Use ZymCTRL Online
Commercially Available ZymCTRL No-Code Web Server
ZymCTRL: Conditional Language Model for Controllable Generation of Artificial Enzyme Sequences
ZymCTRL is a conditional protein language model that generates enzyme sequences from an enzyme class (EC) prompt. It was developed by Geraldene Munsamy, Sebastian Lindner, Philipp Lorenz and Noelia Ferruz, with affiliations at Basecamp Research, Friedrich-Alexander-Universität Erlangen, the University of Girona and the University of Bayreuth, and was presented at the Machine Learning for Structural Biology Workshop at NeurIPS 2022. The model has been released publicly on Hugging Face.
Enzymes drive the chemistry of life and industry, but designing new ones for a chosen reaction is hard. Searching sequence space blindly is inefficient, and most approaches cannot be steered toward a specific catalytic function.
ZymCTRL addresses this by learning the relationship between sequence and enzyme class. Each training sequence is preceded by its EC tag, so at generation time a user can request sequences for a desired reaction type.
Property | Detail |
|---|---|
Architecture | Transformer decoder (CTRL/GPT-2 style), 36 layers, about 738M parameters |
Conditioning | EC control tag, tokenized by level |
Training data | BRENDA enzyme sequences via UniProt (36,276,604 after filtering) |
Context length | 1024 residues |
Training compute | About 15,000 GPU hours on 48 NVIDIA A100 80GB GPUs |
Why Controllable Enzyme Generation Matters
Prompting by EC number turns enzyme design into a targeted request: ask for a class, receive candidate sequences. Because EC labels are tokenized one level at a time, related subclasses share tokens, which helps the model relate similar reactions.
How ZymCTRL Generates Enzyme Sequences
Control tag prompt: The EC label is placed at the start of the sequence as a conditioning prefix.
Autoregressive decoding: The model was trained with a next-token negative log-likelihood objective and generates residues one at a time.
Sampling settings: The authors sample with top_k 9, temperature 1, top_p 1 and a repetition penalty of 1.2, chosen to match the amino acid distribution of natural sequences.
ZymCTRL Evaluation Results
The authors compared 11,439 generated and 11,439 natural sequences with several independent predictors.
Metric | Generated | Natural |
|---|---|---|
Predicted globular (IUPred3) | 97.7% | 99.3% |
OmegaFold LDDT | 60.01 | 84.78 |
Top-level EC accuracy (ProteInfer) | 80.9% | 81.2% |
Full four-level EC accuracy | 54.0% | 62.0% |
Generated sequences had an average identity of 53.1% to the training set, and about 12.5% exceeded 90% identity to a training sequence.
What is Tamarind Bio?
Tamarind Bio is a no-code bioinformatics platform built to give life scientists and researchers access to powerful computational tools. Many cutting-edge machine learning models are hard to deploy and use. Tamarind provides an intuitive, web-based environment that removes the complexity of high-performance computing, software dependencies and command-line interfaces.
The platform is designed for biologists, chemists and other researchers who may not have a background in programming or cloud infrastructure but want to run models on their own data. Key features include:
A user-friendly graphical interface for setting up and launching experiments
A robust API for integration into existing research pipelines
An automated system for managing and scaling computational resources
Tamarind treats information and data security as a top priority, as detailed in its Trust Center and Terms of Service.
Accelerating Discovery with ZymCTRL on Tamarind Bio
Class-targeted enzyme design: Generate candidate sequences for a reaction type specified by its EC number.
Diversifying enzyme families: Produce sequences that are distant from natural ones, as the authors report on average.
Rare class exploration: The authors suggest fine-tuning for underrepresented classes, where model confidence is lower.
Candidate triage: Filter outputs with structure and function predictors, the same way the authors evaluated them.
How to Use ZymCTRL on Tamarind Bio
Log in and open the tool: Sign in at tamarind.bio and open ZymCTRL.
Choose an EC class: Enter the enzyme class label (for example a four-level EC number) that defines the reaction you want.
Set generation parameters: If exposed, set the number of sequences and sampling options such as top_k, temperature and repetition penalty.
Run the job: Submit and wait for generation to finish.
Download outputs: Retrieve the generated amino acid sequences.
Filter and interpret: Check novelty against known sequences and predicted structure confidence before choosing candidates.
Use downstream: Send selected sequences to structure prediction or experimental testing.
Things to Keep in Mind
The paper reports no experimental validation of catalytic activity; function is inferred from computational predictors only.
Predicted structure confidence is lower for generated sequences than for natural ones.
Confidence drops for underrepresented EC classes.
Sequences longer than 1024 residues were truncated in training, and generated lengths were matched to natural ones.
Function assessment relied on a single orthogonal predictor, ProteInfer, which has its own limitations.
Source: Munsamy G, Lindner S, Lorenz P, Ferruz N. ZymCTRL: a conditional language model for the controllable generation of artificial enzymes. Machine Learning for Structural Biology Workshop, NeurIPS 2022. Paper PDF. Model: huggingface.co/nferruz/ZymCTRL.