E0178: Reflections on the False Security of Current Refinement Protocols at Moderate Resolution

E0178

Reflections on the False Security of Current Refinement Protocols at Moderate Resolution. Charles W. Carter Jr., Department of Biochemistry and Biophysics, CB 7260 University of North Carolina at Chapel Hill, Chapel Hill, NC 27599-7260

The initial refinements of rubredoxin demonstrated that protein structures could be improved by adjusting atomic coordinates in accordance with indications generated from differences between observed and calculated structure factors. Soon after, the joint imposition of these crystallographic constraints and stereochemical constraints gave rise to a long and familiar development of increasingly automated refinement procedures. Although this workshop will focus primarily on the use of high resolution data for refinement, I want to explore the limits encountered with successive use of molecular replacement phases at moderate resolutions between 2.0-3.0Å. Using "borrowed" phases carries significant risks. Refinement of the probe structure has produced a carefully balanced set of atomic positions satisfying complicated non-linear constraints, and this solution is bound to persist when that structure is used to phase yet another structure. As noted by Bricogne and others, model bias persists because the amplitude constraints alone are imposed, forcing premature agreement of amplitudes with incorrect phases.

We have found repeatedly that refined molecular replacement models produce solutions with poorer statistics than models of the same protein refined from maps phased with experimental phases, despite the fact that these structures had significantly larger r.m.s. deviations from the target structure. In such cases, model-independent map refinement using Maximum Entropy Solvent Flattening with phase permutation and NCS averaging, when available, produce satisfactory maps at resolutions between 2.0Å and 3.0Å. Models built into these maps using real-space refinement procedures generally have R-factors around 30%. Subsequent refinements using several widely used programs failed to improve the R_free cross-validation statistic significantly, suggesting that the model independent phasing had essentially converged to the correct phases. Two maximum-likelihood refinement procedures, CNS (Brünger) and BUSTER/TNT (Bricogne), achieved modestly lower Free R-factors. In two cases, we have compared the resulting structures with those obtained independently by SIR phasing from selenomethionine substituted crystals, followed by the same density modification procedures. R.m.s. deviations on the order of 0.2Å suggest that the model-independent map refinements are highly effective at minimizing model bias. (Supported by NSF (MCB 9304674) and by NIH (GM 48519)).