It integrates generative models across DNA, RNA, and proteins.
"The Fleshers have made a move"
In new work, we lay out a vision for a high-level programming language for generative biology, called Proto.
Proto composes generative and predictive models spanning DNA, RNA, proteins, ligands, and their interactions, which we use to design complex biological functions. 1/n
Users are excited about the new Proto language for programming generative biology because they see the research as cool, huge, and inspiring.
No Digg Deeper questions have been answered for this story yet.
Current biological design workflows have made tremendous advances, but are highly specialized, siloed, and hard to compose, preventing them from generating more complex systems. Unlocking multi-objective and multi-scale design is a key goal of our work. 2/n
This project had incredible scope, led by an extremely bold team of @aditimerch @dan_guo_ @bviggiano_13 et al. Excited to see what you program!
Preprint: https://www.biorxiv.org/content/10.64898/2026.06.22.733870 About: https://proto.evodesign.org/about GitHub: https://github.com/evo-design/proto-language Q&A: https://arcinstitute.org/news/proto n/n
A key idea is that biological design reduces to four primitives: sequences, constraints, generators, and optimizers. Proto abstracts these primitives to enable compositionality at a high level of abstraction, where generative models bridge the gap to low-level sequences. 3/n
None of this would have been possible without the support of @BrianHie, my brilliant co-first authors @dan_guo_ and @bviggiano_13, and the rest of the Proto team, @leba1127, @ev_hur, @tinabmai, @peteryin21, and @samuelhking! Incredibly grateful to have spent the last few months working alongside these incredible people in our small corner of Arc!
Proto has connections to both probabilistic and visual programming languages like Pyro or Scratch. We specify a generative design campaign with a factorized probability distribution using theory from energy-based models, which also has a nice graphical representation. 5/n
You can also write Proto programs in a web browser, including a chat interface that helps you understand or write new programs: https://proto.evodesign.org/prototypes. And here’s a video demo to learn more: https://youtu.be/8UnOrC_0jmo 10/n
Developing Proto was also a huge engineering lift, as the current ecosystem of biological AI models has conflicting assumptions and dependencies. @bviggiano_13's thread describes a lot of the challenges we overcame to make Proto practically useful: https://x.com/bviggiano_13/status/2069435319364309027 8/n
Proto has a minimal programmatic interface in which generators produce sequences, constraints score them, and optimizers steer toward more desirable designs. Not only do we implement a software API, but also a UI and an MCP/agentic interface. 4/n
We are making available an incredibly useful library of biological tools at https://proto.evodesign.org/tools. You can run structure prediction, inverse folding, seq2func models, etc. all in a browser! Take a look at @dan_guo_'s thread: https://x.com/dan_guo_/status/2069450157973016995 9/n
We also show that human steering of AI coding agents can produce Proto programs of high complexity. What would it look like to diversify the human proteome or design complex regulatory logic this way? 7/n
We experimentally validated Proto-generated eukaryotic and prokaryotic systems from all modalities of the central dogma (alternatively spliced RNA introns, specific interactions involving protein repressors and DNA promoters). See @aditimerch’s thread: https://x.com/aditimerch/status/2069449573916135509 6/n
Importantly, Proto is fully open-source, including the software, a web interface, and standardized access to 120+ tools, at http://proto.evodesign.org. We hope it lowers the barrier to composing these models and makes generative biological design more accessible.
AI in biology has advanced quickly: generators, structure predictors, and sequence-to-function models are each powerful in isolation. But combining them into a unified design campaign still often means stitching together a bespoke pipeline.
Traditional biological engineering composes physical parts from nature using intuition, screening, and build-test-learn cycles. Proto instead defines modularity at a functional level and uses generative modeling to bridge that abstraction down to sequence.
Because Proto is natively multi-objective, we also co-designed interacting components: synthetic promoter–repressor pairs in E. coli.
Despite not using a DNA-binder-specific model, several repressors were functional and preferentially recognized their target designed promoters, from just a well plate of candidates.
To test generality, we first reimplemented five published campaigns in Proto: symmetric proteins, de novo monomers, CRISPR-Cas systems, antibody CDRs, and multi-kilobase chromatin designs. Despite varied models, design objectives, and optimizers, all were expressible in Proto.
We next applied Proto to more synthetic tasks where close natural analogs are less available. We designed introns with cell-line-specific splicing in human cells by inverting AlphaGenome and SpliceTransformer, using models built to predict splicing as constraints to design it.
Testing only tens of candidates, roughly a third of our designs spliced in the intended direction.
Proto rests on four primitives: sequences, constraints, generators, and optimizers. Constraints score designs, generators propose candidates, and optimizers steer the search toward better ones.
These are early steps. Proto is only as capable as the models it composes, and designs require careful experimental testing. However, the framework and experiments reinforce each other: new models can plug in easily, and reproducible model combinations make experimental readouts a cleaner signal for improving models over time. We're excited to see Proto grow and evolve as the Bio ML field advances.
Finally, we paired Proto with AI coding agents to specify larger systems, including a diversified signaling cascade and a cancer-targeting circuit design that satisfied complex functional specifications.