Proto (Generative Biology)

Proto is an open-source programming framework for generative biology developed by researchers at Stanford University and the Arc Institute. It lets a user assemble biological sequence generators, predictive models, scoring rules, and search algorithms into a common propose-score-refine workflow. Its stated scope includes DNA, RNA, proteins, ligands, and interactions among them. Proto is therefore an orchestration and design system, not a new foundation model of its own.^[1][3]

The team released Proto on June 23, 2026 alongside a bioRxiv v1 preprint by Aditi T. Merchant, Daniel Guo, Ben Viggiano, Lucas Brennan-Almaraz, Evelyn Hur, Tina Mai, Peter Yin, Samuel H. King, Euan A. Ashley, and Brian L. Hie.^[1][2] As of August 15, 2026, that manuscript had not been certified by peer review, and no independent replication of its experimental results had been identified. Claims about its design performance consequently describe the authors' reported preprint results, not an established consensus.

Proto connects generative AI with tools from bioinformatics through a Python library, a graphical web interface, and an agent-facing interface. The released study includes two new programs that received wet-lab testing: cell-line-selective introns and bacterial promoter-repressor pairs. Five prior design workflows and three larger agentic examples were computational demonstrations only.^[1][3]

Release and scope

Proto was produced by the Laboratory of Evolutionary Design, with affiliations to Stanford and Arc Institute. The June 2026 paper describes the project as a high-level language that shifts a design specification away from literal parts and toward desired functions. A program can ask for properties such as a predicted fold, binding interaction, splice pattern, or regulatory activity, then use one or more existing models to propose and score sequences.^[1][2]

This abstraction does not remove the need for biological knowledge or experiments. The project site's own description says that model quality varies across problems and that generated candidates still need expert interpretation, construction, and testing.^[3] The manuscript's disclosure states that Brian Hie had outside interests in Arpelos Biosciences and Genyro as a scientific co-founder; the other authors declared no competing interests.^[1]

Language and execution model

The paper defines four conceptual primitives. They separate what is being designed, how candidates are produced, what makes a candidate desirable, and how the search proceeds.^[1]

Scroll sideways for more →

A program can combine several constraints into a single energy score. In the paper's formulation, a generator provides a prior over sequences and the constraints reweight that distribution toward lower-energy candidates. This makes multi-objective design a sum of constraint energies, equivalent to a product of experts. Available search patterns include rejection sampling, gradient-based optimization, and simulated annealing using Markov chain Monte Carlo. A program can also chain several optimization stages, allowing the output of one stage to seed the next.^[1]

The software implementation has seven core runtime abstractions rather than only four. Sequence, Segment, and Construct hold candidate data and organize regions of a larger construct. Constraint, Generator, and Optimizer implement the pluggable operations. Program orchestrates the stages. This is an implementation expansion of the four-part conceptual model, not a separate language design.^[4]

Under the language, proto-tools gives each integrated model or computational program an isolated environment. It handles dependencies, model weights, reference databases, data exchange, GPU selection, and parallel execution. The authors report integrations for more than 120 tools, although that catalog size and the behavior of every integration have not been independently audited.^[1][5] Programs can run on local hardware or use optional remote containers through Modal. Some workloads therefore require substantial external compute even though the language itself is a Python package.^[4][5]

The public interfaces include local Python programs, JSON program descriptions, a browser-based graphical canvas, and an agentic workflow in which coding agents construct or revise programs from natural-language instructions. The project also documents Model Context Protocol access.^[1][3] Agent-generated code remains a design specification; it does not convert a model prediction into experimental evidence.

Evidence reported in the preprint

The paper contains three different levels of evidence that need to be kept separate. The first group re-expresses earlier computational workflows in Proto. The second group includes newly built and assayed sequences. The third uses agents to compose larger programs but stops at computational prediction.^[1]

Computational recapitulations

The authors rebuilt workflows for symmetric protein homo-oligomers, Protein Hunter monomers, CRISPR-Cas system generation, antibody complementarity-determining regions, and a 20-kilobase chromatin-accessibility pattern whose predictions encoded the word PROTO in Morse code. These examples exercised different generators, modalities, and constraints.^[1]

They do not provide new wet-lab validation for those five tasks. Their main evidentiary value is software expressivity: the programs could reproduce the computational structure of previously described campaigns. A favorable structure score, chromatin prediction, or sequence filter in this part of the study is not a measurement of biological function.

Cell-line-selective introns

The first novel experimental program designed introns intended to splice differently between K562 cells and either SH-SY5Y or HepG2 cells. The program combined AlphaGenome splice-site-usage predictions, SpliceTransformer scores, random mutation, and MCMC search. Because AlphaGenome did not supply a direct SH-SY5Y output, the researchers used its neural-cell-ontology prediction as a proxy for that cell line.^[1]

The team selected 65 candidate ProtoIntrons across four directional objectives and embedded them in an mScarlet dual-reporter construct. It then measured isoforms with next-generation sequencing after introducing the libraries into the corresponding cell lines. Among the candidates that were successfully assayed by sequencing, 32 percent showed statistically significant differential splicing in the intended direction. Three of the four objectives yielded a significant design; the HepG2-mis-spliced and K562-spliced objective did not.^[1][8]

One reported candidate, PI-KS1, showed 36 percent splicing in SH-SY5Y and 71 percent in K562. PI-S1 and PI-H1 showed intended-direction cell-line differences of 11 and 30 percentage points, respectively. AlphaGenome placed 85 percent of observed junctions within 20 base pairs of a predicted junction, but the authors found that no individual scoring constraint cleanly separated successful candidates from failures.^[1]

The paper compares the 32 percent figure with a less than 7 percent rate reported by another computational alternative-splicing study after testing roughly 1,000 sequences. It also acknowledges that the experimental systems differed.^[1] The comparison is therefore contextual, not a controlled head-to-head benchmark. The relevant denominator is also the successfully sequenced subset, not all 65 selected candidates.

Promoter-repressor pairs

A second program co-designed bacterial DNA promoters and protein repressors. For promoter proposals, the workflow used Evo 2, rejection sampling, promoter-calculator scores, sigma-70 motif constraints, and operator-site placement. The researchers selected 93 promoters for synthesis and successfully synthesized and cloned 86. In the reported E. coli eGFP assay, 45 of 86 produced expression more than tenfold above the no-promoter control, while 71 of 86 exceeded the PLtetO1 promoter. Composite design energy had a moderate inverse correlation with measured promoter activity, with Spearman rho of -0.38 and P = 3.0 x 10^-4.^[1][8]

The repressor program used several stages combining Evo 2, Boltz-2, AlphaFold 3, LigandMPNN, NA-MPNN, DeepPBS, and Rosetta-based scores. Across 12 promoter targets, 46 percent of tested repressor designs significantly reduced eGFP relative to non-targeting controls. Two promoter targets had no successful design. Nine percent of repressors exceeded a 1.5-fold reduction; the tetR positive control produced 2.38-fold repression in the same study. Adding a second operator site improved repression in a follow-up test.^[1][8]

These results demonstrate activity for measured sequences in that reporter system. They do not show that every generated repressor works, nor do predicted structural novelty and low similarity establish function. The paper compares its hit rate with earlier protein-DNA design studies, but those studies used different targets, assays, and selection procedures. Its description of a leading success rate is an author claim rather than an independent ranking. The team also reported that no one computational score separated its functional and nonfunctional repressors.^[1]

Agentic computational demonstrations

The study used coding agents to write and revise more complex Proto programs. In one demonstration, the authors curated 249 human protein complexes spanning 797 genes. Programs used ESM3 to propose variants and AlphaFold 3 to rescore multimer structures. For complexes with a known reference, the authors reported median predicted RMSD of 1.9 angstroms, TM-score of 0.80, pTM of 0.65, pLDDT of 71, and 80 percent sequence identity. These are structure-model outputs and similarity measures, not assays of complex formation or function.^[1][8]

A second agentic program redesigned parts of the beta-2 adrenergic signaling pathway. A third specified a proposed non-small-cell-lung-cancer-selective lentiviral circuit built around an HSV-TK payload, with predicted layers for binding, regulatory DNA, splicing, and post-transcriptional control. The paper reports favorable model scores for individual parts, but it did not build or test either system.^[1] In particular, the lung-cancer example is a computational design exercise, not evidence of tumor selectivity, treatment efficacy, or clinical safety.

The authors describe structure prediction as a plausibility filter rather than a guarantee of function. That qualification is important because these agentic examples are much more complex than the reporter constructs tested in the same manuscript.^[1]

Software, access, and data

The proto-language and proto-tools repositories are public under the MIT License. Their package metadata identifies version 0.1.0, Python 3.10 or later, and Alpha development status.^[4][5] PyPI hosts a 0.1.0 source distribution and universal wheel for proto-language, both uploaded on June 23, 2026.^[7] The repository README still recommends GitHub installation and says a direct PyPI install is planned, so its installation text lagged behind the published artifact when checked in August 2026.^[4][7]

The MIT licenses apply to Proto's code, not automatically to every model, dataset, or external program it can call. Project documentation lists ESM3, AlphaGenome, AlphaFold 3, and X3DNA among dependencies that require separate license acceptance, access approval, authentication, or registration.^[6] A user must also supply appropriate hardware or configure remote execution for resource-intensive tools. Open source access to the orchestration layer does not mean that every integrated weight set is unrestricted or locally practical.

The public supplement contains three data packages. Data S1 includes intron scores, AlphaGenome prediction tracks, and plasmid and primer sequences. Data S2 provides tested promoter and repressor sequences, scores, and cloning information. Data S3 lists the human proteins, complexes, and structure identifiers used in the diversification demonstration.^[8] The manuscript says raw sequencing data will be deposited in public repositories upon publication and that experimental materials are available on reasonable request.^[1] The supplementary design files are useful for inspection, but they are not the promised future raw sequencing deposit.

Limitations and safety

Proto's central limitation is the gap between a computational score and biological function. The study directly tested alternative-splicing reporters in selected human cell lines and promoter-repressor reporters in bacteria. It did not experimentally validate all supported modalities, the five recapitulated workflows, or any agentic multi-system example. Results from those assays also do not establish organismal, therapeutic, or clinical performance.^[1]

The paper is a first-version preprint from the system's developers. No peer-reviewed Proto publication or independent experimental replication was available by August 15, 2026. The authors themselves report failure modes: one intron objective produced no significant design, two promoter targets lacked a successful repressor, and individual in-silico metrics did not reliably divide hits from misses. Better models and faster optimizers may improve the framework, but every design still requires an appropriate validation plan.^[1][3]

On safety, the authors say they built Proto for research and education. They state that the hosted API and web application permit usage monitoring and that the hosted agentic interface adds safeguards beyond those in its underlying language models. The paper does not publish a safety benchmark, threat model, or detailed evaluation of those controls.^[1] It also does not establish that hosted monitoring applies to locally executed open-source code. Claims about safeguards should therefore remain limited to the hosted services described by the authors.

Relationship to earlier biological design languages

Proto follows earlier efforts to make biological engineering programmable. Eugene provides a domain-specific language for specifying parts, devices, composition rules, and constraints.^[9] Cello compiles a circuit specification into a design using characterized biological parts and predictive models.^[10] The Synthetic Biology Open Language provides a community standard for representing and exchanging designs.^[11]

Proto's authors position their system differently: generators and constraints express desired functions and search beyond a fixed catalog of literal parts. They also aim to span several molecular modalities in one workflow.^[1] That distinction explains the design goal, but it does not show that Proto replaces Eugene, Cello, SBOL, or specialized pipelines. Each addresses a different layer of biological design. Peer-reviewed reviews of machine learning for functional protein design likewise emphasize the importance of combining computational generation with experimental testing, but they predate Proto and do not independently evaluate it.^[12]

References

1. ^Aditi T. Merchant et al., *A high-level programming language for generative biology with Proto*, bioRxiv v1, posted June 23, 2026. biorxiv.org/...2026.06.22.733870v1
2. ^Arc Institute, *Proto: A programming language for generative biology*, June 23, 2026. arcinstitute.org/...proto
3. ^Proto project, *About Proto*, accessed August 15, 2026. proto.evodesign.org/about
4. ^Evolutionary Design, *proto-language*, GitHub repository, accessed August 15, 2026. github.com/...proto-language
5. ^Evolutionary Design, *proto-tools*, GitHub repository, accessed August 15, 2026. github.com/...proto-tools
6. ^Evolutionary Design, *Gated models*, proto-tools documentation, accessed August 15, 2026. github.com/...gated-models.md
7. ^Python Package Index, *proto-language 0.1.0*, accessed August 15, 2026. pypi.org/...proto-language
8. ^Merchant et al., *Proto supplementary material, Data S1-S3*, bioRxiv, June 23, 2026. biorxiv.org/....22.733870v1.supplementary-material
9. ^Anton Bilitchenko et al., *Eugene - a domain specific language for specifying and constraining synthetic biological parts, devices, and systems*, PLOS ONE 6(4), 2011. doi.org/...journal.pone.0018882
10. ^Timothy S. Jones et al., *Genetic circuit design automation with Cello 2.0*, Nature Protocols 17, 2022. doi.org/...s41596-021-00675-2
11. ^Michal Galdzicki et al., *The Synthetic Biology Open Language (SBOL) provides a community standard for communicating designs in synthetic biology*, Nature Biotechnology 32, 2014. doi.org/...nbt.2891
12. ^Jennifer Listgarten et al., *Machine learning for functional protein design*, Nature Biotechnology 42, 2024. doi.org/...s41587-024-02127-0