GST Post Molecular Glue
Unlike binder design against single proteins, de novo design of protein binders targeting a protein·small molecule complex remains a largely unsolved challenge—primarily as accurately capturing the presence and influence of the ligand in the binding process remains one of the largest hurdles to overcome. Current all-atom generative models can be conditioned on other types of ligands, yielding potent binders for small molecules, metal ions, and DNA (https://www.biorxiv.org/content/10.1101/2025.04.06.647261v1,https://github.com/yehlincho/BoltzDesign1,https://www.science.org/doi/10.1126/science.adl2528 https://github.com/baker-laboratory/rf_diffusion_all_atom). Redesign methods (meaning those that take as input a structure and output a sequence that should consistently conform to the original fold) such as LigandMPNN can effectively capture the small molecule context - a recent paper showed it can redesign binders with X better binding affinity and switch-like behaviour in the presence of a small molecule (find 1-2 LigandMPNN papers). We combined gave the ligand-unaware BindCraft ligand context by first selecting potential molecular glue hotspots with MaSIF-Neosurf, then redesigning the initial BindCraft backbones with LigandMPNN.
Case Study 2: Ethacrynic acid and GST-P1
For our second application, we aimed to show a system based on the binding of a ligand on the surface of a protein and the associated change in binding interface. For this we explored conditional binding using the well-known protein Glutathione S-transferase P1 (GST-P1), often used as an affinity tag, and its inhibitor ethacrynic acid, aiming to design proteins that bind GST-P1 only in presence of ethacrynic acid.
Design rationale: We leveraged MaSIF-neosurf software to identify GST-P1 surface regions whose binding potential changed significantly upon ethacrynic acid binding.
AI-enhanced sequence design: Using BindCraft and LigandMPNN, we developed multiple sequences specifically tailored to recognize and interact with these new molecular "hotspots."
Validation results: Experimental results via BLI highlighted several designs with precise nanomolar binding affinity occurring only in the presence of ethacrynic acid. These results confirmed the efficacy of our computational pipeline in creating accurate conditional interactions.
Design Strategy :
Given the lack of conformational change a different approach was used to identify hotspots for subsequent binding design. The MaSIF-neosurf pipeline (https://github.com/LPDI-EPFL/masif-neosurf, https://www.nature.com/articles/s41586-024-08435-4) was used to determine the predicted binding interface both in presence and in absence of ethacrynic acid. The difference between both values was used as a score to determine the optimal binding hotspots. A visual representation of the hotspot identification is shown below :
Computational generation :
We first used our in-house Rosetta-free BindCraft with the Default settings and filters, again only ranging the inter-chain contacts weights (0.1, 0.5, 1.0). Once we had a set of 50 promising “in silico binders”, we pushed them through our second pipeline: we first folded the ternary complex (binder-ligand-target) with Boltz-1 (default parameters), then redesigned the binder with LigandMPNN. We now have the full ligand context! To actually select for binders in a glue-like complex, we developed a custom score weighting the ligand-protein and ligand-target interactions more positively than the protein-target ones. This score accounts for the number of contacts, hydrogen bonds, salt bridges, all weighted by the interaction confidence from Boltz-1 (ipTM).
The same BLI based approach as described previously was used to screen the designed variants. This allowed to identify a design with a K_d of 238nM in presence of ethacrynic acid while showing negligible interaction in absence of it.
