Generation of Conformational Ensembles

An ensemble of 50 structural models was created using ESMFlow in order to take protein flexibility into account and capture a representative landscape of conformational states. Compared to a single static model, this method was used to better characterize the target protein's structural heterogeneity, offering a more thorough foundation for later binding site identification.

Identification and Analysis of Binding Sites

To find conserved functional interfaces, binding site prediction was carried out throughout the generated conformational ensemble (ScanNet). A geometric deep learning framework that combined chemical and spatial characteristics from each ensemble member was used to characterize these sites. To make sure the chosen target site was reliable and structurally well-defined, additional post-prediction analysis was carried out to evaluate the site's druggability and consensus across the different conformations.

Design of Binders

An all-atom generative diffusion model (RFDiffusion3) was used to design de novo protein binders. The main architectural constraint was the binding site that was established in the preceding step. The model was trained to create new binder backbones that were optimized for high-affinity interactions with the target site by using the geometric and chemical context of the identified interface.

Computational Assessment and Verification

To determine their structural stability and anticipated binding affinity, the generated binder candidates underwent a thorough computational evaluation. The thermodynamic consistency of the designed protein-protein interfaces was evaluated using the Boltz-2 biophysical scoring framework, which is based on deep learning. Candidates were only chosen for additional consideration if they showed favorable binding energy metrics and optimal folding stability.