Physically realistic homology models built with ROSETTA can be more accurate than their templates

Abstract

We have developed a method that combines the ROSETTA de novo protein folding and refinement protocol with distance constraints derived from homologous structures to build homology models that are frequently more accurate than their templates. We test this method by building complete-chain models for a benchmark set of 22 proteins, each with 1 or 2 candidate templates, for a total of 39 test cases. We use structure-based and sequence-based alignments for each of the test cases. All atoms, including hydrogens, are represented explicitly. The resulting models contain approximately the same number of atomic overlaps as experimentally determined crystal structures and maintain good stereochemistry. The most accurate models can be identified by their energies, and in 22 of 39 cases a model that is more accurate than the template over aligned regions is one of the 10 lowest-energy models.

Fig. 1.

Low-rmsd models have low energies. Representative data from the 1acf–1awi (Left) and 1b07–2sem (Right) query–template pair centroid and full-atom searches are shown. Each point represents one trajectory. Green points represent the energies and rmsd values of the low-resolution population. Cyan points are the lowest 15% by energy subset of models that were selected for full-atom refinement. Magenta points represent the energies and rmsd values of the final models after full-atom refinement. For reference, the results of five trajectories of idealized minimized 1acf and 1b07 native structures (described in ref. 17) are shown in black. The orange vertical lines represent the rmsd of the template based model; the energy of the template-based model is out of range of this plot.

Fig. 2.

Comparison of atomic clashes observed in native structures, models folded using rosetta with constraints, models generated with modeller, and models generated using a fixed template and rosettasvr modeling protocol. The y axis reports the average number of clashes per test case in aligned regions over the 39 test cases in the benchmark set.

Fig. 3.

The rmsd of refined models selected according to lowest energy, lowest energy*, lowest rmsd, top cluster, and top cluster* compared with the initial fixed template-based model for each protein in the test set. (A) Data were obtained by using 3dpair alignments (structure-based). (B) Data were obtained by using psi-blast alignments (sequence-based). The height of the bars represents the number of cases where the rmsd improved (blue bars), became worse (red bars), or remained unchanged (yellow bars) with respect to the fixed template-based model over aligned regions and over the complete chain. Complete chain models were generated by using rosettasvr modeling.

Fig. 4.

Comparison of predicted buried side-chain (A) and exposed side-chain (B) conformations over aligned regions between the low energy and low energy* models folded with constraints, models produced by using modeller and models produced with rosettasvr. The height of the bars in bins 1, 2, 3, and 4 represent the average number over the test set of side chains where the χ angle was predicted within 0–10°, 10–20°, 20–30°, and 30–40° of the native value, respectively. Green bars, low-energy models folded with constraints; blue bars, low-energy* models folded with constraints; red bars, models built with modeller; yellow bars, models built with rosettasvr.

Fig. 5.

Examples of successful test cases. (A) 1b07 query, 2sem template. The native 1b07 structure is shown in dark gray with the unaligned regions shown in green, the fixed template model in blue, and the model folded with constraints is shown in magenta. Backbone superposition of the native structure and the two models are shown. (B) 1pva query, 1ahr template. The aligned regions of the native 1pva structure, the fixed template model, and the model folded with constraints are shown in gray, blue, and magenta, respectively. The 38-residue unaligned region is shown in gray, green, and red for the native structure, the template-based model, and the model folded with constraints, respectively. This figure was made with pymol (23).

Fig. 6.

The 1pva–1ahr query–template pair full-atom models generated using the boinc distributed computing resource. (A) Cyan points represent the energy and rmsd values of the full-atom models; 54,267 full-atom models are shown. Magenta points represent models that contained nonhelical backbone torsion angles at position 65 and are a subset of the 30% lowest-energy models used for clustering. For reference, the results of 20 trajectories of idealized minimized 1pva native structures (described in ref. 17) are shown in black. The boxes labeled 1 and 2 denote the locations of the models belonging to the largest and second-largest cluster, respectively. The orange vertical line represents the rmsd of the template-based model; the energy of the template-based model is out of range of this plot. (B and C) Models belonging to the largest cluster (B) and second-largest cluster (C) correspond to the lowest-energy and most accurate models in the population, respectively. (B) Average cluster rmsd = 2.5 Å. (C) Average cluster rmsd = 1.4 Å. Models are shown in gray, the native 1pva structure is shown in green, and the irregular helix is shown in magenta. Phe-65, which has nonhelical backbone torsion angles, is shown in blue. The helix in Inset is rotated by 90° to better show the superposition of the backbones and the Phe-65 side chains.