Reproduce “Designing Protein Structures from Natural Language” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Joseph Lee, Yanjing Li, Bo Qiang, Gregory Lau, Allyson Ettinger, Noah A. Smith, Rory Donovan-Maiye, David Baker and 2 more
Prompteina lets designers specify global and residue-level protein goals in natural language, then generates structure and sequence together. Its text pathways combine shared context with residue-level control. In a GM2A campaign, 9 of 89 tested designs were classified as binders; the strongest had an apparent dissociation constant of 107 nM across three replicates, validating binding for selected designs against one target.
Key insights
- A shared natural-language interface can express both broad structural properties and localized constraints, enabling combinations that otherwise require distinct conditioning formats.
- For binder design, specifying interface properties such as interface size and buried surface area contributed more than adding fold and shape descriptions alone.
- Agent-generated specifications can be revised using computational feedback; Prompteina showed larger gains than comparison methods on the hardest target subset, though predicted confidence scores do not establish experimental binding.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.