Daily AI Paper

De novo protein design has made real progress, but it still has a fundamental usability problem. If you want a model to generate a protein with a certain fold, a specific secondary-structure pattern, a disulfide bond, a catalytic site, or a binding interface, many existing systems require you to encode that goal in a special-purpose representation. One model might take a fold label, another might take a residue-level structural sketch, and another might take a separate interface specification. That works when the property has been anticipated in advance, but it becomes limiting fast. The broader the design problem, the more awkward the conditioning interface becomes. This paper asks a simple but powerful question: what if protein design could work more like text-to-image generation? In image models, you can describe many different visual goals in natural language, and the same interface can steer style, objects, composition, and fine details. The authors bring that idea into protein structure generation with Prompteina, a text-conditioned all-atom protein generator that can take natural-language prompts and produce both structure and sequence. Why does that matter? Because proteins are not just shapes. A useful protein has to satisfy multiple constraints at once: global fold, local residue interactions, surface features, and often a specific functional task like binding another molecule. If language can serve as one shared interface for all of that, then a designer or an agent can express goals directly, without translating them into model-specific codes. That makes the system more flexible, easier to use, and potentially much more compatible with iterative, language-driven design workflows. The key challenge is data. A model cannot learn to follow rich natural-language prompts unless it sees protein examples paired with descriptions that actually talk about those properties. Existing text-protein datasets mostly connect sequences to high-level functional descriptions. That is useful, but it is not enough for fine-grained structural control. So the authors build a new resource called Prompteina-Data, with about 1.8 million protein structures and 52.4 million captions. These captions describe both global properties and residue-specific properties. Some describe fold class or compactness; others describe things like disulfide bonds, catalytic sites, or protein-protein contacts, including the relevant residue positions when available. To build those captions, the authors draw from experimental structures in the Protein Data Bank, predicted structures from the AlphaFold Database, and synthetic dimer data. They then transform database annotations and structure-derived measurements into natural-language descriptions. Importantly, they vary the caption content and length so the model learns that a prompt can specify one property or several at once. In total, they generate an average of 29 captions per entry. They also check the quality of these generated captions with an LLM judge on a sample of 80, and report that the captions match the underlying annotations closely, with high recall, precision, and accuracy. The model itself builds on a pretrained all-atom structure generator called Complexa. Prompteina adds two ways to condition on text. The first is global conditioning: a pooled embedding of the prompt is broadcast to every residue, so the entire protein can be steered toward a global property like fold class or overall compactness. The second is residue-level conditioning: individual residues attend to the token states of the prompt, so the model can connect specific words or phrases to specific positions in the protein. This is the crucial part for things like disulfide bonds, catalytic residues, or interface contacts. The architecture is trained in stages. First, the new text-conditioning components are adapted while the denoiser is frozen. Then the whole model is jointly trained on the large text-structure dataset. After that, the authors add an auxiliary alignment objective. This is a form of supervision that tries to align text spans with the residues they refer to. For example, if a caption mentions a disulfide bond, the model can be taught which residues in the structure correspond to that phrase. That idea is similar in spirit to token-level supervision in text-to-image models, where a caption word can be aligned to a region of the image. Here, the alignment head helps the model learn finer residue-level control. At inference time, Prompteina takes a prompt and a desired protein length. For binder design, it can also take a target structure and optional target contact residues. It then samples backbone coordinates and side-chain latent variables, and a frozen decoder turns those into an all-atom structure and amino-acid sequence. The system also uses classifier-free guidance to control how strongly the text steers generation, which is a standard and useful trick in generative modeling. The experiments are organized around three big questions: can the model follow structural prompts, can it control binder design, and does that help in real prospective experiments? First, they test structural control across 18 tasks in five groups: fold and topology, secondary structure, tertiary structure, global shape and surface properties, and combinations of fold class with secondary-structure or disulfide counts. They compare Prompteina against six baselines, including text-guided models and property-specific models that use more specialized inputs. The important point is that each baseline is only evaluated on the tasks its conditioning interface can express, so the comparison is fair in that sense. The result is that Prompteina outperforms the baselines on 15 of the 18 tasks, and it supports more properties than any single baseline. For categorical tasks, they measure accuracy; for count-valued tasks, they use Spearman correlation between the requested and realized values. The broader takeaway is not just that Prompteina wins on a few benchmarks, but that a single natural-language interface can span a wider design space than the usual task-specific controls. They also test whether the generated proteins actually match the intended function descriptions better. The paper says Prompteina generates proteins that more closely match descriptions of their intended functions, which suggests the language conditioning is not only steering geometry but also improving semantic alignment between prompt and output. The second major set of experiments is binder design. Here the model is used to steer the binder-target interface directly. Across 20 targets, Prompteina can control the number of residues at the interface with extremely strong agreement to the requested values, with correlation above 0.99. That is a striking result because it shows the text interface can reach down to a very local structural property, not just global fold categories. Then the authors move to agent-guided binder design. A language-model agent writes the design specifications based on whatever each model can accept. Prompteina outperforms every baseline on at least 13 of the 20 targets, and it improves interface confidence by 14% over the strongest baseline. Interface confidence here is a model-based measure of how plausible the designed binding interaction looks. So this result suggests that natural-language conditioning is not merely convenient; it can help generate better candidate binders. They also test iterative refinement, where the agent revises designs over multiple rounds rather than generating everything independently in one shot. Prompteina benefits especially from this setup. The gain from iterative design over non-iterative generation is 57% larger than the next-best baseline on the hardest targets. That matters because difficult targets are exactly where a flexible, language-driven interface should help most: the agent can inspect feedback, adjust the prompt, and ask for a revised design without translating everything into a new specialized representation. Finally, the paper includes a prospective experimental validation against GM2A, a therapeutic target with no reported de novo binders. This is a strong proof-of-concept because it is not just retrospective scoring. They run a design campaign, assay 89 designs by surface plasmon resonance, and identify nine binders. The strongest has an apparent dissociation constant of 107 nM across three replicates. For de novo protein design, that is a meaningful experimental result: it shows the model can contribute to real binder discovery on a target that had not previously yielded de novo binders. What should we take away from all this? The main claim is that natural language can serve as a powerful conditioning interface for protein structure generation. The value is not just that humans can speak to the model more easily. It is that language provides a unified way to express many kinds of constraints, from coarse global properties to residue-level interactions, and it works well enough to support both computational benchmarks and a real wet-lab campaign. There are still limitations to keep in mind. The text-structure dataset is large, but it is still assembled from existing structures and annotations, so the model inherits whatever biases and coverage gaps those sources contain. The caption alignment is useful, but it is only available for many annotations, not necessarily all. The model also depends on a pretrained structure generator, so its capabilities are built on that foundation. And while the results are strong across the reported tasks, protein design remains a hard problem: success on one target or one benchmark does not mean the system will generalize equally well to every functional class or every biological context. Still, the direction is important. This work shows that if you give a protein generator enough paired text and structure data, and you train it to associate words with residues as well as with whole-protein properties, then language can become a practical design interface. That opens the door to more expressive protein design systems, and to agentic workflows where a language model can reason about a target, write a prompt, refine it, and hand it directly to the structure generator. In that sense, Prompteina is not just another protein model. It is an argument that natural language can become a general-purpose control layer for protein engineering.