Use EnzyGen2 Online
Commercially Available EnzyGen2 No-Code Web Server
EnzyGen2: Co-Designing Sequence and Structure of Functional De Novo Enzymes
EnzyGen2 is a protein foundation model that jointly co-designs enzyme sequences and C-alpha backbone structures, conditioned on a ligand, functionally important residues and an NCBI taxonomic identifier. It was developed by Zhenqiao Song, Huichong Liu, Yunlong Zhao, Yang Yang and Lei Li at Carnegie Mellon University, UC Santa Barbara and MIT, and described in the 2026 bioRxiv preprint "Co-designing sequence and structure of functional de novo enzymes with EnzyGen2". No code repository or license was stated in the text we read.
Designing proteins that bind small-molecule ligands is hard, especially for enzymes that depend on substrate and cofactor binding. Most experimentally validated pipelines generate a backbone first and then design a sequence, which misses the coupling between the two, and most co-design models lack explicit ligand-binding constraints. Ligand-aware training data is also scarce, with only about 20,000 publicly accessible protein-ligand complexes.
Attribute | Detail |
|---|---|
Task | Joint sequence and backbone co-design of ligand-binding enzymes |
Conditioning | Ligand, functionally important residues, NCBI taxonomic identifier |
Size | About 730 million parameters |
Pretraining data | 720,993 proteins from PDB and Swiss-Prot, 9,002 taxonomic clusters |
Speed | About 400x faster than comparison pipelines (72x to 1,420x per pipeline) |
How EnzyGen2 Co-Designs Enzyme Sequences and Structures
Interleaved architecture: three blocks, each with eleven Transformer layers initialized from ESM2 650M and one k-nearest-neighbor equivariant GNN layer.
Ligand interaction module: three kNN-EGNN layers and a feed-forward head model protein-ligand binding.
Multi-task training: a weighted sum of masked sequence, masked structure and binding losses, pretrained in three stages.
Functional residues: defined automatically as positions present in over 30% of aligned sequences within a taxonomic group.
Family fine-tuning: optional fine-tuning on a family such as CAT, AadA or TPMT, using sequence and structure losses.
Candidate selection
Designs are sampled with nucleus sampling and filtered by ESP score above 0.6, an adversarial discriminator score above 0.6 and pLDDT above 80, then binned by identity to natural enzymes (30 to 80%), ranked by log-likelihood and reviewed by experts for active-site geometry.
EnzyGen2 Results: In Silico Benchmarks and Wet-Lab Validation
Evaluation | Result |
|---|---|
Vs. Inpainting (10 families) | Higher ESP in 7/10, RMSD under 2 A in 9/10 |
Vs. RFdiffusion3/LigandMPNN | Higher ESP in 8/10, RMSD under 2 A in 9/10 |
CAT (chloramphenicol acetyltransferase) | 7 of 20 designs conferred chloramphenicol survival |
AadA | 8 of 20 designs conferred aminoglycoside resistance |
TPMT | 6 of 10 expressed well; activity 2% to 73% of AtHMT |
Family-specific models reached average ESP above 0.6 and pLDDT above 90 with over 90% of designs under 2 A RMSD. CAT-17 grew at 500 ug/mL chloramphenicol where wild-type EcCAT did not, and AadA-2 grew at up to 2,400 ug/mL spectinomycin. TPMT-10 reached a kcat of 1.02 per second versus 0.33 for CtHMT.
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 EnzyGen2 on Tamarind Bio
De novo enzyme design: generate new sequences and backbones for a target substrate or cofactor.
Antibiotic-resistance enzyme families: design functional variants of enzymes such as CAT and AadA, as demonstrated in the paper.
Methyltransferase engineering: explore TPMT-like designs with measurable catalytic activity.
Taxonomy-aware design: steer designs toward the sequence preferences of a chosen organism group.
How to Use EnzyGen2 on Tamarind Bio
Open the tool: log in to tamarind.bio and open EnzyGen2.
Specify the ligand: provide the small-molecule ligand the enzyme should bind.
Add functional residues: if exposed, supply functionally important residues or a taxonomic identifier.
Set sampling: if exposed, adjust nucleus sampling and the number of designs.
Run the job: submit the design run.
Download outputs: save the designed sequences and backbone structures.
Filter and validate: apply thresholds such as pLDDT above 80 and review active-site geometry before experimental testing.
Things to Keep in Mind
Ligand constraints act in training and candidate selection, not inside the generative process.
Only C-alpha backbones are modeled; full-atom modeling is future work.
Experimental validation covers three enzyme families, and CAT and AadA were assessed mainly by survival, MIC and LC-MS rather than kinetics.
Candidate selection included expert review, so it is not fully automated.
Pretraining negatives are randomly sampled small molecules, not confirmed non-binders.
The paper reports slightly inconsistent parameter counts and identity ranges in different sections.
Source: Song Z, Liu H, Zhao Y, Yang Y, Li L. Co-designing sequence and structure of functional de novo enzymes with EnzyGen2. bioRxiv, 2026, DOI 10.64898/2026.03.02.709205, https://www.biorxiv.org/content/10.64898/2026.03.02.709205v2.