Fig 1.
Comparison of average RMSD change using various alignment methods.
A total of 100 models were produced for each alignment method. The average RMSD of the models were normalized to the average RMSD of the models produced with the knowledge-based alignment (black). Values above 1 represent an alignment method that produced on average worse models while values below 1 represent an alignment method that produced on average better models. For (A) all receptors regardless of family, the knowledge-based modeling performs the best regardless of region analyzed. When split between (B) Class A and (C) Classes B, C, and F, the majority of the improvements are found in the Classes B, C, and F where template availability is limited.
Fig 2.
Comparison of single template modeling methods with peptide insertion.
Using only a subset of receptors and templates that were available in our original GPCR modeling benchmark (yellow in S2 Table), 100 models were generated using either a single high identity template or the best template available below 40%. Original results from the 2013 benchmark [22] are displayed in black. Using the hybridize code with the same original templates dramatically improved the results across all measures (medium grey). Using a low identity template in hybridize (light grey) expectedly worsened the results compared to the high identity template but was either better or comparable with the original threading alone algorithm.
Fig 3.
Comparison of average RMSDs for single versus multiple template homology modeling.
Using either one template or ten templates, 1000 models were generated for each target and the average RMSD was calculated over the TM region (A), ECL2 (B), and the full model (C). Values that fall above the diagonal performed better when using multiple templates and values that fall below the diagonal performed better with a single template. Targets are colored by class.
Fig 4.
Examples of results obtainable with RosettaGPCR.
In all cases, the crystal structure is colored grey and the model is blue. Three different ECL2 loops structures are presented: disordered (A), β-sheet (B), and lipid-binding (C). RosettaGPCR performs well on loops containing the conserved disulfide. For lipid receptors lacking the conserved disulfide (C) multiple templates (blue) perform worse than using a single template with similar structure (green), in this case the LPA receptor. Extracellular loops 3 (D) and 1 (E) also perform quite well with this method. In general, RosettaGPCR can model the TM region of most receptors below 2 Å (F). However, for receptors like rhodopsin with complex loop structures and termini (red), the model (cyan) fails to capture the overall conformation (8.0 Å RMSD).
Fig 5.
Comparison of model accuracy using various numbers of starting templates.
For each target, 1000 models were generated using either 1, 5, 10, or all available templates. The average RMSD is plotted for the TM region, ECL2, and the Full Model.
Table 1.
Results of Novel Structure Prediction from Various GPCR Modeling Servers.
Blind predictions were carried out on C5aR1 (PDB ID 6C1R [39]), Y1R (PDB ID 5ZBQ [40]), PTAFR (PDB ID 5ZKQ [41]), and D2R (PDB ID 6CM4 [42]). 100 models were generated for each target with RosettaGPCR and the best scoring model was used for analysis. The RMSD of each model for the various servers were calculated for the TM region, ECL2, and the full model. The best performing model in each evaluation criteria is bolded. No data is available for GPCR-SSFE for D2R.
Fig 6.
Model accuracy with templates over multiple sequence identity ranges.
Three receptors (4ZUD [43], 3PBL [44], and 2RH1 [5]) were identified that had at least 5 templates in each identity range (30–39%, 25–29%, 20–24%, and 15–19%). Using the identified 5 templates for each identity range, 100 models were generated and the RMSD of these models is displayed in box and whisker plots for each region (TM, ECL2, and Full Model).