See real binder design runs, then run yours

We used BoltzGen to design VHH nanobodies against a lymphocyte surface receptor, steering them toward a membrane distal surface so they land away from the receptor's functional contact face and do not interfere with native signaling.

The run generated 2000 designs. The best scored ipTM 0.974, and one recurring epitope emerged across 32 independent designs. We narrowed to a final panel of 12 and confirmed them with AlphaFold2 and Boltz-2, two structure predictors independent of the design step; ten of the twelve held up strongly under both. The lead nanobodies sit 11 to 18 angstroms clear of the functional face, so they bind the receptor without blocking it.

We used BoltzGen to design de novo minibinders aimed at one partner interface on a scaffolding protein, so a binder competes with the natural partner for the same surface.

The campaign produced 20,000 unique designs. The top ranked shortlist scored ipTM 0.93 to 0.98 with interface PAE well under one angstrom, on binders 65 to 80 residues long, and the best designs covered up to about 93 percent of the target epitope. The value is a deep ranked pool to sample from rather than a single hero design.

We used RFdiffusion to generate de novo binder backbones against a target, building each fold from scratch through diffusion rather than grafting a binder onto an existing scaffold.

The run produced 60 de novo backbones across five parallel jobs, each an independently generated scaffold positioned against the target surface. RFdiffusion returns backbone geometry, the starting fold you then take into sequence design and an independent structure predictor for scoring. The value is a pool of de novo binder scaffolds to carry forward rather than a single backbone.