Latent generative search unlocks de novo design of untapped biomolecular interactions at scale

Introduces latent generative search for binder design, using reward-guided search at inference time to steer the Proteina-Complexa generative model, which codesigns sequence and structure together in a continuous latent space. Compared with current methods that rely on an inverse-folding step, this framework merges sequence and structure generation into a single latent-space process and removes that step. Supported by the framework description and the subsequent large-scale screening results, representing a core methodological advance.

In a screen of more than one million designs by multiplexed phage display, the method produced more validated binders than every other method tested, and its codesigned sequences surpassed post hoc redesign. This is a validation result obtained by direct comparison with multiple methods in the same screening system, rather than computational prediction alone. Evidence comes from multiplexed phage display screening of more than one million designs, a large-scale experimental validation.

Delivered high-affinity binders across therapeutic receptors, a viral attachment protein and intracellular signalling targets. Indicates the method is not limited to a single target class but covers multiple target types. Stated as the achievement of high-affinity binders, constituting experimental validation.

Generated the first de novo proteins that bind a free carbohydrate, including one that discriminates between blood-group antigens. Opens a new designable space for polar, solvent-exposed epitopes and small, flexible ligands, including carbohydrates, a target class that had largely resisted de novo binders. Presented as a first-of-its-kind result with a concrete example discriminating blood-group antigens, constituting experimental validation.

The results are intended for research and development settings that require binders against polar, solvent-exposed epitopes and small, flexible ligands including carbohydrates, and apply to target classes such as therapeutic receptors, a viral attachment protein, intracellular signalling targets and a free carbohydrate; their value lies in providing a designable route for targets that had largely resisted de novo binders and in laying groundwork for further validation and development across more target classes.

Readers may still watch how reward-guided search performs across different target classes, how broadly the advantage of codesigned sequences applies, and how generalizable free-carbohydrate-binding proteins are to a wider range of glycan targets; in addition, the current text is summary-level and lacks figures and supplementary material, so assessing specific design success rates and affinity values would require the full paper data.