EnzyAdapter: Substrate-Aware Enzyme Design

• EnzyAdapter is a substrate-conditioning module that integrates explicit substrate chemistry into pretrained motif-scaffolding models for enzyme backbone design.
• It employs layerwise cross-attention between enzyme node features and substrate graph embeddings, preserving key catalytic motifs while customizing the scaffold.
• A two-stage, parameter-efficient training method enables significant improvements in designability, catalytic efficiency, and binding affinity.

EnzyAdapter is the substrate-conditioning module at the core of EnzyControl, a framework for enzyme backbone generation that makes a pretrained motif-scaffolding model aware of both conserved functional motifs and small-molecule substrates. In the formulation used by EnzyControl, the target is to generate a scaffold $\mathbf{S}$ conditioned on functional sites $\mathbf{M}$ and a substrate graph $\mathcal{G}$, i.e. $p(\mathbf{S}\mid \mathbf{M}, \mathcal{G})$. EnzyAdapter is introduced as a lightweight, modular bridge that augments FrameFlow, a pretrained SE(3) flow-matching motif-scaffolding model, with an explicit substrate-conditioning pathway rather than rewriting the full generator (Song et al., 29 Oct 2025).

1. Origin and problem setting

EnzyAdapter emerges from a specific limitation in computational enzyme design: prior motif-scaffolding and enzyme-generation systems either ignore substrate information during generation or use substrate information only as a post hoc filter. In that regime, backbone generation can preserve catalytic motifs or produce structurally plausible scaffolds, but it cannot directly tailor the generated backbone to a specified substrate. EnzyControl addresses this limitation by conditioning generation jointly on MSA-annotated catalytic sites and substrate identity, and EnzyAdapter is the module that operationalizes the substrate side of that conditioning (Song et al., 29 Oct 2025).

The broader system is built around EnzyBind, a dataset with 11,100 experimentally validated enzyme-substrate pairs curated from PDBbind and annotated with MSA-derived functional sites. Within this setting, functional sites are identified by MSA over enzyme families, using conserved residues as motif anchors, and the scaffold is then generated around those fixed sites. EnzyAdapter conditions that generation on substrate chemistry, so the resulting model is not merely motif-preserving but substrate-sensitive. This suggests that the module is best understood not as a generic fine-tuning add-on, but as the substrate-aware component that closes the gap between motif scaffolding and substrate-specific enzyme design (Song et al., 29 Oct 2025).

2. Conceptual role inside EnzyControl

The architectural role of EnzyAdapter is sharply delimited. EnzyBind supplies experimentally grounded enzyme-substrate supervision; MSA-derived motifs preserve catalytic residues; EnzyAdapter injects substrate awareness; and the two-stage training scheme aligns the new conditioning signal with the pretrained structural prior. The paper presents this division of labor explicitly, with EnzyAdapter functioning as the dedicated pathway through which substrate chemistry influences backbone generation (Song et al., 29 Oct 2025).

A central misconception is to treat EnzyAdapter as equivalent to naïve multimodal fusion. The paper reports that a simple strategy such as concatenating substrate features with protein node features before passing them into the backbone model was not sufficiently effective. EnzyAdapter was therefore introduced as a more principled alternative: it leaves the strong pretrained structural prior of FrameFlow intact while adding a dedicated conditioning mechanism. A second misconception is to interpret the module as pose-conditioned ligand modeling. The substrate is not represented as a 3D ligand conformer, because the binding pose is generally unknown at generation time; instead, EnzyControl uses a chemical graph $\mathcal{G}$ as the substrate representation (Song et al., 29 Oct 2025).

3. Architecture and information flow

EnzyAdapter has two main components: substrate feature initialization and an adapter block inserted into each layer of the pretrained motif-scaffolding network. The initialization stage begins from a substrate chemical graph $\mathcal{G}$, encoded by a pretrained Uni-Mol encoder. The paper states that Uni-Mol was trained on 209 million molecular conformers and is used here as a general-purpose chemical representation learner. Its output is mapped into the feature space of the enzyme generator by a small trainable projector composed of two linear layers plus Layer Normalization, yielding

$$\mathbf{S}_0 = \text{Projector}(\text{UniMol}(\mathcal{G})), \quad \mathbf{S}_0 \in \mathbb{R}^{D_s}.$$

The Uni-Mol encoder is frozen during training, which the paper describes as a way to preserve pretrained chemistry knowledge and avoid overfitting on the relatively small enzyme-substrate dataset (Song et al., 29 Oct 2025).

The second component is the EnzyAdapter block proper. Rather than concatenating substrate and protein features, the block uses cross-attention. At layer $k$, the current protein node features $\mathbf{h}_k$ serve as queries, while the substrate embedding $\mathbf{S}_0$ provides keys and values:

$\mathbf{M}$0

with

$\mathbf{M}$1

The resulting cross-attention output is passed through a linear layer and concatenated with the IPA output before being reintegrated into the backbone generator. The paper describes this design as model-agnostic and modular, with the implication that the same conditioning mechanism could in principle be attached to other backbone-generation architectures as well (Song et al., 29 Oct 2025).

Functionally, this information flow lets the generator preserve catalytic residues while shaping the surrounding scaffold so that it is better suited to the intended substrate. The technical significance of the design is that substrate information is injected repeatedly, layer by layer, rather than being compressed into a single early fusion step (Song et al., 29 Oct 2025).

4. Training paradigm and parameter-efficient adaptation

EnzyControl trains EnzyAdapter with a two-stage procedure. In the first stage, only the projector and EnzyAdapter are trained, while the rest of the prediction network remains frozen. The stated purpose is to align substrate features with enzyme structure without disturbing the pretrained motif-scaffolding behavior. This stage therefore isolates the substrate-conditioning problem from broader structural adaptation (Song et al., 29 Oct 2025).

In the second stage, the full prediction network and enzyme embedder are fine-tuned using LoRA with rank $\mathbf{M}$2 and scaling factor $\mathbf{M}$3, while the projector and EnzyAdapter continue to be updated under the generation loss. The paper emphasizes that this staged scheme balances preservation of the pretrained structural prior against learning the new substrate-conditioning task. A plausible implication is that EnzyAdapter is not simply a fixed adapter appended to a frozen model; rather, it serves as the anchor around which increasingly global adaptation is introduced in a controlled manner (Song et al., 29 Oct 2025).

This training strategy is central to the module’s identity. The first stage establishes a stable alignment between substrate embeddings and the protein generator’s hidden space. The second stage then allows joint refinement of the overall model without abandoning parameter-efficient adaptation. In the paper’s presentation, EnzyAdapter is therefore inseparable from the staged optimization protocol that makes substrate-aware conditioning stable in practice (Song et al., 29 Oct 2025).

5. Empirical performance and ablation evidence

The value of EnzyAdapter is primarily established through EnzyControl’s benchmark results and ablations. On EnzyBind, EnzyControl achieves the best overall performance, with designability $\mathbf{M}$4, corresponding to a 13% relative improvement over the second-best model. The paper also reports a 13% improvement in catalytic efficiency ($\mathbf{M}$5) and a 10% improvement in EC match rate over the strongest baselines. It further reports improved binding affinity on both EnzyBind and EnzyBench (Song et al., 29 Oct 2025).

The paper provides more specific comparative observations. Relative to EnzyGen on EnzyBench, EnzyControl achieves the best reported binding affinity and pLDDT, with around a 3% binding-affinity improvement over the second-best model on EnzyBind and a 1.5% improvement in the EnzyBench average binding-affinity comparison, while maintaining competitive ESP scores. Because EnzyAdapter is the component that introduces substrate awareness into the pretrained generator, these aggregate gains are attributed in substantial part to its conditioning mechanism (Song et al., 29 Oct 2025).

The ablation study makes the module’s contribution more explicit. Removing EnzyAdapter consistently degrades performance, with the largest drop in EC match rate, and also lowers designability, catalytic efficiency, binding affinity, and ESP score. The paper interprets this pattern as evidence that EnzyAdapter is critical for producing enzymes that align with the intended functional profile rather than merely producing plausible backbones. The model also shows strong sensitivity to motif annotation quality, supporting the broader claim that accurate MSA-derived functional sites and substrate-aware conditioning are jointly essential. This suggests that EnzyAdapter should not be viewed as an isolated substitute for functional annotation; its strongest effect appears when coupled to reliable catalytic-site constraints (Song et al., 29 Oct 2025).

6. Relation to other adapter paradigms and terminological boundaries

The term “adapter” is used widely across machine learning, but EnzyAdapter belongs to a distinct class of modules. In information retrieval, adapter methods such as Houlsby or Pfeiffer adapters are bottleneck layers inserted into frozen transformers for parameter-efficient transfer; for example, adapter-tuned SPLADE optimizes only about 2% of the training parameters while matching or surpassing full fine-tuning in sparse retrieval (Pal et al., 2023), and cross-language dense retrieval studies report updating roughly 0.5% to 10% of parameters depending on configuration (Yang et al., 2022). EnzyAdapter differs in mechanism and objective: it is a cross-attentive substrate-conditioning pathway inside a motif-scaffolding backbone generator rather than a generic bottleneck fine-tuning layer (Song et al., 29 Oct 2025).

The name should also be distinguished from several enzyme-related adapter systems in adjacent literatures. ProCALM uses conditional adapters in ProGen2 for function- and taxonomy-conditioned protein sequence generation (Yang et al., 2024). ERBA, the Enzyme-Reaction Bridging Adapter, is a staged multimodal adapter for kinetic-parameter prediction, combining Molecular Recognition Cross-Attention, Geometry-aware Mixture-of-Experts, and Enzyme-Substrate Distribution Alignment (Wang et al., 13 Mar 2026). In algorithmic differentiation, EnzymeExternalFunctionHelper is a CoDiPack integration layer for Enzyme-generated derivative routines rather than a protein-design module (Sagebaum et al., 2023). These systems share the adapter label only at a high level of modularity; they address different tasks, operate on different modalities, and implement different integration strategies.

Within this landscape, EnzyAdapter is specifically the module that converts a pretrained motif-scaffolding backbone generator into a substrate-aware enzyme backbone generator. Its distinguishing features are the use of frozen Uni-Mol features for graph-based substrate representation, layerwise cross-attention from protein nodes to substrate embeddings, and a two-stage training scheme that first aligns the conditioning pathway and then refines the broader model with LoRA. In that restricted but technically precise sense, EnzyAdapter is the substrate-conditioning mechanism that defines EnzyControl’s attempt to generate enzyme backbones that are simultaneously structurally designable, catalytically meaningful, and substrate-compatible (Song et al., 29 Oct 2025).