Can protein structures be designed directly from natural language descriptions?

One text interface replaces per-property inputs

Most protein structure generators control design through a dedicated input format for each property: a categorical label for fold, a per-residue string for secondary structure, a bond matrix for disulfides. Each new objective needs its own representation, which caps how many objectives one model can serve. Prompteina's premise is that natural language can act as a single interface for all of them, the way text-to-image models express many visual attributes without task-specific controls.

The model does not merely use text as a filter at inference time. It conditions the denoising process directly on text embeddings, generating backbone coordinates, side-chain latents, and amino acid sequence together. That differs from earlier text-guided protein work that either steers sampling with a separate structure-to-caption model or generates sequences that must later be folded into structures. Because both modalities are conditioned on the same prompt, the paper argues, control extends to residue-level features rather than stopping at overall fold.

A 1.8M-structure caption dataset underpins control

Fine-grained text control requires text that describes fine-grained structure, and the authors found existing text–protein datasets paired sequences with only high-level functional descriptions. They built Prompteina-Data: 1.8M structures and 52.4M captions covering both functional and structural properties, drawn from PDB chains (611K entries, 19.5M captions), AlphaFold Database monomers (288.7K entries, 7.5M captions), and synthetic Teddymer dimers (903K entries, 25.4M captions).

Captions were generated from database annotations and properties computed directly from atomic coordinates, spanning ten groups such as fold and topology, secondary structure, and ligand-related features. Where possible, the dataset records which phrases in a caption correspond to which residues. That phrase-to-residue correspondence is what later lets the model learn that a specific phrase should control a specific part of the structure, rather than only its overall shape.

Staged training keeps fold accuracy while adding text

Prompteina adapts Complexa, a pretrained all-atom structure generator, rather than training from scratch. New text pathways — pooled conditioning through adaptive layer normalization and token-level conditioning through cross-attention — are first trained with the denoiser frozen. The denoiser is then unfrozen for text-conditioned continual pretraining on Prompteina-Data, and a final stage fine-tunes an alignment head that ties caption phrases to residues.

The alignment stage is deliberately narrow. The authors report that higher learning rates or unfrozen query and key projections degraded fold accuracy within one to two epochs, so they trained only the value and output projections in the last five cross-attention layers, matching all fourteen layers with one third as many trainable parameters. Cross-attention output is gated by a per-layer scalar initialized to zero, so text influence grows from nothing instead of disrupting the pretrained generator at the start.

Broader property coverage than baselines

On 18 structural properties, Prompteina outperformed baselines on 15 and generated proteins that more closely matched descriptions of their intended functions. The comparison matters because the baselines are strong and specialized: RFdiffusion3 and BoltzGen combine many controls in one model but give each property a distinct input and its own training annotations, while ESM3 uses dedicated token tracks with separate tokenization and noise schedules per property.

The paper's claim is not that text beats every specialized control on its home property. It is that one interface covers a broader range of objectives while matching or exceeding existing approaches on properties the baselines also support. That distinction is the practical one for a reader deciding whether to add a new design objective: with a per-property interface, a new objective means new annotations and a new input pathway; with text, it means new captions.

Binder gains concentrate in interface specification

In binder design, Prompteina led baselines on at least 13 of 20 targets and improved interface confidence by 14% over the strongest baseline. The measure here is ipSAE min, an interface confidence score computed on generated sequences, not an experimental binding measurement. Prompteina can accept interface size as a residue count and buried surface area, inputs that neither RFdiffusion3 nor BoltzGen accepts.

An ablation isolates where the gain comes from. Rerunning all 20 targets with successively richer prompts, the expected top-k ipSAE min was 0.362 with fold only, 0.364 when shape was added, and 0.455 with fold, shape, and interface properties. Shape specification added little; interface specification produced the largest gain. Within each target, larger realized buried surface area was associated with higher interface confidence, which is a correlation the authors use to motivate the ablation rather than a demonstrated cause.

Iterative feedback helps most on hard targets

The paper also tests a workflow in which a language-model agent reads design scores and writes new specifications for the next round. Across 20 targets, four rounds of 50 designs were compared against matched-budget independent generation. Prompteina benefited more from this loop than either baseline, and the advantage was largest on the hardest targets, where its gain over non-iterative generation was 57% larger than the next-best baseline's.

The proposed explanation is that a broader interface gives the agent more properties to revise. In the Claudin-1 example, the agent first explored alternative topologies and then requested a larger interface, improving the best design after each round. Paired with either baseline, the agent still found productive patterns, but each baseline's interface offered fewer properties to change. This is the authors' interpretation of the mechanism, supported by the pattern of gains rather than by a direct test of why they occur.

Nine GM2A binders, including a 107 nM hit

The experimental test targeted GM2A, a protein with no reported de novo binders and none in a prior 30-design campaign. The team ran iterative optimization with a 1,000-design budget, selected 90 designs for testing, and 89 were expressed and assayed by SPR. Nine were classified as binders, eight with fitted apparent KD values, and six sub-micromolar binders spanned diverse architectures — a 6.7% assay hit rate. The strongest binder had an overall apparent KD of 107 nM across three replicates.

These are apparent KD values from SPR, not equilibrium constants from a full thermodynamic analysis, and the campaign is described as a proof of concept against one target. The designs also realized the agent's requested architecture: a two-layer alpha/beta fold around GM2A residue W94, with five strands, a hydrophobic interface, and 21 interface residues versus 22 requested. That agreement between prompt and structure is the paper's clearest evidence that residue-level text specifications can be executed, though it is shown for one target rather than across the 20-target benchmark.

Where text conditioning earns its place

The paper's most decision-relevant comparison is not text versus specialized controls in general, but which design objectives benefit from a unified interface. On shared structural properties, Prompteina matches or exceeds baselines; its distinctive advantage appears when an objective has no existing input format, as with interface size and buried surface area, and when an agent needs to revise specifications across rounds. The ablation showing a large gain from interface properties and little from shape points to the same priority: specify the interface, not just the fold.

The evidence also sets the boundary. The 20-target binder benchmark uses ipSAE min, a computational confidence score, and the experimental validation covers one target with apparent KD values from SPR. The iterative-feedback advantage is measured in silico. Within those limits, the paper supports a specific judgment: natural language is most valuable as a conditioning interface when the design objective is residue-specific and when a human or agent will iterate on specifications using feedback, which is exactly where fixed per-property inputs run out of expressive room.

Sources used in this answer