EnzyControl: Adding Functional and Substrate-Specific Control for Enzyme Backbone Generation

1 Introduction

• Enzyme design requires substrate-specific binding and conserved functional sites, but prior approaches neglect functional-site conservation, substrate interactions during generation, or enzyme-family-tailored benchmarks.The paper notes that filtering by substrate after generation does not incorporate substrate information into the generation process.Source: §1 · p. 2
• EnzyControl combines MSA-identified conserved functional motifs, substrate conditioning through EnzyAdapter, and two-stage training to generate enzyme backbones.The adapter uses a cross-modal projector and cross-attention; it is trained first, followed by LoRA fine-tuning of the full model.Source: §1 · p. 2
• EnzyBind provides 11,100 curated enzyme-substrate pairs with MSA functional-site annotations and an enzyme-family-based EC consistency evaluation framework.The framework evaluates whether generated samples match the target native enzyme’s EC number.Source: §1 · p. 2
• 0.7160 of generated backbones are designable, a 13% relative improvement over the second-best model; EnzyControl also reports 13% higher catalytic efficiency than suboptimal baselines.The introduction also reports a 10% improvement in EC match rate and a 3% binding-affinity improvement over the second-best model on EnzyBench.Source: §1 · p. 2

Enzyme Design Applications

• Semi-rational and rational design optimize or repurpose natural enzymes, whereas de novo design embeds catalytic motifs into synthetic scaffolds.EnzyControl is described as bridging rational and de novo design with substrate-aware backbone generation and MSA-based active-site annotation.Source: §2 · p. 3

Methods for Motif-Scaffolding Task

• Motif-scaffolding designs the remainder of a protein—the scaffold—around a prespecified functional-residue arrangement, but existing generative models cannot incorporate substrate specificity.RFdiffusion and FrameFlow condition scaffold generation on motif structure and/or sequence.Source: §2 · p. 3

3 EnzyBind: High-Quality Enzyme Dataset

• EnzyBind provides experimentally validated enzyme-substrate complexes with precise pocket structures and substrate conformations for catalytic backbone generation.The dataset addresses the lack of precise pocket information in sequence-and-SMILES datasets and the lack of wet-lab validation in a synthetic dataset.Source: §3 · p. 3
• The PDBBind curation and preprocessing pipeline yields approximately 11,000 complexes across six fundamental catalytic types.Complexes that could not be processed by RDKit were excluded, and remaining PDB files were cleaned using a standardized procedure.Source: §3 · p. 3, Figure 1 · p. 3, Figure 2 · p. 3
• MSA with MAFFT identifies evolutionarily conserved regions among enzymes sharing a second-level EC number to annotate functional sites.The method builds on prior work [31].Source: §3 · pp. 3–4

4.1 Preliminary: Flow Matching

• Flow matching estimates a vector field between noise and data distributions, then enables backward sampling by numerically integrating the learned field.The model estimates the field with a least-squares objective.Source: §4.1 · p. 4

4.2 EnzyControl’s Base Network

• The model generates scaffold residues conditioned on functional sites and a substrate graph, targeting a backbone that preserves the sites and fits the substrate.Its backbone representation uses residue frames in SE(3).Source: §4.2 · p. 4

4.3 EnzyAdapter and Two-Stage Training Strategy

• EnzyAdapter is modular and model-agnostic, using per-layer cross-attention to inject substrate features separately from enzyme node features.The authors report that simply concatenating substrate and node features was insufficiently effective.Source: §4.3 · pp. 5–6
• Substrates are represented as chemical graphs, encoded by frozen pretrained Uni-Mol, then projected into the enzyme representation space.The encoder is frozen because the downstream dataset contains approximately 11,000 enzyme–substrate pairs.Source: §4.3 · p. 5
• Training first updates the projector and EnzyAdapter with the prediction network frozen, then fine-tunes the network and enzyme embedder using LoRA while continuing adapter updates.The first stage aligns substrates with corresponding enzymes; the second uses generation loss for joint optimization.Source: §4.3 · p. 6

5.1 Setup

• Evaluation uses a unified ProteinMPNN-to-ESMFold pipeline to compare EnzyGen and motif-scaffolding baselines on structural, functional, binding, and multi-objective success measures.Success requires matching the native EC number, scTM > 0.5, scRMSD < 2Å, and better binding affinity than the native counterpart.Source: §5.1 · pp. 6–7

5.2 Main Results

• EnzyControl reaches 0.7160 designability versus Proteina’s 0.6328 and exceeds the second-best model by 10% in EC match rate and 15% in kcat.The comparison covers structural and functional metrics on EnzyBind.Source: Table 1 · p. 7, §5.2 · p. 7
• Binding affinity improves by 2.8% and ESP score by 3.1% over the second-best method.Source: §5.2 · p. 7

5.3 Ablation Study and Analysis

• At 50% motif perturbation, EC match rate falls 2.4%, predicted kcat drops 9.0%, and binding affinity changes from −6.93 to −6.61 kcal/mol.Source: Table 4 · p. 8, §5.3 · p. 9
• EnzyControl achieves state-of-the-art EC match rates at the third and fourth levels, although current annotations use second-level EC categories.Source: Figure 7 · p. 8, §5.3 · p. 9
• Removing MSA input or EnzyAdapter degrades sample quality, with adapter removal producing the largest drop in EC match rate.The ablation figures compare kcat and binding-affinity distributions; adapter removal shifts kcat left and alters the binding curve particularly around -6.Source: §5.3 · pp. 8–9, Figure 5 · p. 8, Figure 6 · p. 8

5.4 Quantitative Analysis of Enzyme Function

• Generated sequences are approximately 30% shorter than RFDiffusion’s at comparable kcat, while the 2cv3 case shows 51% better binding affinity and nearly 8× catalytic efficiency.Docking also found more substrate interaction bonds for the EnzyControl sample than for RFDiffusion.Source: §5.4 · pp. 9–10, Figure 9 · p. 10, Figure 10 · p. 10
• On unseen EC categories and novel substrates, zero-shot binding affinity averages –7.0125 and –6.7292 kcal/mol, respectively.Source: §5.4 · p. 9, Figure 8 · p. 10

• EnzyControl achieves leading performance across structural and functional metrics on the EnzyBind benchmark.The work also compares EnzyControl against existing methods.Source: §6 · p. 10
• EnzyControl uses substrate information to guide backbone refinement and MSA-identified conserved motifs to help preserve catalytic features.Source: §6 · p. 10
• Additional analyses report strong generalization capabilities, including zero-shot performance.Source: §6 · p. 10

• EnzyControl generates enzyme backbones without modeling the conformations they adopt when bound to substrates.Source: A Limitations · p. 17
• The method produces diverse backbone samples but struggles to balance diversity with designability.Source: A Limitations · p. 17
• Improving designability without sacrificing diversity remains an open challenge and future direction.Source: A Limitations · p. 17

B More Experimental Results

• Tables 8–10 provide test-set enzyme-family descriptions and EC-family breakdowns for pLDDT and ESP comparisons.The reported results are broken down by EC family because the main-text results do not distinguish families.Source: B More Experimental Results · p. 17, Table 8 · p. 18, Table 9 · p. 19
• A post-hoc pipeline uses RFDiffusion binder design to add a complementary partner to a generated single-chain enzyme without modifying EnzyControl’s core architecture.Source: B More Experimental Results · p. 17
• Preliminary results indicate stable complex formation between a designed enzyme and its binder.Source: B More Experimental Results · p. 17
• The proposed multi-substrate extension generates a substrate-aware embedding for each substrate, aggregates them, and passes them to the backbone generator.The authors propose this direction for supporting multi-step or cofactor-dependent catalytic reactions.Source: B More Experimental Results · p. 17
• The current model represents substrate input as an embedding rather than through explicit spatial modeling, motivating two explored substrate-aware optimization strategies.One strategy selects among multiple generated backbones by docking affinity; the other branches from an annotated catalytic motif, completes and docks candidates, then reuses the best-scoring motif variant.Source: B More Experimental Results · p. 18
• Predicted binding affinity changed from -6.92 before optimization to -8.38 afterward in a prototype docking-guided strategy.The authors describe docking-guided optimization as substantially improving substrate compatibility and suggest integrating it into the main generative loop as future work.Source: B More Experimental Results · p. 18

C Enzyme Commission number

• EC numbers classify enzymes by the chemical reactions they catalyze, and EnzyControl uses these categories to guide backbone generation with specific functions.The figure illustrates EC numbers in BRENDA.Source: C Enzyme Commission number · pp. 18–19, Figure 11 · p. 19

D.1 Data Licenses

• EnzyBind is released under CC BY 4.0, while users must also comply with the licensing terms of PDB and PDBbind, from which the dataset is derived.Source: §D.1 · pp. 19–20

D.2 Complex Preprocessing

• The train-test split clusters enzyme sequences with CD-HIT and assigns clusters randomly so training and test enzymes are disjoint.Source: §D.2 · p. 20
• Preprocessing filters and standardizes 11,100 enzyme-substrate complexes; the model cannot process multi-chain or symmetric enzymes, so only substrate atoms within 10 Å of an enzyme atom are retained.Source: §D.2 · p. 20

D.3 Multiple Sequence Alignment

• Functional sites are identified by aligning sequences from the same enzyme family with MAFFT and selecting conserved residues using an identity threshold of τ = 0.3.Source: §D.3 · p. 20
• The identified functional sites are encoded as a binary vector matching the original sequence length, with 1 marking a functional-site position and 0 other positions.Source: §D.3 · p. 20

E.1 Backbone Frame Representation

• Each residue is represented by a local frame constructed from idealized N, Cα, C, and O backbone coordinates using the rigid3Point procedure.The procedure follows AlphaFold2 and assumes standard chemical bond lengths and angles.Source: §E.1 · p. 21

E.2 Model Architecture of Projector

• The projector uses two linear layers and layer normalization to decompose substrate embeddings and output aligned substrate features.Its architecture is shown in Figure 15.Source: §E.2 · p. 21, Figure 15 · p. 21

E.3 Edge and Backbone Update

• Node embeddings are linearly projected to half their dimension before edge-level reasoning, and pairwise edge features are updated through an MLP followed by layer normalization.The edge-update input combines projected source and target node embeddings with the current edge feature.Source: §E.3 · p. 22
• A learned rigid-body transformation updates each residue frame by combining a quaternion-derived rotation with a translation, then composing the transform with the current frame.The update progressively refines backbone geometry across Evoformer layers.Source: §E.3 · p. 22
• Known structural motifs remain fixed anchors during denoising, while surrounding regions are generated using a binary mask to distinguish clamped and generated structure.The motif region is overwritten with its known value at each denoising step.Source: §E.3 · p. 23

F.1 Additional Training Details

• Training uses LoRA with rank 16 and scaling factor 32 across selected attention and embedding projections; node and edge embeddings have dimensions 256 and 128.The model supports up to 2000 residues and embeds 1000 timesteps using sinusoidal and learned positional encodings.Source: §F.1 · p. 23
• The IPA module has six blocks with eight-head attention, alongside a two-layer sequence Transformer with four heads per layer.These configurations were selected empirically to balance computational efficiency and modeling capacity.Source: §F.1 · p. 23

F.2 Evaluation Details

• Self-consistency is measured by comparing generated backbones with ESMFold-predicted all-atom structures using TM-score and aligned RMSD.TM-score uses tmtools, and RMSD follows the FrameFlow alignment procedure.Source: §F.2 · p. 23
• EC match rate is the proportion of designs whose predicted EC number matches the native enzyme, while catalytic efficiency is reported as mean predicted kcat.CLEAN predicts EC numbers from sequence; the kcat model takes sequence and substrate SMILES as input.Source: §F.2 · p. 23
• Substrate binding is assessed by GNINA docking over the whole enzyme structure and by mean ESP scores from a model using sequence and substrate inputs.The cited metric description identifies CLEAN, UniKP, EnzyGen ESP, and Gnina as prediction or docking tools with experimental validation examples or complementary validation roles.Source: §F.2 · pp. 23–24
• Table 11 is identified as presenting model configuration and hyperparameter settings.Source: Table 11 · p. 24
• Diversity is the fraction of generated designs assigned to unique structural clusters, while novelty is the average maximum TM-score to native proteins, with lower scores indicating greater novelty.Both measures use Foldseek.Source: §F.2 · pp. 23–24