A central challenge in de novo protein design is generating proteins with desired structural and functional properties. Existing design models typically require a dedicated input representation for each property, limiting the design objectives they support. Lee and colleagues develop Prompteina, the first all-atom structure generation model that supports new and broader design objectives specified in natural language, analogous to text-to-image generation. To train Prompteina, they construct Prompteina-Data, a paired text–protein structure dataset of 1.8M structures and 50M+ captions describing global and residue-specific structural and functional properties. They adapt Proteina-Complexa, a pretrained structure generation model, through text-conditioned pretraining and finetuning that aligns residues with their corresponding text. Prompteina outperforms baselines on 15 of 18 structural properties and generates proteins that more closely match descriptions of their intended functions. In binder design, it leads baselines on at least 13 of 20 targets and improves interface confidence by 14% over the strongest baseline. With iterative in silico feedback, Prompteina's gain over non-iterative generation is 57% larger than the next-best baseline on the hardest targets. Finally, they experimentally validate Prompteina designs against GM2A, a target with no reported de novo binders. They identify nine binders, with the strongest at an apparent KD of 107 nM. These results establish natural language as an effective and powerful conditioning interface for protein structure generation.

Prompteina-Data: Large-Scale Text–Structure Pairs

Training a text-conditioned protein structure generator requires paired data describing both global and residue-specific properties. Existing text–protein datasets pair sequences with only high-level functional descriptions. The authors construct Prompteina-Data from experimental structures in the Protein Data Bank, predicted structures in the AlphaFold Database, and Teddymer synthetic dimers. Each chain is treated as a separate entry, giving about 1.8M entries. The dataset includes up to 59 annotations per entry, derived from database fields and atomic coordinates. Annotations span global properties including fold class and compactness, and residue-specific properties with residue positions such as disulfide bonds, catalytic sites, and protein–protein contacts.

At inference, prompts may specify one property or combine several, so the authors vary caption content and length in the training data. For each entry, they sample several subsets that differ in the number and choice of annotations, then prompt Qwen3-235B-A22B to write one caption from each subset. In total, they generate 52.4M captions, averaging 29 captions per entry. An LLM judge compares generated captions with ground-truth annotations, achieving 97.8% recall, 89.8% precision, and 95.7% accuracy on a validation sample.

Text-Conditioned Structure Generation Architecture

Prompteina builds on the pretrained Complexa model, adding two text-conditioning pathways and an auxiliary text–residue alignment objective. Complexa's denoiser generates backbone coordinates and one side-chain latent per residue before a frozen variational autoencoder decodes both into an all-atom structure and amino acid sequence.

For global properties, a pooled embedding from frozen Qwen3-Embedding-8B is broadcast through adaptive layer normalization, giving every residue the same conditioning signal. For residue-specific properties, each residue attends to frozen Qwen3-8B token states via cross-attention before protein self-attention in each denoiser layer. Two-layer MLPs map both representations to denoiser dimensions. The combination of pooled conditioning and cross-attention performs best across tasks.

Text–Residue Alignment

In text-conditioned pretraining, the denoiser is trained using a flow-matching loss without direct supervision of which residues correspond to each text span. Prompteina-Data contains residue positions for many annotations, which the authors use to construct text–residue alignment annotations. For each annotation selected for a caption, its associated residue positions form the annotation's residue span. They locate corresponding text spans using regular expressions and TF-IDF-weighted token matching.

An auxiliary alignment head is trained on these annotations. For N protein residues and T caption tokens, alignment annotations are represented as a binary matrix M where M[n,i] = 1 when token i is aligned to residue n. The alignment head predicts a logit A[n,i] using cross-attention updates for every residue–token pair. Aligned residue–token pairs are sparse, so a positive weight w+ is set to the ratio of negative to positive pairs. The alignment loss uses a weighted binary cross-entropy formulation. After alignment fine-tuning, Prompteina gains finer control over features such as disulfide-bond and salt-bridge counts while retaining global fold control without the alignment head at inference.

Training and Inference

The authors first adapt the Complexa denoiser to natural-language conditioning using newly initialized parameters for pooled and token-level text conditioning. These parameters are first trained with the denoiser frozen before unfreezing it for joint text-conditioned continual pretraining on Prompteina-Data, following staged training approaches from text-to-image models. For alignment fine-tuning, the denoiser is frozen while text projectors, cross-attention layers, and alignment head are updated.

At inference, Prompteina takes a prompt and protein length and, for binder design, a target structure and optional target contact residues. Prompteina samples backbone coordinates and side-chain latents with the same stochastic differential equation as Complexa and uses classifier-free guidance to adjust text-conditioning strength. The frozen variational autoencoder decodes these into an all-atom structure and amino acid sequence.

Structural Property Control

The authors test Prompteina on prompts for structural properties across five groups: fold, shape, secondary structure, tertiary structure, and combined fold with secondary or tertiary properties. Across 18 tasks, Prompteina leads evaluated text-guided baselines on 15 tasks and supports more properties than any single baseline. For property-specific models using dedicated labels or structured inputs, Prompteina achieves competitive results on CATH class (96.0% accuracy), CATH architecture (51.1%), compactness (76.7%), and secondary structure content (helix 0.94, strand 0.88, sheet 0.64 Spearman correlation). Prompteina also generates structures that more closely match descriptions of intended functions, evaluated using ProTrek similarity against functional specifications from diverse protein families.

Binder–Target Interface Steering

At finer structural scales, Prompteina steers the number of residues at the binder–target interface across 20 targets with Spearman correlation exceeding 0.99. The authors evaluate agent-guided protein binder design where the same language-model agent writes specifications based on the properties supported by each model's conditioning interface. Prompteina outperforms every baseline on at least 13 of 20 targets and raises interface confidence by 14% over the strongest baseline. When interface properties are explicitly specified, Prompteina achieves higher binding confidence than baselines that lack this capability. The language interface preserves rich design intent for generation, allowing the agent to express complex multi-property specifications that fixed input representations cannot capture.

Iterative Binder Prompt Optimization

The authors allow an agent to iteratively refine designs using in silico feedback from structure prediction models. Prompteina shows larger gains over matched-budget i.i.d. generation than either baseline, especially on the most difficult targets. On the hardest targets, Prompteina's gain over non-iterative generation is 57% larger than the next-best baseline. This demonstrates that Prompteina's language interface enables effective iterative refinement, integrating language-based reasoning into the protein design loop.

Experimental Validation Against GM2A

As a proof of concept, the authors conducted a prospective, agent-guided design campaign against GM2 ganglioside activator protein (GM2A), a therapeutic target implicated in GM2 gangliosidosis. GM2A has no reported de novo binders and was one of three targets without a hit in a prior campaign. In their campaign, nine of 89 assayed designs were classified as binders by surface plasmon resonance, with the strongest yielding an overall apparent KD of 107 nM across three replicates. The best single replicate achieved 68.8 nM. This experimental validation confirms that Prompteina's natural-language interface produces functional binders for challenging targets.

Conclusion

Prompteina establishes natural language as an effective conditioning interface for protein structure generation, enabling structural and functional control across scales and integrating language-based reasoning and iterative refinement into protein design. The model's ability to jointly generate all-atom structures and sequences directly from free-form text, combined with residue-level alignment for fine-grained control, represents a significant step toward unified, flexible protein design. Code, data, and model weights are available on GitHub.