather than similar proteins), align them as an MSA, and train an order-agnostic discrete diffusion generator with an MSA transformer conditioned on the molecule, with a frozen substrate-enzyme discriminator providing a differentiable guidance loss.

Key results

• SENZ achieves the highest average predicted log10(kcat) (0.380) on the test set, above ground truth (0.363) and ProGen2 (0.352), with pLDDT 62.8 vs ZymCTRL 62.5.
• Retrieval improves catalytic capability: more retrieved enzymes raise predicted kcat at the cost of foldability (pLDDT).
• Removing discriminator guidance leaves generated kcat roughly equal to the retrieved enzymes, showing guidance is necessary for higher kcat but lowers pLDDT.

Measured effects reported in the paper

Surprising findings

• Generated enzymes are predicted to outperform ground truth natural enzymes in catalytic capability, suggesting natural enzymes may not be the most efficient.
• A trade-off exists between predicted catalytic capability and foldability as more retrieved enzymes or guidance are used.

Limitations

• Catalytic capability is evaluated only in silico with a kcat predictor (UniKP) rather than wet-lab validation.
• Foldability decreases as retrieval count and discriminator guidance increase, a stated trade-off.
• Variable-length handling relies on gap tokens, so output sequences can be shorter than the fixed length.

Experimental setup

Datasets: RHEA substrate-enzyme relation dataset

Benchmarks: Enzyme-Substrate Relation Dataset (RHEA-derived) test split

Metrics: predicted log10(kcat) via UniKP pLDDT (ESMFold) BLASTp identity Wasserstein distance cluster count difference repeat amino acid length Wasserstein distance success rate AutoDock-Vina docking score

Baselines: ProtGPT2 ProGen2 ZymCTRL NOS LigandMPNN Ground truth enzymes Random sequences Single-position mutation

Builds on: EvoDiff NOS MSA Transformer RSA RetMol

Closest papers by meaning

• Generative Enzyme Design Guided by Functionally Important Sites and Small-Molecule Substrates (ICML 2024 · similarity 0.90)
• EnzyControl: Adding Functional and Substrate-Specific Control for Enzyme Backbone Generation (NeurIPS 2025 · similarity 0.87)
• CLIPZyme: Reaction-Conditioned Virtual Screening of Enzymes (ICML 2024 · similarity 0.85)
• SurfPro: Functional Protein Design Based on Continuous Surface (ICML 2024 · similarity 0.85)
• CARE: a Benchmark Suite for the Classification and Retrieval of Enzymes (NeurIPS 2024 · similarity 0.85)
• Advancing Retrosynthesis with Retrieval-Augmented Graph Generation (AAAI 2025 · similarity 0.85)