EnzyBind: Enzyme Backbone Design Dataset

• EnzyBind is a curated dataset of 11,100 enzyme–substrate pairs that supports substrate-specific enzyme backbone generation while preserving catalytic motifs.
• It integrates precise structural data, MSA-derived functional-site annotations, and substrate chemical graphs to guide conditional enzyme design.
• Benchmark metrics like scTM, EC Match Rate, and designability demonstrate its effectiveness as a standardized evaluation tool in enzyme design.

EnzyBind is a curated dataset and benchmark for substrate-specific enzyme backbone generation introduced in “EnzyControl: Adding Functional and Substrate-Specific Control for Enzyme Backbone Generation” (Song et al., 29 Oct 2025). It is derived from PDBbind and contains 11,100 experimentally validated enzyme–substrate pairs, curated to support conditional enzyme design in which generation is conditioned jointly on a target substrate and on functionally important catalytic motifs. In the formulation used by EnzyControl, EnzyBind supplies the ingredients for learning the conditional distribution $p(\mathbf{S}\mid \mathbf{M}, \mathcal{G})$, where $\mathbf{M}$ is the set of annotated functional-site residues, $\mathcal{G}$ is the substrate chemical graph, and $\mathbf{S}$ is the scaffold to be generated (Song et al., 29 Oct 2025). The dataset is therefore not merely a collection of protein–ligand complexes; it is a substrate-aware enzyme-design resource with structure, ligand context, EC-family information, and MSA-derived functional-site annotations.

1. Definition and scientific scope

EnzyBind was introduced to address a specific gap in enzyme design data: existing resources were not suitable for learning or evaluating enzyme backbone generation with explicit substrate specificity and catalytic-site preservation (Song et al., 29 Oct 2025). The motivating claim is that enzyme design differs from generic protein design because enzymes must preserve functional motifs, bind specific small molecules, and support the reaction class reflected by their EC number. A useful dataset in this setting therefore requires, for each example, a protein structure, a bound small-molecule substrate, spatial binding context, and functional-site annotations.

The dataset is positioned against several prior resource types. EnzymeMap and ReactZyme provide protein sequences and SMILES but lack precise pocket information; they are described as mainly used for EC prediction rather than enzyme backbone generation. EnzymeFill includes precise pocket structures and substrate conformations, but it is characterized as a synthetic dataset whose reliability is limited by the absence of wet-lab experimental validation. EnzyBind is meant to fill this gap by providing experimentally validated enzyme–substrate complexes with precise pocket structures and substrate conformations, curated for catalytic backbone generation (Song et al., 29 Oct 2025).

This scope also clarifies what EnzyBind is not. It is not a generic affinity benchmark organized around measured $K_d$ or $K_i$, nor is it a reaction-only dataset like KEGG-derived metabolite transformation graphs. In particular, the enzymatic link prediction framework ELP predicts molecule-to-molecule transformation links in a KEGG biochemical network rather than enzyme–substrate binding events, and its graph is explicitly not bipartite over enzymes and compounds (Jiang et al., 2020). EnzyBind instead centers on experimentally grounded enzyme–substrate complexes.

2. Construction from PDBbind and curation workflow

EnzyBind is derived from PDBbind, which is used as the upstream structural source because it is a curated database of protein–ligand complexes derived from the Protein Data Bank and therefore provides bound ligand structures in complex with proteins together with experimentally determined 3D structures (Song et al., 29 Oct 2025). The resulting EnzyBind corpus contains 11,100 experimentally validated enzyme–substrate pairs, with the appendix giving the exact total as 11,100.

The curation pipeline described in the paper is selective but not fully exhaustive. The first explicit filter is that complexes that could not be processed using RDKit were excluded. The remaining PDB files were then cleaned following a standardized procedure. The preprocessing steps stated in the appendix are: molecule standardization using Open Babel, following the preprocessing pipeline from EquiBind; correct hydrogen placements on enzymes; and add missing hydrogens using the reduce tool (Song et al., 29 Oct 2025).

A model-driven structural filter is applied because the authors’ model cannot process multi-chain enzymes or symmetric complexes containing repeated enzyme units. To address this, they retain only the substrate atoms within $10 \, \text{\AA}$ of any enzyme atom. The paper states that this ensures each sample represents a physically relevant interaction while excluding redundant or ambiguous structural data. A plausible implication is that EnzyBind is curated toward tractable single-enzyme/substrate interaction contexts rather than full multimeric assemblies.

The split protocol is based on sequence disjointness rather than chronology. Enzyme sequences are clustered using CD-HIT, clusters are assigned randomly to training or test, enzyme–substrate pairs are then sampled accordingly, and the goal is to ensure that enzymes in train and test are disjoint (Song et al., 29 Oct 2025). Exact split sizes are not reported in the extracted text. The paper also explicitly does not report several criteria that would matter for exact reconstruction of the raw-to-final pipeline, including a crystallographic resolution cutoff, an explicit experimental-method filter, explicit exclusion rules for cofactors, inhibitors, transition-state analogs, or metal ions, a detailed ligand-disambiguation protocol, residue or atom completeness thresholds, protein length cutoffs, ligand size cutoffs, or train/validation/test counts (Song et al., 29 Oct 2025).

3. Annotations, representations, and stored information

A defining feature of EnzyBind is that each example is enriched with annotations needed for substrate-aware motif scaffolding. Protein structure is represented as a sequence of residue frames

$$\mathbf{T}=[T^{(1)},\dots,T^{(N)}], \quad T=(\mathbf{r},\mathbf{x})\in \mathrm{SE}(3),$$

where $\mathbf{r}\in \mathrm{SO}(3)$ is a residue-local rotation and $\mathbf{x}\in \mathbb{R}^3$ is a translation (Song et al., 29 Oct 2025). The appendix states that each residue frame is constructed from backbone atoms $\mathbf{M}$0 using AlphaFold2’s rigid3Point-style frame construction: $\mathbf{M}$1

The main added annotation is the functional-site motif, derived through multiple sequence alignment (MSA). Enzymes are grouped by the same second-level EC number, aligned using MAFFT, and conserved residues are identified using the identity threshold

$\mathbf{M}$2

Residues “that appear consistently across all aligned sequences” are treated as functional sites, following the EnzyGen approach, and are encoded as a binary vector of sequence length, with 1 indicating a functional site and 0 otherwise (Song et al., 29 Oct 2025). In the model notation, these residues form the motif set

$\mathbf{M}$3

Substrates are represented primarily as chemical graphs rather than as fixed bound conformers during generation. The substrate encoder pipeline is

$\mathbf{M}$4

where $\mathbf{M}$5 is the substrate chemical graph, $\mathbf{M}$6 is a frozen pretrained molecular encoder, and the $\mathbf{M}$7 is a trainable module of two linear layers plus LayerNorm (Song et al., 29 Oct 2025). For $\mathbf{M}$8 prediction in evaluation, the paper also states that the substrate’s SMILES representation is used as input to UniKP.

Each EnzyBind entry therefore effectively includes enzyme structure, enzyme sequence, substrate structure, EC label sufficient to assign at least a second-level family, MSA-derived functional-site annotations, and a motif mask or binary vector (Song et al., 29 Oct 2025). This is more richly annotated for conditional enzyme design than sequence-only enzyme resources or reaction-only datasets such as Boost-RS, which models a sparse binary enzyme–compound matrix derived from KEGG and uses EC, KO, fingerprints, and compound–compound relationships as auxiliary information but does not include structural complex geometry (Li et al., 2021).

4. Dataset scale, EC coverage, and benchmark role

The exact dataset size reported in the appendix is 11,100 enzyme–substrate complexes (Song et al., 29 Oct 2025). The dataset is said to cover six fundamental catalytic types, and the paper reports family-level analyses over second-level EC families including 1.1, 1.6, 1.14, 2.1, 2.3, 2.5, 2.7, 3.1, 3.2, 3.4, 3.5, 3.6, 4.1, 4.2, 5.6, 5.99, and 6.2. Exact counts per EC class are not provided.

EnzyBind is used both as a training corpus and as a benchmark dataset. The paper states explicitly that FrameFlow is used as the pretrained foundation model and is fine-tuned on EnzyBind, and that baselines, including EnzyGen, are evaluated or retrained on EnzyBind for fair comparison (Song et al., 29 Oct 2025). Thus EnzyBind functions simultaneously as a curated data source, a supervision source for motif and substrate conditioning, and a standardized evaluation substrate for controlled enzyme backbone generation.

This role distinguishes EnzyBind from EnzyBench, which is an external benchmark from the EnzyGen study and is not described as a subset of EnzyBind (Song et al., 29 Oct 2025). EnzyBench, by contrast, is a cross-family benchmark spanning 3,157 fourth-level enzyme families and is designed around substrate-aware generative enzyme design with evaluation via ESP score, Gnina docking affinity, and pLDDT (Song et al., 2024). A plausible implication is that EnzyBind is more tightly curated around experimentally validated enzyme–substrate complexes from PDBbind, whereas EnzyBench is broader in family coverage and benchmark orientation.

The formal task on EnzyBind is conditional enzyme backbone generation or motif scaffolding with substrate conditioning. Given functional sites $\mathbf{M}$9 and substrate $\mathcal{G}$0, the model must generate a scaffold $\mathcal{G}$1 such that the full backbone $\mathcal{G}$2 is compatible with both motif and substrate. The target vector field is conditioned on motif and substrate: $\mathcal{G}$3 The paper gives the flow-matching loss over translation and rotation components as

$\mathcal{G}$4

5. Evaluation protocol and benchmark metrics

The EnzyBind evaluation pipeline is standardized across methods. For each input, a method generates 20 backbone structures; ProteinMPNN designs 5 sequences per backbone; ESMFold predicts all-atom structures; and all reported metrics are computed on the ESMFold-predicted structures (Song et al., 29 Oct 2025). This makes EnzyBind a benchmark for comparing enzyme-generation pipelines under a common downstream structure-recovery and assessment protocol.

Structural and self-consistency metrics include scTM, defined as TM-score between generated backbone and ESMFold-predicted all-atom structure; scRMSD, defined as aligned $\mathcal{G}$5-RMSD; Designability, defined as the fraction of generated backbones with

$\mathcal{G}$6

and also the fraction with

$\mathcal{G}$7

Functional metrics are EC Match Rate, using CLEAN to test whether the predicted EC number matches the native enzyme’s EC number; $\mathcal{G}$8, defined as the average predicted catalytic rate constant from UniKP using sequence and substrate; Binding Affinity, defined as the docking score from Gnina, where lower is better; and ESP Score, defined as the enzyme–substrate interaction score from the EnzyGen/ESP model, where higher is better (Song et al., 29 Oct 2025). Additional metrics include AAR, RMSD relative to native structures, Diversity, and Novelty; the appendix defines Novelty as the average of the maximum TM-scores between each generated enzyme and all native proteins, with lower values corresponding to more novel designs.

A composite metric, Success Rate, counts a design as successful only if all four of the following hold: EC match, $\mathcal{G}$9, $\mathbf{S}$0, and better binding affinity than the native counterpart (Song et al., 29 Oct 2025). This composite criterion is notable because it combines structural plausibility, predicted function, and substrate compatibility.

The same paper reports EnzyControl’s main EnzyBind benchmark results: scTM > 0.5 = 0.8848, Designability = 0.7160, EC Match Rate = 0.5041, $\mathbf{S}$1, Binding Affinity = $\mathbf{S}$2, ESP Score = 0.7334, and Success Rate = 0.1195 (Song et al., 29 Oct 2025). These numbers define the current performance reference within that benchmark. The paper further states that EnzyControl achieves 13% improvement in designability and 13% in catalytic efficiency compared to baseline models in the abstract, while the main text reports a 23% relative improvement in scTM > 0.5 over the second-best model, a 10% relative improvement in EC Match Rate over the second-best model, and a 23% improvement in Success Rate over the second-best model (Song et al., 29 Oct 2025).

6. Empirical significance, ablations, and limitations

Ablation studies are used to test whether the two EnzyBind-specific signals—MSA motif annotations and substrate conditioning—materially contribute to performance. When EnzyAdapter is removed but MSA motifs are retained, scTM > 0.5 drops from 0.8848 to 0.8748, Designability from 0.7160 to 0.7067, EC Match Rate from 0.5041 to 0.4761, $\mathbf{S}$3 from 2.9168 to 2.5833, Binding Affinity worsens from $\mathbf{S}$4 to $\mathbf{S}$5, and ESP from 0.7334 to 0.7205 (Song et al., 29 Oct 2025). When MSA conditioning is removed but EnzyAdapter remains, the scores become 0.8719, 0.6863, 0.4764, 2.4615, $\mathbf{S}$6, and 0.7183, respectively. Removing both yields 0.8684, 0.6784, 0.4627, 2.4492, $\mathbf{S}$7, and 0.7168. These ablations indicate that both substrate input and MSA-derived motif annotations are functional components of the benchmarked task, rather than optional metadata.

The paper also perturbs motif residues to test annotation sensitivity. At 0% perturbation, designability is 0.7160, EC match is 0.5041, $\mathbf{S}$8, and binding affinity is $\mathbf{S}$9; at 50% perturbation, the corresponding values are 0.7023, 0.4918, 2.6540, and $K_d$0; and at 100% perturbation, they are 0.6863, 0.4764, 2.4615, and $K_d$1 (Song et al., 29 Oct 2025). The authors interpret this as evidence that performance is sensitive to motif fidelity and that the MSA-based annotation strategy is essential within the benchmark.

The dataset’s limitations are specific and consequential. First, the work focuses on generating enzyme backbones only, “without modeling the specific conformations these backbones adopt when binding to substrates” (Song et al., 29 Oct 2025). Thus, although EnzyBind contains bound complexes, the benchmarked model uses substrate embeddings rather than explicit bound 3D ligand placement during generation. Second, functional sites are inferred from MSA conservation at the second-level EC classification, not from curated catalytic-residue databases; the motif labels are therefore useful but approximate. Third, because the framework cannot process multichain or symmetric complexes directly, preprocessing retains only substrate atoms within $K_d$2 of enzyme atoms and effectively simplifies interaction contexts toward single-chain enzyme scaffolds. Fourth, the benchmark emphasizes functional specificity and designability over unconstrained structural diversity; the paper notes that EnzyControl lags behind RFDiffusion and Chroma in diversity and novelty on EnzyBind (Song et al., 29 Oct 2025).

These caveats are important when situating EnzyBind among related enzyme-design resources. EnzyPGM addresses substrate-specific enzyme design with a stronger emphasis on explicit pocket–substrate interaction modeling through Residue-atom Bi-scale Attention and reports an average 0.47 kcal/mol lower Vina score than EnzyGen on its EnzyPock benchmark (Lin et al., 27 Jan 2026). GENzyme instead adopts a reaction-conditioned design paradigm in which a catalytic reaction is the primary conditioning signal, and it generates a catalytic pocket, a full enzyme, and a putative enzyme–substrate complex, but its binding stage is a post hoc docking-and-screening module rather than an end-to-end learned catalytic complex model (Hua et al., 2024). By contrast, EnzyBind is narrower and more explicit in its role: it is a curated dataset of experimentally validated enzyme–substrate complexes designed to support and evaluate substrate-aware motif-scaffolding systems.

EnzyBind is publicly released on Zenodo at

$K_d$3

under CC BY 4.0, with the additional note that users must also comply with the licensing terms of PDB/PDBbind because the dataset is derived from those sources (Song et al., 29 Oct 2025). This public release, together with the accompanying code release for EnzyControl, makes EnzyBind not only a benchmark concept but an operational resource for computational enzyme design research.