Prompteina: all-atom protein structure generation from natural language prompts

Conditioning on paired text and 1.8M structures yields 15 of 18 property wins, higher binding confidence on 13 of 20 targets, and a 107 nM wet-lab binder.

Primary source: arXiv:2610.13917: https://arxiv.org/abs/2610.13917

What’s new: De novo protein design models usually rely on specialized input representations for each targeted property, from overall fold geometries to residue-level catalytic sites. Designing Protein Structures from Natural Language (arXiv:2610.13917, 9 Oct 2026) introduces Prompteina, an all-atom structure generation model that accepts multi-scale design objectives specified entirely in natural language. The authors assemble Prompteina-Data, a corpus of 1.8 million structures and over 50 million synthesized captions spanning global architecture and residue-specific functional annotations. Prompteina augments the pretrained Complexa denoiser with both global text conditioning and a token-to-residue cross-attention head supervised by an auxiliary text-residue alignment objective.

In evaluations across five structural property groups, Prompteina outperforms prior specialized baselines on 15 of 18 tasks. For binder design across 20 diverse targets, it leads baselines on 13 targets while raising interface confidence metrics by 14%. When paired with iterative in silico feedback, its improvement over single-pass generation exceeds the next-best baseline by 57% on the hardest design targets. Crucially, the authors validate designs experimentally in the wet lab against GM2A, a challenging target lacking reported de novo binders, isolating nine functional binders with the top candidate achieving an apparent binding affinity ($K_D$) of 107 nM.

Why it matters: Natural language provides a flexible interface that unifies global topological targets, catalytic geometry, and interface contacts into a single prompting paradigm. Moving from bespoke geometric descriptors to text-conditioned flow matching enables researchers to integrate automated LLM reasoning and iterative refinement directly into structural biology workflows, backed by physical binding validation.

Caveats: Generating structures one chain at a time limits native co-folding of complex multimeric assemblies. The training annotations rely heavily on synthetic captions derived from LLM interpretation of structural databases, which may propagate descriptive biases. Apparent binding affinities were confirmed on a single target protein family (GM2A), so wet-lab generalizability across broader therapeutic classes remains to be shown.