EvoDiff is Microsoft’s open-source framework for generating protein sequences and evolutionary alignments with discrete diffusion models. It treats protein design as a generative modeling problem over amino acid sequences rather than continuous structures. The repository has about 688 stars on GitHub and was updated in January 2026.
Proteins are strings over a twenty-letter alphabet, so the continuous diffusion methods used for images do not apply directly. EvoDiff adapts diffusion to this discrete setting, learning from large collections of natural protein sequences and multiple sequence alignments that capture evolutionary relationships. The trained models can generate novel, plausible protein sequences unconditionally, or conditioned on evolutionary context from an alignment, helping researchers explore sequence space beyond what nature has sampled. Released by Microsoft Research under the MIT license, the repository includes the models and code needed to reproduce the published approach to sequence-level protein generation.
In practice, researchers use EvoDiff to propose candidate sequences for a desired function, then filter the proposals with structure prediction tools and, ultimately, laboratory assays. By generating from evolutionary patterns rather than copying known proteins, it can suggest sequences that are novel yet still biologically plausible.
EvoDiff is free and open source under the MIT license. Researchers can download the code and models from GitHub and run them on their own GPU infrastructure at no licensing cost.
EvoDiff is for computational biologists, protein engineers and ML researchers working on generative models for biology. It fits teams designing new protein sequences in silico, or studying how diffusion methods can capture evolutionary patterns in sequence data.