Computational & AI Journal of Chemical Information and Modeling
Four leading cofolding models often place ligands well yet get the receptor's functional conformational state wrong, limiting their use for state-selective design
- Four current AI cofolding models were compared on kinase and Class A GPCR complexes, using structures deposited after every model's training cutoff so the test is genuinely out-of-sample.
- Ligand poses frequently looked accurate while state-defining features, such as kinase activation-loop and DFG/DYG arrangements or GPCR intracellular loops, were predicted incorrectly.
- Adding state-annotated templates or filtered alignments helped only in some systems, confidence scores tracked ligand placement but not pocket accuracy, and the one allosteric complex was missed by every model.
A low ligand RMSD was never proof the pocket is right, and this study shows why that matters for kinase and GPCR work. If your design hypothesis hinges on DFG-out or an inactive-state loop, treat a cofolded pose as a hypothesis, not a structure. Before trusting it, I'd compare against docking into a state-matched experimental structure and check whether the model just recalled the nearest training-set state. The allosteric miss is a single case, so I wouldn't generalize it.
Paper: Journal of Chemical Information and Modeling ↗ Benchmarking Protein–Ligand Cofolding Models: Correct Ligand Placement Is Decoupled From Accurate Protein Conformation