evedesign: unified framework for accessible biosequence design

What is evedesign?

evedesign is a unified open-source framework for biosequence design that formalizes conditional design problems in a method-agnostic way. The framework allows users to seamlessly build and execute complex multiobjective design workflows, including supervised and unsupervised models, from a standardized set of specifications and operations. An interactive web-based user interface facilitates end-to-end biomolecular design for a broad scientific audience and is publicly available at https://evedesign.bio.

This repository implements the core interfaces for standardizing the interaction with biomolecular models, generation of nucleotide sequences, and many other utility functions for structure handling, sequence space embeddings, etc.

Please also check evedesign-server for automated pipeline execution from declarative design specifications and the REST API, as well as evedesign-ui for the interactive user interface.

Publication

Hopf TA, Gazizov A, Garcia Busto S, Eschbach E, Lee S, Mirdita M, Orenbuch R, Belahsen K, Ross D, Sander C, Steinegger M, d'Oelsnitz S, Marks D. evedesign: accessible biosequence design with a unified framework. bioRxiv (2026) doi:10.64898/2026.03.17.712115

Installation

Use the following command to install evedesign with support for all currently implemented models. You can remove any of the options if you do not need the respective model. Please see specific instructions for Boltz-2 further below.

pip install evedesign[evmutation2,esm2,mpnn,umap,gpytorch] 

Boltz2 installation

GPU-based (CUDA, recommended)

For structure predictions with Boltz-2 and CUDA, install the boltz2fold-cuda extra with uv:

uv pip install evedesign[boltz2fold-cuda]

PyPI's default torch ships with CUDA support (tested on a CUDA 13 build, which runs on CUDA 12.x+ drivers via forward compatibility). If you need a specific CUDA build to match an older driver, you might explore adding the matching PyTorch index, e.g.:

    uv pip install evedesign[boltz2fold-cuda] \
        --extra-index-url https://download.pytorch.org/whl/cu126 \
        --index-strategy unsafe-best-match

uv is required for this install path: boltz and its dependencies ship overly-conservative version pins that are overridden in [tool.uv] in pyproject.toml, and pip does not honor those overrides.

CPU/MPS-based

For CPU/MPS-only use (no CUDA), install the boltz2fold extra instead of boltz2fold-cuda.

Getting started

Please refer to some of our examples how to use evedesign. We are continuously extending these as new models are added to the framework.

To implement your own models in the framework, please have a look at our existing reference implementations (e.g. EVmutation2, ESM-2, ProteinMPNN, Gibbs sampler) as well as the underlying model interfaces and description of molecular systems and instances.

We are happy to help if you have any questions!

Currently available models and methods

Biomolecular models, embedders and restraints

Supervised models

Samplers

Analyzers

Nucleotide sequence generation

Roadmap and contributing

We plan to continuously add more models, restraints, oracles and samplers to the framework, e.g. de novo 3D structure generation with BoltzGen or BindCraft.

We are actively looking for further contributors to develop our framework jointly with the community. If you are interested or feel like an important model is missing from the framework, please get in contact with us!

License

evedesign is released under the MIT license.

Contact

For general questions or inquiries about evedesign please reach out to hello@evedesign.bio.