GIRAF is an in-development, folding-model-guided hallucination framework for de novo protein design. It can be used to design scaffold-based binders, including VHHs, scFvs, affibodies, and DARPins.

GIRAF takes a target sequence and a chosen protein scaffold as input, keeps the framework fixed, and optimizes only the designable regions, for example the CDRs, by iteratively co-folding the target with the candidate binder and updating the sequence using gradients from the folding model. In this way, the method directly optimizes model-based confidence metrics such as iPTM and pLDDT. In practice, we use protenix-v1 and AF3.

A key feature of GIRAF is that it relies on a single folding model and does not require a separate inverse folding model. We believe that combining a suitable scaffold, for example an FDA-approved Fv framework, with folding-model-based interface hallucination is a good strategy for designing desirable proteins. .

For final wet-lab selection, we score candidates extensively for developability using AbNatiV, CamSol, NanoMelt, and PyRosetta, and select those with a strong overall balance of properties.