One language interface steers protein structure and binder design

Can one protein generator let designers express both broad structural goals and residue-level contact details without inventing a new input format for each property?

The question matters because agents that research targets and propose designs can only communicate those decisions through the controls a generator accepts.

Prompteina pairs a large text–structure dataset with a pretrained all-atom generator so natural-language prompts control whole-protein and residue-level design properties.

The researchers built Prompteina-Data from about 1.8 million protein-chain entries and 52.4 million captions, then adapted the pretrained Complexa all-atom generator. Pooled text conditioning carries global properties, while token-level cross-attention and text–residue alignment connect phrases with particular residues.

The evaluation compared Prompteina with six baselines across 18 structural tasks and tested binder design on 20 targets.

Prompteina led on 15 structural tasks and tracked requested binder interface-residue counts closely across all 20 targets. In the main binder comparison, it had the highest predicted interface-confidence score in all three sequence conditions, including after sequence redesign.

Together, these results support the narrower conclusion that one language interface can express and use diverse structural controls in these tests.

The structural and binder comparisons rely partly on computational proxies: interface-residue count is not binding affinity, and predicted interface confidence is not experimental binding.

The binder-score calibration is not fully independent and has limited target and negative-example coverage, so these results do not establish general hit rates or broad protein function.