[cs.LG] A central challenge in de novo protein design is generating proteins with desired structural and functional properties.
Editor’s note: As we expand outward from Earth to other worlds we are almost certainly going to encounter things we did not expect to find – things that are unlikely or impossible on Earth. Life on other worlds may arise from a totally different set of chemical pathways than was the case on Earth. Or it may follow a very similar path. Or both. How do we estimate what could exist such that we are better prepared to search for the unexpected?
The genetic code of all life on our home world (with a few minor exceptions) is based on a genomic code consisting of 4 standard nucleotides – and the universe of proteins that it can code for. It is possible that Earth life may have once used a different assortment of letters in its instructional alphabet but the evidence suggests that the current code has been in use for quite some time. But the genomics of life on other worlds might be different and use a different genetic alphabet.
As we tinker with earthly genomics for commercial and health reasons we are discovering novel ways to tweak the standard genomic model to alter the outcome of a genetic sequence and its end products. Whether the new sequences developed in synthetic biology (SynBio) will work is another matter. But it does give us insight into how genetic and biochemical sequences work (and do not work) here and, by extension, how they might work elsewhere.
The use of natural language design models certainly offers to enhance the speed and output of such research.
Existing design models typically require a dedicated input representation for each property, limiting the design objectives they support. We 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, we 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. We 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, we experimentally validate Prompteina designs against GM2A, a target with no reported de novo binders.
We 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.
Paired text–structure data and text-conditioned structure generation. (a) Structural and functional annotations are curated and (b) passed to a large language model for caption generation. (c) Residues are aligned to their corresponding caption text spans for supervised alignment finetuning. (d) Captions are embedded into pooled and per-token text representations, then projected to condition the structure generation model. (e) The denoiser is first trained on caption–structure pairs with a flow-matching objective, then fine-tuned with supervised text–residue alignment (§2.2).– [cs.LG]
Joseph Lee, Yanjing Li, Bo Qiang, Gregory Lau, Allyson Ettinger, Noah A. Smith, Rory Donovan-Maiye, David Baker, Hanna Hajishirzi, Pang Wei Koh
Comments: 53 pages, 9 figures. Code, data, and model weights: this https URL
Subjects: Machine Learning (cs.LG); Biomolecules (q-bio.BM)
Cite as: arXiv:2610.13917 [cs.LG] (or arXiv:2610.13917v1 [cs.LG] for this version)
https://doi.org/10.48550/arXiv.2610.13917
Focus to learn more
Submission history
From: Joseph Lee
[v1] Fri, 9 Oct 2026 17:51:55 UTC (16,987 KB)
https://arxiv.org/abs/2610.13917
Astrobiology,
