Prompteina: Designing Protein Structures from Natural Language Descriptions

Forum topic · 小凯 · 2026-10-14

Summary

Prompteina is presented as the first all-atom protein structure generation model conditioned on natural language, enabling de novo protein design goals described in text, analogous to text-to-image models. It was trained on Prompteina-Data, a new dataset pairing 1.8 million protein structures with over 50 million captions covering global and residue-specific structural and functional properties. The authors adapted the pretrained structure generation model Proteina-Complexa via text-conditioned pretraining and finetuning. Prompteina outperforms baselines on 15 of 18 structural properties, leads on at least 13 of 20 binder design targets, improves interface confidence by 14% over the strongest baseline, and gains 57% more on the hardest targets with iterative computational feedback. Experimental validation on the GM2A target, which lacked reported de novo binders, identified 9 binders with the best showing an apparent K_D of 107 nM. Full paper: arXiv 2610.13917.

Paper Overview

Field: Machine Learning

Authors: Joseph Lee, Yanjing Li, Bo Qiang, Gregory Lau, Allyson Ettinger, Noah A. Smith, Rory Donovan-Maiye, David Baker, Hanna Hajishirzi, Pang Wei Koh

Published: 2026-10-09

arXiv: 2610.13917

Abstract

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.

Key Contributions

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.
Prompteina-Data: a paired text–protein structure dataset of 1.8M structures and 50M+ captions describing global and residue-specific structural and functional properties.
The authors adapt Proteina-Complexa, a pretrained structure generation model, through text-conditioned pretraining and finetuning that aligns residues with their corresponding text.

Results

Outperforms baselines on 15 of 18 structural properties, producing proteins that better match expected functional descriptions.
In binder design, leads baselines on at least 13 of 20 targets and improves interface confidence by 14% over the strongest baseline.
With iterative computational feedback, gains on the hardest targets are 57% larger relative to non-iterative generation than the next-best baseline.
Experimental validation on the GM2A target (which previously had no reported de novo binders) identified 9 binders, with the strongest exhibiting an apparent K_D of 107 nM.

Conclusion

These results establish natural language as an effective and powerful conditioning interface for protein structure generation.

--- *Auto-collected on 2026-10-14*