Designing Protein Structures from Natural Language: How Prompteina Turns Text Prompts Into All-Atom Proteins
TL;DR: Researchers at the University of Washington, Allen Institute for AI, National University of Singapore and Howard Hughes Medical Institute described Prompteina, a text-conditioned model that jointly generates protein sequence and all-atom structure from natural-language prompts.
Researchers have described a protein-design system that accepts plain written descriptions and returns complete protein structures with sequences. The work was posted on October 9, 2026, under the title Designing Protein Structures from Natural Language, and summarized on October 12, 2026, by Natural language can steer all-atom protein design across diverse objectives | Zotpaper. The central result is a model called Prompteina that produces an all-atom structure together with its amino-acid sequence in response to text embeddings from a prompt.
A single language interface for many design goals
De novo protein design seeks proteins with chosen structural and functional characteristics. The paper frames that task as spanning an overall fold or function down to characteristics of single residues within a chain.
Earlier structure-generation systems could be steered, but each characteristic generally needed its own engineered conditioning format. The paper gives two illustrations: a desired fold might be supplied as one categorical label, while desired secondary structure might be supplied as a per-residue string marking helix, strand or loop.
Prompteina replaces those separate formats with one natural-language interface. The same prompt channel can express broad objectives and fine-grained residue-level objectives. The authors compare the approach to text-to-image generation, where language provides a unified way to request diverse visual attributes without task-specific controls.
According to the arXiv posting, Prompteina is presented as the first all-atom structure-generation model that supports new and broader objectives stated in natural language. The model takes an input prompt and jointly creates structure and sequence conditioned on the text embeddings of that prompt.
Who built Prompteina
The author list on the arXiv posting names Joseph Lee, Yanjing Li, Bo Qiang, Gregory Lau, Allyson Ettinger, Noah A. Smith, Rory Donovan-Maiye, David Baker, Hanna Hajishirzi and Pang Wei Koh.
The affiliations listed are the University of Washington, the National University of Singapore, the Allen Institute for AI (Ai2) and the Howard Hughes Medical Institute. David Baker is listed with the University of Washington and the Howard Hughes Medical Institute. Noah A. Smith is listed with the University of Washington and the Allen Institute for AI.
The October 12 summary identifies Lee and colleagues as the developers and names the University of Washington, National University of Singapore, Allen Institute for AI and Howard Hughes Medical Institute as participating organizations. The arXiv record describes the document as 54 pages with 9 figures. The listing hosted by arXivDaily for paper 2610.13917 points to code, data and model weights at github.com/jleechung/prompteina.
How Prompteina differs from earlier text-linked methods
The paper distinguishes direct conditioning of both sequence and structure from two narrower uses of text described in prior work.
One earlier pattern used text to guide structure generation only at inference time, associated in the paper with Ingraham and colleagues in 2023. A second pattern used text to generate sequences that then had to be folded into structures, associated with Dai and colleagues in 2025 and Pei and colleagues in 2026.
Prompteina follows a different route. It directly conditions generation of both modalities on text. That design is intended to give robust control over structure and sequence together, rather than shaping one modality after the other.
The practical consequence highlighted in the paper is combined control of global characteristics such as fold class and residue-specific characteristics such as particular disulfide bonds and binder-target contacts. Fold class is therefore requested through the same language channel as disulfide-bond placement and contact specification.
Prompteina-Data: 1.8 million structures with tens of millions of captions
Training such a model required paired examples linking structures to language at two scales. Existing text-protein datasets, as characterized in the paper, paired sequences with only high-level functional descriptions.
To address that gap, the team constructed Prompteina-Data. The dataset is reported as containing 1.8 million structures. It contains more than 50 million captions, specified elsewhere in the paper as 52.4 million captions.
Those captions describe both functional and structural properties. They cover global properties and residue-specific properties. The October 12 digest describes the captions as derived from database annotations, scientific literature and properties computed from structures.
The arXiv abstract describes the resource as a paired text-protein structure dataset. The scale — 1.8 million structures and 52.4 million captions — is intended to support direct, fine-grained control rather than high-level functional labeling alone.
From Proteina-Complexa to text-conditioned generation
Prompteina did not start from random initialization. The team adapted Proteina-Complexa, described as a pretrained structure-generation model.
Adaptation used text-conditioned pretraining and finetuning. The finetuning step aligned residues with their corresponding text. In other words, portions of the prompt that refer to particular positions are linked to those residues during training.
The input during generation is a natural-language prompt. The output is an all-atom protein structure and its sequence. Conditioning is implemented through the prompt's text embeddings, which influence both outputs jointly.
Controlling folds, disulfides and contacts in words
The paper presents natural language as a way to request diverse objectives without building a new input representation for each one. Global properties and residue-specific properties travel through the same prompt.
Global properties include fold class. Residue-specific properties include specific disulfide bonds. They also include binder-target contacts, which specify how a designed binder should touch its target.
The figure in the paper illustrates that range with a single interface covering global and residue-specific properties. The Zotpaper summary likewise describes joint generation of sequences and structures conditioned on descriptions of desired properties.
Performance across 18 structural properties
The evaluation reports structural-property performance on 18 properties. Prompteina outperformed baselines on 15 of those 18 structural properties.
The paper also reports functional alignment. The generated proteins more closely matched descriptions of their intended functions than baseline outputs. The October 12 digest restates that finding as experimentally confirmed generation from textual specifications that outperformed specialized baselines.
Those two results together address structure and function. Structural-property tests measure whether requested geometric and architectural characteristics appear in the output. Functional-description matching measures whether the outputs correspond to the intended functional language.
Binder design across 20 targets
Binder design is a major test in the paper. A binder is a designed protein intended to attach to a chosen target. The evaluation spans 20 targets.
Prompteina led baselines on at least 13 of those 20 targets. The metric highlighted is interface confidence, which concerns the predicted binder-target interface.
Prompteina improved interface confidence by 14% over the strongest baseline. The October 12 summary frames the same program as producing experimentally confirmed binders while outperforming specialized baselines.
Across the 20-target comparison, the same agent specified designs for every model. Prompteina achieved the highest interface confidence under that shared specification procedure. That setup is shown in the paper's overview figure comparing models when prompts come from a common agent.
Agentic design: from literature research to structure
The paper also connects Prompteina to agentic protein design. In that workflow, a language-model agent researches target biology and writes text prompts describing candidate binders. Prompteina then realizes those research-informed specifications directly as generated structures.
The advantage stated for this arrangement is direct transfer of specifications. Language-model agents can pass specifications to the structure generator in natural language. They do not have to translate decisions into fixed, model-specific inputs that would constrain which design choices they can express.
That point is linked to work on agentic protein design associated with Shanehsazzadeh in 2026. Prompteina is presented as complementary to that line of work because it widens the range of design decisions an agent can convey.
Iterative in silico feedback on the hardest cases
Beyond single-shot generation, the paper tests iterative in silico feedback. In that loop, computational feedback informs subsequent generation rounds.
Prompteina's gain over non-iterative generation was 57% larger than the next-best baseline's gain on the hardest targets. In other words, iteration helped all models considered, but the additional benefit for Prompteina exceeded the next-best improvement by 57% where targets were most difficult.
The Zotpaper digest and arXiv abstract both place this result alongside the 15-of-18 and 13-of-20 comparisons. Together they portray advantages in direct generation and in refinement through feedback.
Experimental test against GM2A
The team moved one design campaign from computation to experiment. The target was GM2A, described as a target with no reported de novo binders.
Experimental validation identified nine binders against GM2A. The abstract reports the strongest at an apparent K_D of 107 nM. The overview figure reports the best replicate at an apparent K_D of 68.8 nM and describes the nine as diverse binders.
The arXiv abstract presents the nine binders as the outcome of Prompteina designs against GM2A. The October 12 digest likewise reports experimentally confirmed binders to GM2A from textual specifications.
GM2A therefore serves as a first-demonstration target without prior reported designed binders. Nine distinct designed proteins showed binding, with the strongest apparent K_D values reported in nanomolar units in the abstract and figure noted above.
What was released with the paper
The arXiv record for the October 9 posting states that code, data and model weights are provided. The arXivDaily page for the same paper gives the location as github.com/jleechung/prompteina.
The paper itself runs 54 pages with 9 figures, according to the arXivDaily listing. The author count is ten. The institutional count is four organizations across the United States and Singapore.
The Zotpaper checks for the October 12 digest note that its automated review found reported numbers on named benchmarks. The same automated checklist states that it did not locate a code link, weights link, dataset link, compute details or stated limitations in the text and links available to it, while cautioning that absence in its review does not mean the materials do not exist.
Why language conditioning changes the workflow
The traditional workflow required a dedicated input representation for each property. A fold label, a secondary-structure string and other fixed formats each had to be engineered separately. That limited the design objectives any single model could support.
The proposed workflow keeps the generator fixed and varies the language. New objectives are expressed as new sentences or phrases rather than new encodings. Global fold language and residue-level contact language enter together.
Direct conditioning matters for that workflow because both outputs respond to the same prompt. Sequence generation does not proceed independently and folding does not happen afterward as a separate stage. Structure and sequence emerge jointly from the text embeddings.
Residue-text alignment matters because prompts can refer to specific positions. The finetuning stage links residues to corresponding text so that local instructions affect local structure and sequence.
Context for United States science and technology readers
For readers tracking AI for biology, the notable claim is breadth without re-engineering. One model handles fold class, disulfide bonds, binder-target contacts and functional descriptions through prompts.
The notable evidence consists of four linked findings reported by the authors: 15 of 18 structural properties favoring Prompteina, closer functional-description matching, leadership on at least 13 of 20 binder targets with a 14% interface-confidence improvement, and nine experimental GM2A binders with a strongest apparent K_D of 107 nM in the abstract.
The iterative-feedback finding adds a fifth element: a 57% larger gain over non-iterative generation than the next-best baseline on the hardest targets. All five findings are reported in the October 9 arXiv posting and reflected in the October 12 research digest.
Readers who want to inspect methods, baselines and additional controls can consult the full 54-page document with 9 figures through the arXiv posting and the repository link listed on the arXivDaily page. The repository location given there is the place to look for the released code, data and weights referenced in the abstract.