Generalization and Usability of Co-Folded GPCR–Ligand Complexes: A Physics-Guided Assessment

Deep learning co-folding models for end-to-end protein–ligand structure prediction mark a major advance beyond AlphaFold2, yet their reliability for decision-making in drug discovery remains unclear. Here, we benchmark Boltz, a state-of-the-art co-folding model, using a curated set of ligand-bound human G protein-coupled receptors (GPCRs) from families unseen during training. We find that the receptor backbones are generally predicted with reasonable accuracy, but ligand poses often deviate significantly from experimental structures. We then evaluate physics-based refinement with rigid-receptor (Glide) and induced-fit docking (IFD-MD) methods, which recover more than half of the misplaced ligands to near-experimental accuracy. As conventional evaluations for co-folded structures focus on distance-based metrics such as root-mean-squared deviation (RMSD), which can miss subtle but consequential binding-site errors, we carry out a further assessment of Boltz performance using free-energy perturbation (FEP+), which is both accurate and sensitive to starting-structure quality, on curated congeneric ligand series with known binding affinities that target the GPCRs. A significant fraction of the 14 congeneric series tested in this fashion fail to reproduce experimental binding affinities via FEP+ when employing the Boltz generated complex, even when the binding-site RMSD is low in some cases. IFD-MD rescues these failures and restores retrospective FEP signals to native-like level for all of these series. Together, these results delineate current generalization and usability limits of co-folded GPCR–ligand complexes and motivate a workflow that pairs deep learning predictions with physics-based refinement and validation before high-stakes decisions in drug discovery.