Human G protein-coupled receptors (hGPCRs) are the largest drug target superfamily (Hauser et al., 2017). One third of FDA-approved drugs target ∼14% hGPCRs (Hauser et al., 2018) and this protein class has a further, untapped pharmacological potential. Unfortunately, rational drug design is hampered by the lack of experimental structures for more than 90% hGPCRs¹ (Munk et al., 2019; Qu et al., 2020). Structural insights into ligand poses for these cases can be obtained by computational modeling. Reliable predictions can be made by docking approaches on homology models, based on templates sharing overall sequence identity >35–40% and/or high conservation of binding site residues (Beuming and Sherman, 2012; Kufareva et al., 2014). For lower resolution models, however, the uncertainty in the structure (and particularly in the orientation of side chains) decreases the accuracy of the docking predictions and, thus, follow-up molecular dynamics (MD) simulations are recommended (Kufareva et al., 2014; Cavasotto and Palomba, 2015; Esguerra et al., 2016; Heifetz et al., 2016; Fierro et al., 2017; Lupala et al., 2018; Rodríguez-Espigares et al., 2020). In an effort at addressing this issue, we have developed a Hybrid Molecular Mechanics/Coarse-Grained (MM/CG) simulation approach (Neri et al., 2005, 2008; Leguèbe et al., 2012; Marchiori et al., 2013; Sandal et al., 2015; Capaldi et al., 2018; Alfonso-Prieto et al., 2019; Fierro et al., 2019). The receptor/ligand interactions are described in atomistic detail, including explicit water molecules in the binding site (MM region), while the rest of the receptor is coarse-grained (CG region) (Schneider et al., 2018). The all-atom force fields used (Schneider et al., 2020) for the MM part of the protein and water are the Amber14SB (Maier et al., 2015) and TIP3P (Jorgensen et al., 1983), respectively, whereas the ligand can be described using either GAFF or GAFF2 (Wang et al., 2004; Case et al., 2020). The CG region is described by a Gō-like (Gō and Abe, 1981) potential. A region at the interface between the MM and CG parts couples the two levels of resolution (Figure 1). The membrane is described implicitly by introducing five potential walls (Leguèbe et al., 2012; Schneider et al., 2018). Two planar walls coincide with the height of the head groups of the membrane lipids, two hemispheric walls cap the extracellular and intracellular ends of the protein and prevent water evaporation, and the last wall follows the initial shape of the interface between protein and membrane, mimicking the effect of the lipid acyl tails (Figure 1). The scheme can be used also for GPCRs other than those from Homo sapiens. This approach turned out to be able to reproduce the ligand poses for four different hGPCRs (Schneider et al., 2020). These include the adenosine 2A receptor in complex with caffeine (i.e., the example case of the webserver), the human bitter receptor 16 in complex with phenyl-β-D-glucopyranoside, the β2-adrenergic receptor with adrenaline, and the dopamine D3 receptor with eticlopride. The structures of these hGPCR/ligand complexes were either experimentally determined, taken from all-atom MD trajectories or predicted based on templates with decreasing resolution, up to extremely low sequence identity (<15%) (Schneider et al., 2020). Retrospective validation against available X-ray structures and mutagenesis data confirmed that the MM/CG approach can predict correct ligand poses and identify experimentally determined binding residues² (Schneider et al., 2020), regardless of the model resolution. In addition, the MM/CG simulations can provide insights into the flexibility of receptor–ligand interactions and hydration of the binding cavity, at a lower computational cost than all-atom molecular dynamics simulations.

The increasing number of applications and requests to use the MM/CG approach has prompted us to develop the Hybrid MM/CG webserver, a publicly accessible web interface aimed at preparing and running short MM/CG simulations of hGPCR/ligand complexes (available since December 2019). To the best of our knowledge, this is the only webserver dedicated to this task. It complements other excellent online resources (Table 1) aimed at preparing all-atom or coarse-grained molecular dynamics (MD) simulations of GPCRs and GPCR/ligand complexes (or in general membrane proteins).

The Hybrid MM/CG webserver requires only the coordinates of the receptor/ligand complex (as PDB file). These may come from experimental structures, simulation snapshots, or computational models generated with other webservers, such as the ones listed in the “Input” section. Furthermore, we have linked our GOMoDo webserver³ for modeling and docking of GPCRs to the Hybrid MM/CG webserver, so that the complexes generated with GOMoDo can be directly transferred. The user is first guided through a set of user-friendly forms for preparing the setup files. The procedure requires only a few minutes. Then, the user runs the initial MM/CG simulation steps (up to 10 ns) directly on the server. The resulting files can be downloaded and the MM/CG simulation can be continued using local resources.

As an application case, we considered the human adenosine 2A receptor (hA_2AR) in complex with its antagonist caffeine (CFF). CFF was proposed to have a protective and therapeutic effect against Parkinson’s disease (Chen et al., 2001; Sonsalla et al., 2012; Petzer and Petzer, 2015; Nazario et al., 2017). The input PDB file of the hA_2AR/CFF complex used for the Example can be downloaded from the same page. The demonstration can be run by simply starting the preparation workflow⁹ without uploading any PDB file. Precomputed results for this example case can be found on the server¹⁰ and selected parts of these results can be seen in Figure 3. Figures 3A–C show the whole hA_2A receptor in complex with CFF, with and without water solvating the binding site and the extracellular loops, as well as the walls mimicking the membrane and preventing water evaporation. Figures 3D–F show the ligand and its direct and water-mediated interactions with residues in the binding cavity, as determined by the NGL viewer (Jakalian et al., 2004). The system can be inspected in 3D after the preparation, short equilibration, and production runs. In addition, plots showing the time evolution of the potential energy and temperature during the different simulation steps are generated (see Supplementary Figures 1–3).

The hybrid MM/CG protocol has been successfully used to predict ligand poses in a variety of hGPCRs (Leguèbe et al., 2012; Marchiori et al., 2013; Sandal et al., 2015; Fierro et al., 2017; Capaldi et al., 2018; Fierro et al., 2019). Moreover, the Amber-based MM/CG poses improved significantly relative to those obtained by simple docking, especially for low resolution starting models (Schneider et al., 2020). However, the setup of such simulations has been time-consuming and system-dependent. The webserver presented here automatically prepares all files needed to start the equilibration, in a short time: the hA_2AR/CFF example requires only 6 min. An additional advantage of the Hybrid MM/CG webserver over most of the online services for simulation of GPCRs (Table 1) is the automatic ligand parameterization. Although other webservers can also include ligands (see Table 1), their parameterization usually has to be done externally. The results of the preparation steps can be checked on-the-fly (e.g., ligand structure and full solvation of extracellular loops). No additional software (except the patch for the GROMACS code, which is available for download¹¹) is needed to extend the simulations locally. This makes the method useful for people acquainted with MD simulations and not necessarily familiar with every single MM/CG preparation step.

The server could prospectively be used for setting up MM/CG simulations of different ligands in complex with the same hGPCR. Capitalizing on the limited computational cost of the MM/CG approach, this could pave the way for low-throughput virtual screening efforts. Furthermore, comparative studies of hGPCRs in multiple states, e.g., active/inactive, can be accomplished by setting up several simulations starting from different initial structures. Modeling of allosteric effects is currently not possible, because of the use of the Gō-like model for the part of the receptor pointing toward the intracellular side. We also plan to integrate the Amber-based MM/CG approach used by the webserver with the recently developed open boundary MM/CG for grand canonical simulations (Tarenzi et al., 2017, 2019), aiming at calculating ligand binding free energies. Further planned developments include the extension to other MD codes, the implementation of the CHARMM36m force field (Huang et al., 2017) for the MM/I regions, as well as the adaptation of the scheme to other membrane protein families, such as ion channels.

The Hybrid MM/CG Webserver, available at https://mmcg.grs.kfa-juelich.de/, is free and open to all. There is no login requirement.

JS, MA-P, PC, and AG contributed to workflow design and documentation. JS developed the server. RR designed the interface on the GOMoDo webserver. JS, RR, MA-P, and AG tested the webserver. All authors participated in writing the article.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.