Co-designing sequence and structure of functional de novo enzymes with EnzyGen2

EnzyGen2 is a 730-million-parameter protein foundation model that enables the rapid, simultaneous co-design of sequence and structure for functional de novo enzymes by leveraging multi-task learning on protein-ligand pairs, achieving superior performance and experimental validation across multiple enzyme families compared to state-of-the-art methods.

The Gist

Imagine you are an architect trying to build a brand-new, custom-made key that fits a very specific, complex lock. In the world of biology, that "lock" is a tiny chemical molecule (a ligand), and the "key" is a protein (an enzyme) designed to grab onto it and perform a specific job, like breaking down a toxin or creating a medicine.

For decades, scientists have struggled to design these keys from scratch. Usually, they had to build the shape of the key first, and then try to figure out what material to make it out of. If the shape was slightly off, the key wouldn't work. It was like trying to sculpt a statue out of clay, only to realize halfway through that the clay isn't the right type to hold the shape.

Enter EnzyGen2: The "Master Key" Architect

This paper introduces EnzyGen2, a super-smart AI that changes the game. Instead of building the shape and the material separately, EnzyGen2 designs them at the same time.

Here is how it works, using some everyday analogies:

1. The "Recipe Book" and the "Chef"

Imagine the world's largest library of recipes (proteins) and the ingredients they work with (ligands). Scientists fed EnzyGen2 a massive cookbook containing 720,000 recipes of real proteins and the chemicals they interact with.

• The Old Way: Previous AI tools were like chefs who could only guess the ingredients based on the shape of the pot, or guess the pot shape based on the ingredients. They often got it wrong.
• The EnzyGen2 Way: This AI is a master chef who looks at the dish you want to make (the chemical reaction), the kitchen you are in (the type of organism, like E. coli), and the special tools you need (the ligand), and then instantly writes a brand-new recipe and builds the perfect pot to cook it in.

2. The "Evolutionary GPS"

One of the coolest features of EnzyGen2 is that it doesn't just guess; it knows the rules of nature.

• The Analogy: Imagine you are trying to write a story in a specific language (like French). If you just pick random words, it won't make sense. EnzyGen2 uses NCBI taxonomic identifiers as its "GPS." It knows, "Oh, you want this enzyme for a bacteria? Okay, I'll only use words and grammar rules that bacteria actually use."
• This stops the AI from creating "alien" proteins that look cool on a computer but fall apart in real life. It keeps the design within the realm of what nature has already proven works.

3. The "Double-Check" System

EnzyGen2 is trained with three goals at once (Multi-task Learning):

1. Sequence: "Does the list of amino acids (the letters) make sense?"
2. Structure: "Does the 3D shape fold correctly?"
3. Binding: "Will this shape actually grab the target chemical?"

Think of it like a car factory where the assembly line, the design team, and the quality control inspector all talk to each other in real-time. If the design looks good but the engine won't fit, the system fixes it immediately.

The Results: From Theory to Reality

The team didn't just stop at computer simulations. They took the AI's designs and built them in a real lab. They tested it on three different types of enzymes:

• CAT: A protein that breaks down the antibiotic chloramphenicol.
• AadA: A protein that breaks down another antibiotic.
• TPMT: A protein used to help make other medicines.

The Magic:

• Speed: EnzyGen2 generates designs 400 times faster than previous methods.
• Success: The AI created enzymes that worked just as well as, or even better than, the natural ones found in nature.
• Novelty: These weren't just copies of existing enzymes. They were brand new creations with only about 50% similarity to natural proteins. It's like the AI invented a new type of car engine that runs on a completely different principle than anything we've seen before, but it still drives perfectly.

Why This Matters

This is a huge leap forward for "Green Chemistry" and medicine.

• Medicine: We can design enzymes to create new drugs faster and cheaper.
• Environment: We can design enzymes to eat plastic or clean up oil spills.
• Sustainability: We can create biological factories that produce fuel or materials without pollution.

In a nutshell: EnzyGen2 is a powerful new tool that teaches computers how to speak the language of life. It allows us to design custom biological machines from scratch, solving problems that nature hasn't solved yet, all while doing it faster and more accurately than ever before. It's not just copying nature; it's learning from nature to build something better.

Technical Summary

1. Problem Statement

The design of de novo proteins, particularly enzymes, faces a critical bottleneck: functional targeting. While recent advances allow for the generation of protein backbones and sequences, accurately designing enzymes that bind specific small-molecule ligands (substrates or cofactors) remains a formidable challenge.

• Limitations of Current Methods: Most existing strategies follow a sequential paradigm (generate backbone $\rightarrow$ design sequence), which fails to capture the reciprocal dependencies between sequence, structure, and ligand binding.
• Data Scarcity: High-quality protein-ligand complex data is scarce (only ~20,000 publicly available), hindering the training of ligand-aware generative models.
• The Gap: There is a lack of a general, ligand-aware foundation model capable of simultaneously optimizing protein sequence and 3D structure to achieve specific catalytic functions with high efficiency and novelty.

2. Methodology: EnzyGen2

The authors introduce EnzyGen2, a 730-million-parameter protein foundation model designed for the simultaneous co-design of sequence and structure under ligand-guided functional targeting.

A. Data Curation

To overcome data scarcity, the authors constructed a massive dataset of 720,993 protein-ligand pairs:

• Sources: Merged experimentally validated proteins from the Protein Data Bank (PDB) and Swiss-Prot.
• Ligands: Retrieved from UniProtKB.
• Expansion: This dataset is more than an order of magnitude larger than existing complex datasets.
• Taxonomy: Proteins were clustered by NCBI taxonomic identifiers (9,002 distinct clusters) to encode evolutionary constraints.
• Labeling: Proteins with documented ligands were labeled positive; those without were assigned random small molecules as negative controls.

B. Model Architecture

EnzyGen2 employs an interleaved neural network architecture combining:

1. Transformer Layers: To capture long-range sequence dependencies and global amino acid correlations.
2. k-Nearest Neighbor Equivariant Graph Neural Networks (kNN-EGNN): To model 3D structural geometries and residue correlations in space.
3. Protein-Ligand Interaction Module: A dedicated module to enforce ligand-binding specificity.
4. Heterogeneous Inputs:
   • NCBI Taxonomic Identifiers: To constrain the search space to evolutionarily plausible regions.
   • Functionally Important Residues: Automatically identified via Multiple Sequence Alignment (MSA) to guide specific functions without manual annotation.
   • Ligand Structures: To capture the physics of binding.

C. Training Objective

The model is trained using a multi-task learning framework with a weighted loss function:
$$L = \alpha L_{seq} + \beta L_{str} + \lambda L_{bind}$$

• $L_{seq}$: Masked protein sequence prediction.
• $L_{str}$: Masked backbone structure reconstruction.
• $L_{bind}$: Protein-ligand interaction prediction (binary classification of binding).
• Pretraining: 3 stages involving masked language modeling, structure reconstruction, and finally the full multi-task objective.
• Fine-tuning: Adapted to specific enzyme families (CAT, AadA, TPMT) using family-specific datasets.

D. Candidate Selection Pipeline

A rigorous screening pipeline was used to select candidates for experimental validation:

1. Scoring: Enzyme-Substrate Prediction (ESP) score, Adversarial Discriminator score (to distinguish natural vs. artificial), and AlphaFold2 pLDDT (stability).
2. Filtering: Thresholds set at ESP > 0.6, Discriminator > 0.6, and pLDDT > 80.
3. Diversity: Candidates grouped by sequence identity to natural homologs (30–80% bins) to ensure novelty while maintaining foldability.
4. Expert Review: Geometric alignment of active sites with natural templates.

3. Key Contributions

• First Ligand-Aware Foundation Model: EnzyGen2 is the first large-scale model to jointly co-design sequence and structure explicitly guided by ligand binding and evolutionary taxonomy.
• Massive Dataset Construction: Curated a dataset of ~721k protein-ligand pairs, significantly expanding the available training data for ligand-aware design.
• Speed and Efficiency: The model generates samples 400$\times$ faster than state-of-the-art baselines (e.g., RFdiffusion/ProteinMPNN pipelines).
• Experimental Validation: Successfully designed de novo enzymes for three distinct families (CAT, AadA, TPMT) with catalytic activities rivaling or exceeding natural enzymes, despite low sequence identity.

4. Results

A. In Silico Benchmarks

• Performance: EnzyGen2 outperformed state-of-the-art baselines (Inpainting, RFdiffusion/ProteinMPNN, RFdiffusion2/3 + LigandMPNN) across 10 enzyme families.
  • ESP Scores: Superior in 7/10 categories.
  • Structural Fidelity: Higher proportion of designs with RMSD < 2Å (9/10 categories).
  • Stability: Higher AlphaFold2 pLDDT scores (mean 85.10).
• Ablation Studies: Confirmed that removing any component (Taxonomic ID, Ligand Module, Co-design framework, or Functionally Important Residues) significantly degraded performance. The functionally important residues were identified as the most critical guidance signal.

B. Experimental Validation

The authors experimentally validated designs for three enzyme families:

1. Chloramphenicol Acetyltransferase (CAT):

   • Outcome: 7 of 20 designed variants enabled E. coli survival on chloramphenicol.
   • Performance: Variant CAT-17 showed resistance to 500 µg/mL chloramphenicol, outperforming the wild-type enzyme which failed at this concentration.
   • Structure: RMSD of 0.54 Å against natural homolog.

2. Aminoglycoside Adenylyltransferase (AadA):

   • Outcome: 8 of 20 variants conferred resistance to aminoglycosides.
   • Performance: Variant AadA-2 allowed growth in 2400 µg/mL spectinomycin, a substantial enhancement over the wild-type Salmonella AadA.
   • Structure: RMSD of 1.08 Å against natural homolog.

3. Thiopurine S-methyltransferase (TPMT):

   • Application: Designed for S-adenosylmethionine (SAM) regeneration using methyl iodide.
   • Outcome: 6 of 10 variants were highly expressed.
   • Performance:
     • TPMT-4: High catalytic efficiency ($k_{cat}/K_M = 0.091 \mu M^{-1}s^{-1}$), comparable to natural HMTs.
     • TPMT-10: High turnover number ($k_{cat} = 1.02 s^{-1}$), exceeding natural homologs.
   • Novelty: TPMT-10 aligned with an uncharacterized protein clade (30% identity), demonstrating the model's ability to explore uncharted protein space.

5. Significance

• Paradigm Shift: Moves protein design from sequential (backbone-first) to simultaneous co-design, better capturing the sequence-structure-function relationship.
• Biotechnological Impact: Demonstrates the ability to create high-performance, novel biocatalysts that can outperform natural enzymes, with potential applications in green chemistry, therapeutics, and industrial biotechnology.
• Scalability: The speed (400$\times$ faster) and robustness of EnzyGen2 make it a practical tool for rapid enzyme discovery, potentially accelerating the development of enzymes for reactions not accessible through natural evolution.
• Evolutionary Insight: By leveraging taxonomic identifiers and MSA-derived constraints, the model successfully navigates evolutionary space to find functional solutions that are both novel and biophysically stable.