EnzyControl

EnzyControl is a generative method for designing enzyme backbones that are conditioned on both a catalytic site and the specific substrate the enzyme is meant to act on. Most protein backbone generators produce structures that are designable but functionally agnostic: they can scaffold a motif without any notion of which small molecule the resulting protein should bind or catalyze. For enzymes, this is a critical gap, because catalytic function depends on the precise geometry of active-site residues around a substrate.

Rather than training a new generator from scratch, EnzyControl attaches a lightweight, modular component called EnzyAdapter to FrameFlow, a pretrained SE(3) flow-matching model for motif-scaffolded backbone generation. The adapter injects substrate and catalytic-site information into the frozen generator, adding functional and substrate-specific control while preserving the base model's learned ability to produce high-quality, designable backbones. This adapter-on-a-foundation-model strategy keeps the added parameter and data cost small.

EnzyControl was introduced in October 2025 by Chao Song, Zhiyuan Liu, Han Huang, Liang Wang, Qiong Wang, Jianyu Shi, Hui Yu, Yihang Zhou, and Yang Zhang, a collaboration spanning Northwestern Polytechnical University, the National University of Singapore, The Chinese University of Hong Kong, and the Institute of Automation at the Chinese Academy of Sciences. It is described in an arXiv preprint with code released under the MIT license.

#Key Features

• Substrate-conditioned generation: EnzyControl conditions backbone generation on a specific substrate and catalytic site, producing enzymes tailored to a target reaction rather than functionally agnostic scaffolds.
• Adapter on a frozen foundation model: EnzyAdapter is a lightweight module that plugs into the pretrained FrameFlow generator through cross-attention layers and a small projector, adding control while preserving the base model's designability.
• Two-stage training: The projector and adapter are first trained with the base network frozen to align substrate features, then the full network is fine-tuned with LoRA under the generation loss for consistent optimization.
• EnzyBind dataset: A curated benchmark of 11,100 experimentally validated enzyme-substrate pairs from PDBbind, spanning six catalytic types with MSA-identified conserved functional sites.
• Improved catalytic metrics: Across the authors' evaluations, EnzyControl improves designability and functional quality over baselines, including a reported 15.4% gain in predicted catalytic efficiency.

#Technical Details

EnzyControl builds on FrameFlow, an SE(3) flow-matching model that generates protein backbones by scaffolding a fixed structural motif. EnzyAdapter augments this generator with cross-attention layers and a two-linear-layer projector with layer normalization that fuse substrate representations and catalytic-site context into the backbone-generation trajectory. Training proceeds in two stages: stage one updates only the projector and adapter while the base network stays frozen, and stage two fine-tunes the full model with Low-Rank Adaptation (LoRA). The model is trained and evaluated on EnzyBind, comprising 11,100 enzyme-substrate pairs curated from PDBbind. On the EnzyBench evaluation, EnzyControl reaches a designability of 0.716 (a 13% relative improvement over the next-best method) with 88.5% of designs exceeding scTM 0.5, an EC-number match rate of 0.504, and a 15.4% improvement in predicted catalytic efficiency, while also improving predicted substrate binding affinity.

#Applications

EnzyControl targets computational enzyme designers who need backbones built around a defined active site and a particular substrate, for example when engineering catalysts for a chosen reaction or repurposing a known catalytic motif toward a new substrate. Because it operates as an adapter on an existing backbone generator, it fits into workflows already using flow- or diffusion-based scaffolding, adding substrate awareness without retraining the underlying model. The released code, checkpoints, and EnzyBind dataset let groups reproduce the pipeline and extend it to their own catalytic targets.

#Impact

EnzyControl illustrates how functional, substrate-level control can be retrofitted onto pretrained structure generators through a small trainable adapter, avoiding the cost of training a bespoke enzyme model. Its EnzyBind dataset and EnzyBench benchmark also give the field a substrate-paired resource for measuring functional, not just structural, quality of generated enzymes. As a recent preprint, its improvements are in-silico and await peer review and experimental validation; model checkpoints are distributed separately via Google Drive rather than bundled with the repository.