MetaEnzyme

A group trying to improve a PET hydrolase has a few hundred characterized enzymes to learn from, not the millions of sequences that make general protein models work. Enzyme design is chronically low-resource: functional categories fragment the data, each fold class holds only a handful of experimentally determined structures, and a network trained directly on one enzyme family leaves nothing for the next. MetaEnzyme's response is to stop training enzyme-specific models at all — pretrain one general protein design network on structure, freeze it, and reach each enzyme task through a few meta-learning updates or through no gradient updates whatsoever.

MetaEnzyme was built by Jiangbin Zheng, Han Zhang, Qianqing Xu, An-Ping Zeng and Stan Z. Li at Westlake University's AI Lab and its Center of Synthetic Biology and Integrated Bioengineering, with Zhejiang University, and was presented at ACM Multimedia 2024. Its foundation is UniProt-Net, a cross-modal structure-to-sequence network that reads a protein backbone and emits an amino acid sequence. Three enzyme redesign modes are then driven off that one checkpoint: FuncDesign, which predicts enzyme function and fold class; MutDesign, which ranks the effects of mutations; and SeqDesign, which generates full sequences for a supplied backbone.

The three modes differ in how much adaptation they need. SeqDesign and MutDesign run zero-shot off the fixed pretrained parameters — MutDesign is non-parametric, making no distributional assumption about the query enzyme at all — while FuncDesign uses MAML-style inner and outer loops over held-out fold classes, with a zero-shot arm reported alongside. The same Westlake group later pursued the complementary direction in CrossDesign, which distils a sequence model into a structure encoder instead of pretraining the structure network directly.

#Key Features

• One backbone, three task paradigms: Autoregressive generation, non-autoregressive regression and few-shot classification all draw on the same pretrained UniProt-Net weights instead of a separate architecture per enzyme task.
• Geometry-equivariant structure encoder: A lightweight GVP module produces rotation- and translation-invariant residue features, fused with a local reference frame so the whole encoder-adapter stack is invariant to rigid motion of the input coordinates.
• Decoupled mutation scoring: Wild-type and mutant distributions are computed independently, so one pass over the wild type plus a one-hot mutant encoding replaces per-mutant inference.
• Energy-motivated data augmentation: Backbones are perturbed with Riemann-Gaussian rather than plain Euclidean noise, keeping the augmented structures geometry-aware and consistent with the forces acting on a real protein.

#Technical Details

UniProt-Net has 124.82M parameters. Its geometry-invariant encoder is four GVP layers with scalar and vector node dimensions of 1024 and 256; the structure-sequence adapter and the self-attention sequence decoder each carry eight attention heads at an embedding dimension of 512. The encoder-adapter module is initialized from the ESM-IF inverse-folding model and then held fixed, while pretraining runs on CATH — 18,204 training, 608 validation and 1,120 test structure-sequence pairs — on a single A100 80GB GPU. On sequence recovery, MetaEnzyme reaches 54.94% amino acid recovery on the CATH test set and 62.68% on Ts50, ahead of ESM-IF, PiFold and ProteinMPNN; on a curated set of 178 plastic-degrading enzymes it recovers 64.35%, against 62.56% for ESM-IF. On ProteinGym's 217 deep mutational scanning assays it reaches a Spearman ρ of 43.9% with no finetuning, above VESPA at 43.7% and ESM-IF at 42.2%, and ahead of ESM-2 and other sequence-only baselines. Because scoring is decoupled, throughput on mutant libraries is thousands of times higher than per-mutant inference. Wet-lab assays of roughly 40 guided mutations on the P-protein decarboxylase of the reversible glycine cleavage system gave a Spearman ρ of 70.1% against measured effects.

#Applications

MetaEnzyme is aimed at enzyme engineering campaigns where the target family is too small to train on: prioritizing mutations before a round of cloning and assay, proposing sequences for a known catalytic backbone, and classifying the fold of an uncharacterized enzyme. The PET hydrolase and carbon-fixation case studies point at the settings that motivate it — plastic biodegradation and one-carbon metabolism — but nothing in the method is specific to either, and the decoupled scoring path matters most when a campaign must rank multi-site or higher-order mutants that per-mutant inference would make prohibitive.

#Impact

MetaEnzyme's contribution is organizational as much as architectural: it shows that function prediction, mutation-effect ranking and conditional sequence generation can all be served by one structure-pretrained protein network, and it assembles benchmark sets so those tasks can be compared at all. That framing is what the group carried forward into its later low-resource enzyme work. The practical constraint on adoption is that no code or checkpoints have been released — the repository named in the paper does not exist — so the results stand on the publication rather than on artifacts others can run. The authors list restricted functional-enzyme datasets and the limited scale of the wet-lab validation as open problems.