evedesign: accessible biosequence design with a unified framework

The paper introduces evedesign, a unified open-source framework and web interface that overcomes current limitations in protein engineering by enabling flexible, method-agnostic, and iterative multi-objective biosequence design workflows accessible to both experts and non-experts.

The Gist

Imagine you are a master chef trying to invent a new recipe for a dish that needs to be delicious, healthy, and able to survive being shipped across the ocean without spoiling.

In the world of biology, proteins are the "dishes" (or machines) that keep life running. Scientists want to "cook up" new proteins to cure diseases, clean up pollution, or make better fuels. For years, they've had amazing new "cooking tools" (AI models) to help them design these proteins. But there was a huge problem: every tool spoke a different language.

One tool only understood written recipes (sequences), another only understood 3D shapes (structures), and a third only understood the history of the ingredients (evolution). To use them together, a scientist had to be a master translator, writing custom code to make them talk to each other. It was slow, expensive, and only experts could do it.

Enter "EveDesign": The Universal Kitchen.

This new paper introduces EveDesign, a unified framework that acts like a universal translator and a master kitchen organizer for protein engineering. Here's how it works, using simple analogies:

1. The Universal Language (The "Standardized Interface")

Think of EveDesign as a universal adapter plug. Before, if you wanted to use a British hairdryer in an American socket, you needed a specific adapter. If you wanted to use a German blender, you needed a different one.

EveDesign creates a single "socket" where any protein design tool can plug in. Whether the tool is an AI that learns from evolution, a language model that reads protein "text," or a 3D structure builder, they all speak the same language inside EveDesign. You can swap them out or combine them like Lego blocks without rewriting the whole system.

2. The "Smart Clipboard" (The Multi-Level Instance)

When a chef writes a recipe, they might list the ingredients (sequence), draw a picture of the dish (structure), or write notes about how it tastes (embedding).

EveDesign uses a "Smart Clipboard" (called an Instance). Every time a design is created, this clipboard holds all the information at once:

• The sequence (the letters).
• The 3D shape (the structure).
• The "vibe" or mathematical representation (the embedding).

This means a tool that only understands shapes can pass its work to a tool that only understands letters, and the information never gets lost. It's like passing a note that has a drawing, a translation, and a summary all on the same piece of paper.

3. The Three Magic Moves (Generate, Score, Transform)

Instead of complex coding, EveDesign breaks every design task down into three simple moves, like a game:

• Generate: "Create new ideas!" (The AI invents new protein sequences).
• Score: "Rate the ideas!" (The AI checks: Is this stable? Does it bind to the target? Is it safe?).
• Transform: "Change the format!" (Turn a sequence into a 3D shape, or turn a shape into a list of probabilities).

You can chain these moves together. For example: Generate 1,000 ideas -> Transform them into 3D shapes -> Score them for stability -> Generate a new batch based on the winners. This allows for complex "lab-in-the-loop" workflows where the computer learns from real-world experiments and improves the next round of designs automatically.

4. The "Open-Source Cookbook"

Previously, if you wanted to design a protein, you might need to buy expensive, closed software or hire a computer programmer to build a custom tool for you.

EveDesign is open-source (free for everyone to use and improve) and comes with a web interface. You don't need to be a coder. You can go to a website, type in your goal (e.g., "Design an antibody that binds to this virus but doesn't cause side effects"), and the system runs the complex workflow for you.

Why This Matters (The Real-World Impact)

The paper shows EveDesign working in three real scenarios:

1. Inventing Enzymes: It designed new versions of an enzyme (Chorismate Mutase) that are just as good as nature's best, proving the AI can "dream up" functional biology.
2. Fixing Antibodies: It combined a tool that reads protein "text" with a tool that reads "3D shapes" to find the perfect mutations for antibodies. The text tool found some good changes, the 3D tool found others, and together they found the best ones.
3. Finding Hidden Gems: It used a "teacher" (supervised learning) to scan thousands of natural proteins and find the few that are super-efficient at breaking down cancer-fighting drugs.

The Bottom Line

EveDesign is the "Operating System" for the future of protein engineering.

Just as Windows or macOS allowed anyone to use a computer without knowing how to build the hardware, EveDesign allows biologists, doctors, and researchers to design life-saving proteins without needing to be coding wizards. It turns a chaotic collection of disconnected tools into a single, powerful, collaborative engine that can solve the world's toughest biological problems.

Technical Summary

1. Problem Statement

The field of protein engineering faces a critical bottleneck: while machine learning (ML) methods for protein design have proliferated (including MSA-based models, large language models, inverse folding, and de novo structure design), they remain largely non-interoperable.

• Fragmentation: Existing tools operate on disparate data representations (sequences, structures, embeddings) with bespoke APIs, making it difficult to compare, combine, or swap models.
• Inaccessibility: Real-world design problems often require conditional design (optimizing under constraints like thermostability or T-cell epitope removal) and multi-objective optimization. Current workflows require labor-intensive, custom coding ("glue code") that excludes non-computational researchers.
• Lack of Iteration: There is no standardized infrastructure to support lab-in-the-loop workflows, where experimental data continuously refines successive design rounds.
• Proprietary Barriers: Existing solutions are often proprietary or fragmented, lacking open-source, interactive interfaces that span the workflow from target protein to orderable DNA.

2. Methodology: The evedesign Framework

The authors introduce evedesign, a unified, open-source, multi-layered framework designed to formalize biomolecular design as a conditional modeling problem.

A. Core Conceptual Architecture

• Conditional Modeling: Design is framed as generating molecular instances conditioned on a system specification. A "system" consists of entities (proteins, DNA, ligands) with known attributes (sequence, structure, homologs, binding partners).
• Multi-Level Instance Representation: Unlike frameworks enforcing a single representation, evedesign uses a standardized Instance object that simultaneously encodes:
  • Sequence
  • Positional Embeddings
  • 3D Structure (PDB format)
  • Note: Not all levels need to be populated, allowing seamless flow between sequence-only and structure-based models.
• Three Composable Operations: All workflows are constructed by chaining three standardized operations:
  1. Generate: Produces new design instances conditioned on the system (e.g., autoregressive sampling).
  2. Score: Assigns a quantitative fitness value (e.g., log-likelihood) to instances.
  3. Transform: Maps instances between representation levels (e.g., sequence $\to$ structure, sequence $\to$ embeddings).

B. Software Stack

The framework is delivered as a modular suite:

1. evedesign (Core Python Package): Implements the interface and includes reference models:
   • EVmutation2: A new, lightweight evolutionary model (see below).
   • ESM-2: A protein language model.
   • ProteinMPNN/LigandMPNN: For inverse folding.
   • Utilities: Gibbs samplers, MSA generation (MMseqs2), structure search (FoldSeek), and codon optimization.
2. evedesign_server: A REST API and pipeline runner for distributed execution, allowing self-hosting for privacy-sensitive or commercial deployments.
3. Interactive Web Interface: A React-based frontend (https://evedesign.bio) enabling end-to-end design (target to DNA) without coding.

C. Novel Model: EVmutation2

To address the speed limitations of existing evolutionary models (like EVE) which require per-target fine-tuning, the authors developed EVmutation2:

• Architecture: Couples an order-invariant autoregressive decoder with a compact AlphaFold3-based encoder (14.3M parameters).
• Mechanism: It moves all known tokens to a randomized prefix to maximize context for designed positions. It uses linear projection of pair representations for attention weights rather than full query-key-value attention, reducing parameters while enforcing biologically meaningful pairwise interactions.
• Training: Trained on the OpenProteinSet to reconstruct masked MSA sequences. It achieves performance comparable to EVE on ProteinGym benchmarks without per-target fine-tuning.

3. Key Contributions

1. Unified Interface: A standardized, method-agnostic interface that allows swapping or combining any model (supervised/unsupervised, sequence/structure) without bespoke code.
2. Declarative Workflow System: A system where complex, multi-objective pipelines are defined declaratively, supporting lab-in-the-loop iteration (halting, replaying, and updating with new data).
3. Open-Source Ecosystem: A fully open-source (MIT/AGPLv3) platform with a public web server and self-hostable backend, adhering to FAIR principles (Findability, Accessibility, Interoperability, Reusability).
4. EVmutation2: A new, fast, generative evolutionary model suitable for interactive server deployment.

4. Results and Case Studies

The paper validates evedesign through three distinct case studies demonstrating its flexibility:

• Unsupervised Enzyme Design (Chorismate Mutase):
  • Used EVmutation2 to generate 1,024 new sequences via autoregressive sampling.
  • Result: Generated sequences recapitulated natural amino acid statistics ($r > 0.96$) and showed zero-shot scores comparable to functional natural sequences. The designs interpolated the natural sequence distribution while showing meaningful extrapolation.
• Complementary Antibody Scoring (Sequence vs. Structure):
  • Combined ESM-2 (sequence-only) and ProteinMPNN (structure-conditioned) to score mutations in clinically relevant antibodies.
  • Result: The models provided complementary information. ESM-2 identified beneficial mutations generally, while ProteinMPNN specifically down-scored deleterious mutations at the antibody-antigen interface (within 6Å) that ESM-2 missed. This highlighted the necessity of multi-model workflows for interaction-sensitive tasks.
• Supervised Discovery (Kynureninase Efficiency):
  • Reproduced a study predicting catalytic efficiency ($k_{cat}/K_m$) using Random Forest regression on ESM-2 embeddings.
  • Result: The evedesign pipeline achieved cross-validation performance ($\rho = 0.73$) matching the original study. It successfully ranked unlabelled homologs, recovering the top experimental variants (K1-K3) in the top 10 rankings.

5. Significance and Future Outlook

• Democratization: By providing an interactive UI and standardized APIs, evedesign lowers the barrier to entry for experimental biologists, allowing them to perform complex, multi-objective designs without coding expertise.
• Scalability & Privacy: The self-hostable evedesign_server enables commercial and clinical users to run sensitive design workflows on private infrastructure, addressing data privacy concerns.
• Future of Protein Engineering: The framework is designed as a "living" foundation. Its architecture specifically anticipates iterative lab-in-the-loop workflows, where experimental data feeds back into the design cycle to refine models in real-time—a capability currently missing in ad-hoc tool collections.
• Community Growth: The modular design invites the community to contribute new models and workflows, ensuring the platform evolves alongside the rapidly advancing field of protein ML.

In summary, evedesign bridges the gap between theoretical ML advancements and practical protein engineering by providing a unified, interoperable, and accessible infrastructure for the next generation of biosequence design.