Design Strategy

Our approach aimed to generate a flexible, multi-site binder targeting two distinct interaction hotspots on RBX1.

First, we used RFDiffusion to independently generate short α-helical binding elements positioned at each of the two target hotspots. These helices were then inverse folded using solubleMPNN and structurally evaluated using AlphaFold3.

To ensure high-quality candidates, we applied filtering based on predicted interface confidence, retaining designs with an interface predicted TM-score (ipTM) above 0.7 and an interface predicted alignment error (ipAE) below 10 Å.

Next, the selected helical binders were combined into a single construct by linking them with flexible regions using ProteinGenerator. During this step, we biased the linker sequences toward disordered loop-like conformations to promote flexibility and enable cooperative engagement of both binding sites.

Finally, the full-length constructs were re-evaluated using AlphaFold3, and designs were ranked based on interface quality, with priority given to those showing low interface predicted alignment error (ipAE), indicating confident and well-defined binding.