Conformation-dependent backbone geometry restraints set a new standard for protein crystallographic refinement

Abstract

Ideal values of bond angles and lengths used as external restraints are crucial for the successful refinement of protein crystal structures at all but the highest of resolutions. The restraints in common use today have been designed on the assumption that each type of bond or angle has a single ideal value that is independent of context. However, recent work has shown that the ideal values are, in fact, sensitive to local conformation, and, as a first step towards using such information to build more accurate models, ultra-high-resolution protein crystal structures have been used to derive a conformation-dependent library (CDL) of restraints for the protein backbone [Berkholz et al. (2009) Structure 17, 1316-1325]. Here, we report the introduction of this CDL into the phenix package and the results of test refinements of thousands of structures across a wide range of resolutions. These tests show that use of the CDL yields models that have substantially better agreement with ideal main-chain bond angles and lengths and, on average, a slightly enhanced fit to the X-ray data. No disadvantages of using the backbone CDL are apparent. In phenix, use of the CDL can be selected by simply specifying the cdl = True option. This successful implementation paves the way for further aspects of the context dependence of ideal geometry to be characterized and applied to improve experimental and predictive modeling accuracy.

Keywords: crystallographic refinement; geometry restraints; ideal geometry; protein structure; structural genomics.

Figure 1. The conformation-dependent library (CDL) concept

(A) Ramachandran plots emphasizing the large increase in information content that is associated with shifting from the conventional SVL library in current use ([13]; left hand plot) to the CDL library ([25]; right hand plot) that we have here incorporated into Phenix. Using the color scheme indicated to the right, each plot shows the N-Cα-C bond angle targets for general residues (the 18 non-Gly, non-Pro residues in the case of the SVL and the 16 non-Gly, non-Pro, non-Ile/Val, non-PrePro residues in the case of the CDL). For the CDL, conformation-dependent the N-Cα-C bond angle targets are defined for 10×10° bins of ϕ and ψ whereas for the SVL all are given the single value of 111.0°. Shown in both panels are small white circles marking the ϕ,ψ-angles of the residues shown in panels B and C, and black outlines indicating the regions sufficiently populated so that the CDL library provides actual conformation-dependent values rather than defaulting to a global average value. The global average value for the right hand panel is 110.8°, which can be perceived as having a slightly different hue than the left hand panel color that represents 111.0°. It is of interest to note that the previous adjustments in SVL target values over the last 60 years are equivalent to making such a slight change in hue, while switching from the SVL to the CDL paradigm introduces a rainbow of greater information. (B) The model and 0.86 Å resolution electron density map contoured at 7 ρ_rms showing the evidence for the N-Cα-C bond angle of residue Asn44 in PDB entry 1g6x (with ϕ,ψ-angles=−162°,+106°) that is observed to be 104.5°. (C) Same as B but for residue Asn108 of PDB entry 4ayo (with ϕ,ψ-angles=−122°,−26°) with its 0.85 Å resolution map contoured at 7 ρ_rms and an observed N-Cα-C bond angle of 117.7°. The examples in panels B and C were found using the Protein Geometry Database [26]. (D) Schematic of information content of the backbone CDL showing how a central residue (Yaa) and its C-terminal neighbour (Zaa) define one of 8 residues classes (green lines), and the ϕ,ψ-angles of the residue specify which restraint values to obtain from that class of residue (blue line) for each of the up to 7 backbone bond angles and 5 backbone bond lengths (red lines). The coloring scheme for the CDL plots of each residue type is similar to those in panel A, but with a common background color for simplicity.

Figure 2. Comparing outcomes of the SVL- vs. CDL-based re-refinements of 25976 PDB entries

(A) Shown are the average rms deviations of the backbone bond lengths for structures grouped in 0.1 Å resolution bins. Colors distinguish the results derived from SVL (blue) and the CDL-based (red) calculations. The three pairs of curves are the rms deviations from the library values of the PDB entries as deposited (dashed lines), after re-refinement using Phenix (solid lines), and after regularisation (thick lines). Dotted lines indicate the values obtained if ‘alternate location’ atoms are included in the rmsd values. (B) same as ‘A’ but for backbone angles. (C) same as ‘A’ but for the N-Cα-C bond angle. An inset in panel B shows the absolute changes in R-factors (as measured in percent) for CDL re-refined minus SVL re-refined structures. Changes are shown for R_free(green), R_work (black), and R_free+R_gap (purple) where R_gap is R_free-R_work.

Figure 3. Clashscore filtering improves behavior of lower resolution structures during re-refinement

(A) The rms deviations for the N-Cα-C angle, as a representative indicator of model geometry quality, are plotted as a function of resolution for pdb entries re-refined using the CDL library. The four curves are shown are based on all structures (thick solid), or subsets of the structures that after re-refinement using the SVL had Molprobity clashscore of <9 (thin solid), <6 (dashed) or <3 (dotted). The number of structures in each of the four groups are 25976, 24867, 22052, and 10871, respectively. The clashscore filtering was carried out on PDB entries after discrete modelled water molecules were deleted. (B) A log-scale plot of the number of protein models as a function of resolution in each of the groups shown in figure 3A, with matching line types. The inset shows as a function of resolution what percent of the models at each resolution range were removed by the clashscore <6 filtering that was used for selecting files for the following refinement tests.

Figure 4. Comparing outcomes of the SVL- vs. CDL-based re-refinements of 22052 PDB entries surviving the Clashscore filter

Panels (A) – (C) show results as in Figure 2, comparing the SVL- (blue traces) and CDL-based (red traces) results, except that for each plot only two pairs of curves are shown: the rms deviations from the library values of the PDB entries after re-refinement using Phenix (solid lines), and after regularisation (thick lines). In addition to presenting the average rms deviations as a function of resolution, for panels (A) through (C), boxes are included that represent the range covered from the 25^th to the 75^th percentile values within each resolution bin. Also, as in Figure 2, panel B contains an inset showing the small changes in R_free (green), R_work (black), and R_free+R_gap (purple) that occur upon changing from the SVL to the backbone CDL.