AlphaFold2 Model Refinement Using Structure Decoys

AlphaFold2-predicted protein structures inform many modeling techniques in structural biology, including the interpretation of cryogenic electron microscopy (cryo-EM) maps. However, the accuracy of the AlphaFold2 prediction, the quality of the experimental cryo-EM data, and the reliability of the model's alignment with density may affect the accuracy of the interpretation. In this work, we explored a new refinement approach by generating unbiased structural decoys from the AlphaFold2 model via 3DRobot or elastic network model (ENM)-based ModeHunter. Our hope was that some of the decoys would resemble the true structure, and consequently, that the refinement problem could then be reduced to selecting the correct decoy from the decoy set that most closely resembles the experimental cryo-EM map. We explored a map/model pair from the structure of a lipid-preserved respiratory supercomplex, where AlphaFold2 previously struggled (TM-score: 0.52). In this specific case, we observed that the inherent bias of 3DRobot toward compact decoys limited the AlphaFold2 model enhancement (best decoy TM-score: 0.53), whereas ENM is capable of producing extended decoys that significantly improve the accuracy of the AlphaFold2 model (best decoy TM-score: 0.68).