Fig. 1: Binding site mutagenesis challenges against co-folding models using the CDK2 system (PDB: 1B38).

Predicted binding-site residues are shown as cyan sticks, predicted ligand poses are shown as green sticks, and the original co-crystallized ligand pose is shown as gray sticks. The first row shows each model’s prediction for the wild-type protein-ligand system prior to any modification. The remaining rows show different adversarial challenges where all binding site residues are mutated. In binding site removal, all residues are mutated to glycines effectively removing all ligand-side-chain interactions from the original system. The packing challenge mutates all residues to phenylalanine, removing all native interactions with side-chains and further occupying the pocket with bulky, hydrophobic groups. In the inversion challenge, binding site residues are mutated to residues with dissimilar properties. These mutations should annihilate the binding site and remove the majority of native protein-ligands interactions necessary for binding. However, in many cases the ligand is still predicted within the binding site and can adopt a low RMSD pose, indicating that these co-folding models are not predicting poses based on physics of interactions, but rather learning patterns in global protein structures and sequences.