Benchmarking GPCR homology model template selection in combination with de novo loop generation

Abstract

G protein-coupled receptors (GPCR) comprise the largest family of membrane proteins and are of considerable interest as targets for drug development. However, many GPCR structures remain unsolved. To address the structural ambiguity of these receptors, computational tools such as homology modeling and loop modeling are often employed to generate predictive receptor structures. Here we combined both methods to benchmark a protocol incorporating homology modeling based on a locally selected template and extracellular loop modeling that additionally evaluates the presence of template ligands during these modeling steps. Ligands were also docked using three docking methods and two pose selection methods to elucidate an optimal ligand pose selection method. Results suggest that local template-based homology models followed by loop modeling produce more accurate and predictive receptor models than models produced without loop modeling, with decreases in average receptor and ligand RMSD of 0.54 Å and 2.91 Å, respectively. Ligand docking results showcased the ability of MOE induced fit docking to produce ligand poses with atom root-mean-square deviation (RMSD) values at least 0.20 Å lower (on average) than the other two methods benchmarked in this study. In addition, pose selection methods (software-based scoring, ligand complementation) selected lower RMSD poses with MOE induced fit docking than either of the other methods (averaging at least 1.57 Å lower), indicating that MOE induced fit docking is most suited for docking into GPCR homology models in our hands. In addition, target receptor models produced with a template ligand present throughout the modeling process most often produced target ligand poses with RMSD values ≤ 4.5Terms of use and reuse: academic research for non-commercial purposes, see here for full terms. https://www.springer.com/aamterms-v1

G Protein-Coupled Receptor–Ligand Pose and Functional Class Prediction

“…The internal dataset used to develop our random forest classifier (Figure 1, ligandreceptor complex types 5 and 6) required modeled structures for the 5 DUD-E GPCR targets (AA2AR, ADRB1, ADRB2, CXCR4, and DRD3). For each DUD-E GPCR target, a set of three homology models representing "best" and "normal" cases for template selection were constructed using our benchmarked homology modeling protocol (Table 4) [29][30][31]. We chose to develop multiple homology models for each target (rather than just a single homology model) to reflect the range of information that may or may not be available for any GPCR target at the center of a ligand identification study.…”

“…However, observed RMSD values support the use of the CoINPocket similarity score as a metric for homology model template selection regardless of whether a template structure binding to the same endogenous ligand as the target GPCR being modeled is available. For the external dataset, four modeled structures were generated or retrieved for each target: one was constructed with our in-house GPCR modeling protocol [29][30][31], two were retrieved from GPCRdb [26], and one was retrieved from the AlphaFold Protein Structure Database [33]. When generating the in-house homology model for each target in the external dataset, template structures were selected using the CoINPocket similarity metric [30], and loop modeling was again performed after the determination of loop anchor residues (Table S3).…”

“…Ligands were then docked into homology models of mixed origin: one was constructed with our in-house GPCR modeling protocol [29][30][31], two were retrieved from GPCRdb [26], and one was retrieved from the AlphaFold Protein Structure Database [33]. In contrast to ligand docking runs used to construct the internal dataset, the use of multiple homology models for external set docking allowed for the generation of a larger external dataset that reflected docked poses of active and inactive ligands in the context of structurally variable receptor models.…”

“…Using ligands shown to be active or inactive for each target in prior studies by Guo et al [27] (GPR31) and Borowsky et al [28] (TAAR2), an external dataset containing interaction profiles for ligands docked into modeled structures of GPR31 and TAAR2 was constructed and used to further assess our classifier's accuracy in predicting ligand function based on a given binding mode. In addition to simply allowing for the assessment of whether the classifier could correctly predict the binary activity class (active or inactive) of known active and inactive ligands for these targets, the use of modeled structures from various sources (in-house homology models [29][30][31], GPCRdb [26], AlphaFold [32,33]) during construction of the external dataset ensured that the classifier's performance was assessed in the context of structurally variable GPCR models.…”

BIO-GATS: A Tool for Automated GPCR Template Selection Through a Biophysical Approach for Homology Modeling

“…Numerous benchmarking studies have been conducted by incorporating global and local similarity measures to select the appropriate template for GPCRs. Models based on local similarity measures have produced better results in virtual screening experiments ( Castleman et al, 2019 ; Szwabowski et al, 2020 ). Multiple studies have shown that sequence identity above 30% could result in good GPCR homology models (within 3 Å) ( Shahaf et al, 2016 ; Loo et al, 2018 ; Jaiteh et al, 2020 ).…”