Refinement of crystal structures at ultralow resolution with assistance from AlphaFold modeling and Rosetta optimization

Abstract

Crystals of large macromolecular complexes often diffract quite poorly, typically having high solvent content, relatively feeb le lattice contacts, quite weak subunit associations, and somewhat fl exible interdomain linkages. Although resolution may be limited to dmin > 7-8 Å, the diffraction amplitudes should suffice, in principle, to specify conformational torsion angles; however, at such ultralow resolution, realizing and maintaining a suitable model within the radius of refinement convergence is a challenge. Important insights into biological processes may be obtained, but only if structural validity can be assured. Having successfully refined a four - copy structure of Hsp70 DnaK in the S-state at 7.7 Å resolution as rigid bodies (Wang et al. , Mol. Cell 81 , 3919, 2021), we set out to refine a crystal structure of ryanodine receptor RyR1 at 8.0 Å resolution by having multiple quasi - rigid bodies to comprise the 5037 residues in each protomer of the RyR1 -tetramer as complexed with calstabin. After molecular replacement from a 65% -complete cryo-EM model at 3.6 Å resolution (des Georges et al. , Cell 167 , 145, 2016), the structure was refined from a single rigid-body (R free = 0.53), through five linked rigid bodies (R free = 0.47), and fi nally as 18 linked domains (R free = 0.43) identified in the cryo-EM analysis and then sub-divided as dictated by (F o - F c ) difference map and the R free analysis. We then turned to AlphaFold, presuming that the process had stalled due to incompleteness and uncertainty in the initial model. Trials showed that AlphaFold - predicted domains reduced Rfree when fitted into crystal density. We then systematically identified such AlphaFold - modeled domains and obtained substantial improvement (Rfree = 0.38). Further improvement followed after Rosetta refinement using tight restraints in the phenix.rosetta_refine module (R free