Subtype-selective mechanisms of negative allosteric modulators binding to group I metabotropic glutamate receptors

Abstract

Group I metabotropic glutamate receptors (mGlu₁ and mGlu₅) are promising targets for multiple psychiatric and neurodegenerative disorders. Understanding the subtype selectivity of mGlu₁ and mGlu₅ allosteric sites is essential for the rational design of novel modulators with single- or dual-target mechanism of action. In this study, starting from the deposited mGlu₁ and mGlu₅ crystal structures, we utilized computational modeling approaches integrating docking, molecular dynamics simulation, and efficient post-trajectory analysis to reveal the subtype-selective mechanism of mGlu₁ and mGlu₅ to 10 diverse drug scaffolds representing known negative allosteric modulators (NAMs) in the literature. The results of modeling identified six pairs of non-conserved residues and four pairs of conserved ones as critical features to distinguish the selective NAMs binding to the corresponding receptors. In addition, nine pairs of residues are beneficial to the development of novel dual-target NAMs of group I metabotropic glutamate receptors. Furthermore, the binding modes of a reported dual-target NAM (VU0467558) in mGlu₁ and mGlu₅ were predicted to verify the identified residues that play key roles in the receptor selectivity and the dual-target binding. The results of this study can guide rational structure-based design of novel NAMs, and the approach can be generally applicable to characterize the features of selectivity for other G-protein-coupled receptors.

Introduction

G-protein-coupled receptors (GPCRs) constitute a large superfamily of membrane proteins that transduce endogenous signals across the cell membrane. Metabotropic glutamate receptors (mGlus) are class C GPCRs, and are distinguished by a characteristically large extracellular (VFT) domain that is linked to the seven-transmembrane (7TM) domain by a cysteine-rich domain (Supplementary Fig. S1) [1, 2]. To date, eight subtypes of these receptors (mGlu_1–8) have been identified and can be divided into three groups according to their biochemical, structural and pharmacological properties (Supplementary Table S1) [3]. Group I receptors (mGlu₁ and mGlu₅) are considered promising targets for the treatment of many psychiatric and neurodegenerative disorders including chronic pain, schizophrenia, Alzheimer’s disease, depression, and anxiety [4–12]. In the VFT domain of mGlus, a distinctive ligand-binding site (orthosteric site) exists for endogenous compounds. An allosteric site is present in the 7TM domain, and the binding of modulators results in a conformational change, which affects the binding properties of orthosteric ligands. Based on their functional effects, these modulators can be divided into three types: positive allosteric modulators (PAMs), negative allosteric modulators (NAMs), and silent allosteric modulators (SAMs) [1, 2]. Furthermore, allosteric modulators have attracted more attention than traditional orthosteric mGlu_1/5 compounds. This is because allosteric modulators are safer than orthosteric ligands in drug development, which can reduce the risk of receptor oversensitization by modulating the natural response to the endogenous ligand instead of directly activating the receptor [13–18].

Dipraglurant, a selective NAM of mGlu₅, received orphan drug designation from the US FDA in 2016 for Parkinson’s disease levodopa-induced dyskinesia [19]. In addition, several alkyne-based mGlu₅ NAMs (mavoglurant, basimglurant, and STX107) and non-alkyne-based NAMs (fenobam, HTL14242, RGH-618, and VU0424238) are progressing to clinical studies (Fig. 1a and Supplementary Table S2) [14, 20–27]. For the mGlu₁ receptor, although a number of selective NAMs with high efficacies have been reported for many neurological diseases (Fig. 1b), so far, none of them has been marketed or entered a clinical study, and the binding mode of NAMs in mGlu₁ remains elusive [28–31].

Fig. 1. Chemical structures of NAMs.

a The structures of clinical mGlu₅ NAMs, b the structures of four selective mGlu₁ NAMs and one dual-target NAM (VU0467558).

As in other GPCR-based drug discovery programs, subtype selectivity is an important requirement for the drug candidates acting at mGlu receptors [13, 32–34]. Research from the Gloriam group demonstrates that the difficulty in achieving subtype-selective compounds with the desired properties is a main factor limiting the approval of a marketed drug for mGlu receptors [34]. mGlu₁ and mGlu₅ have a high sequence identity of 78.5% between the 7TM domain of the two receptors (Supplementary Fig. S2) [16, 35]. There are only six non-conserved residues in the two allosteric sites [2]. Moreover, the literature reports the use of mGlu₁ NAM to produce motor and cognitive side effects, as well as shows concerns about the psychotomimetic effects of certain mGlu₅ NAM [15, 36, 37]. Therefore, it is important to develop NAMs with excellent subtype-selectivity profiles for mGlu₁ and mGlu₅. On the other hand, according to research by Emmitte group, dual-target mGlu_1/5 NAMs may achieve an improved safety profile by reducing the occupancy at each receptor required for efficacy [38]. Although a new lead series of dual-target mGlu_1/5 NAMs have been discovered, active compounds that are selective versus the other six members of the mGlu family and that exhibit the properties required for use in vivo are lacking. Therefore, exploring the subtype-selective mechanism of mGlu₁ and mGlu₅ NAMs will provide a considerable insight into the rational design of novel NAMs with single- or dual-target mechanisms of action.

The crystal structures of the TM domain of mGlu₁ and mGlu₅ have been successfully resolved with several NAMs [1, 2, 24, 28, 39, 40], providing good start points for computational modeling to accurately predict the ligand–receptor binding mode [41–46] and provide further valuable insights into the drug-selective mechanism of targets with high homology [47–52]. In our previous study, the shared common features of five clinical NAMs (mavoglurant, dipraglurant, basimglurant, STX107, and fenobam) interacting with 11 residues in the allosteric site of mGlu₅ were explored by molecular dynamics (MD) simulations [53]. However, except for fenobam, the other four compounds are alkyne-based scaffolds.

In this study, three new non-alkyne-based clinical candidates (HTL14242, RGH-618, and VU0424238) of mGlu₅ were selected to improve the shared binding mode of NAMs in mGlu₅ by combining molecular docking and MD simulations [14, 23, 24, 54]. The binding mode of fenobam was further optimized based on the co-crystalized structure (PDB ID: 6FFH) reported in 2018 [39]. For the mGlu₁ receptor, four selective NAMs (FITM, BDBM50301534, EMQMCM, and JNJ16259685) with high efficacies were selected to predict the elusive binding modes of NAMs in mGlu₁ [28–31]. Based on the equilibrated MD trajectories, the ligand–protein binding free energy and interaction fingerprint were analyzed. As a result, the subtype-selective features underlying the binding modes of 12 different NAMs in mGlu₁ or mGlu₅ were identified. Furthermore, the binding modes of a unique dual-target NAM (VU0467558) [38] in mGlu₁ and mGlu₅ (Fig. 1b) were predicted to verify the subtype-selective features and to provide information for the design of novel dual-target NAMs. This study has positive implications for the treatment of psychiatric and neurodegenerative disorders that seriously affect the quality of human life [44, 55].

Materials and methods

Molecular docking

Molecular docking is one of the most frequently used methods to determine the predominant binding pose of a ligand within a protein binding site [56]. Information about the mGlu₁ and mGlu₅ crystal structures is summarized in Supplementary Table S3. The allosteric sites studied in this work are located in the 7TM domain of mGlus, far from the VFT and RCD domains (Fig. S1). For mGlu₁, only one crystal structure of the receptor’s 7TM domain complexed with a selective NAM FITM (PDB ID: 4OR2) was selected for the docking study. For mGlu₅, seven crystal structures are available in the PDB database, including five 7TM domain structures complexed with different NAMs and two apo full structures. To explore the selective binding mechanism of NAMs at the allosteric sites, the high-resolution crystal structure of the first representative NAM (mavoglurant) bound receptor (PDB ID: 4OO9) was selected for docking other NAMs into mGlu₅. The extracellular and intracellular loops (ECL2 and ICL2) in 4OO9 and 4OR2 are missing. They are far from the allosteric site, but ECL2 is involved in ligand entry [2, 13]. Therefore, the missing structures were modeled and refined by means of Prime in Schrödinger software [57].

In this work, the standard precision (SP) Glide Docking and Induced Fit Docking (IFD) [58] were employed to predict the initial conformations of NAMs binding in the allosteric sites of group I mGlus. To guarantee the reliability of the docking method, re-docking and cross-docking studies were performed based on crystal structures 4OR2, 4OO9, 5CGD, and 6FFH (bound with FITM, mavoglurant, HTL14242 and fenobam) [2, 24, 28, 39, 40, 59]. During docking, the Protein Preparation Wizard tool was used to prepare the receptor structures and the size of the grid boxes was set by selecting the original ligands in the allosteric binding site. The 3D structures of the studied NAMs retrieved from the PDB or PubChem database were prepared using LigPrep [60] with the OPLS-2005 force field for energy minimization and with the ionized state assigned by Epik [61] at a pH value of 7.0 ± 2.0. Finally, SP Glide Docking was applied to search the ligand–receptor binding pose and IFD was used to refine the interaction between the ligand and the binding site residues (within 5.0 Å of the ligand).

MD simulations

Preparation for MD simulation

Initial conformations of NAMs binding to mGlu₁ and mGlu₅ from docking or re-docking were prepared for the MD simulations. First, OPM [62] was used to calculate the spatial orientations of the NAM–mGlu_1/5 complexes with respect to the membrane normal defined by the Z-axis. Second, membrane builder in CHARMM-GUI [63–65] was employed to build the solvent box: each of the complex was inserted into the explicit POPC lipid bilayer and immersed in TIP3P water with a 20.0 Å thickness, and NaCl was added with a concentration of 0.15 mol/L according to the physiological concentration [43, 66]. As a result, each system contained ~60,140 atoms per periodic cell with a box size of 74 × 74 × 115 Å. Then, the HF/6-31G* level of Gaussian09 suite [67] was used to implement the geometric optimization and the electrostatic potential calculation of NAMs. The LEaP module in AMBER16 [68, 69] was used to assign the force field parameters for each of the complex, with ff14SB [70] and Lipid14 [71] were used for protein and lipids, respectively. Meanwhile, the conserved disulfide bonds between cysteine residues (Cys644-Cys733 in mGlu₅ and Cys657-Cys746 in mGlu₁) were considered by LEaP. Finally, the force field parameters for NAMs were created by use of the Antechamber program, using gaff [72] atom types and RESP [73] partial charges.

Implementation of MD simulation

All MD simulations were carried out in four stages with GPU-accelerated NAMD (version 2.12) [74]. Energy minimization and equilibration simulations were conducted before the production simulations to eliminate bad contacts formed by solute and solvent water molecules. Firstly, except for the lipid tail all atoms were fixed, 0.1 ns minimization and 0.5 ns equilibration were carried out. Then, ligand heavy atoms, protein Cα atoms and ions were fastened, and the system was further minimized for 0.1 ns and equilibrated for 0.5 ns. Next, a 5 ns equilibration MD simulation was carried out with all of the atoms free. Finally, a 100 ns production MD simulation was conducted in the NPT ensemble at a temperature of 310.0 K and a pressure of 1 bar. The integration step size was 2 fs and the coordinates of the trajectory were saved every 5000 steps. Moreover, periodic boundary condition was employed and electrostatic calculation was based on the particle-mesh Ewald (PME) method with a 10.0 Å nonbonded cutoff. For each complex, multiple parallel trajectories were conducted to ensure the convergence of the simulation. After the computational simulation, all of the MD trajectory analyses were based on CPPTRAJ [75–77], such as the root means square deviation (RMSD) calculation and the extraction of representative snapshots.

Binding free energy analysis

The binding free energies (ΔG_cal) of the NAM–mGlu_1/5 complexes were calculated by the molecular mechanics/generalized Born surface area (MM/GBSA) method implemented in AMBER16 [78, 79]. For each complex, 500 snapshots were extracted from the last 50 ns equilibrium trajectory, and the binding free energy was calculated by:

$$\mathrm{\Delta }{G}_{\mathrm{cal}\left(\mathrm{MM/GBSA}\right)}=\mathrm{\Delta }{E}_{\mathrm{vdW}}+\mathrm{\Delta }{E}_{\mathrm{ele}}+\mathrm{\Delta }{G}_{\mathrm{pol}}+\mathrm{\Delta }{G}_{\mathrm{nonpol}}.$$ 1

The energy terms ΔE_vdW, ΔE_ele, ΔG_pol, and ΔG_nonpol in the equation represent the van der Waals interaction, electrostatic contribution in the gas phase, polar and nonpolar solvent interaction energies, respectively. Besides, ΔE_vdW and ΔE_ele were calculated using the AMBER force field ff14SB, the modified GB model (igb = 2) was used to calculate the electrostatic free energy of solvation (ΔG_pol), and ΔG_nonpol was estimated via the solvent accessible surface area (ΔSASA) using the linear combination of pairwise overlaps (LCPO) method calculated by ΔG_nonpol = 0.0072 × SASA with 1.4 Å Probe radii. And in the binding free energy calculation, solute and solvent dielectric constants were set to 1 and 80.

Then, per-residue decomposition energy ($\mathrm{\Delta }{G}_{\mathrm{cal}}^{\mathrm{per}-\mathrm{residue}}$) was determined to quantitatively analyze the energy contribution of per residue by decomposing the total binding free energy into each residue according to the following equation:

$$\mathrm{\Delta }{G}_{\mathrm{cal}}^{\mathrm{per}-\mathrm{residue}}=\mathrm{\Delta }{E}_{\mathrm{vdW}}^{\mathrm{per}-\mathrm{residue}}+\mathrm{\Delta }{E}_{\mathrm{ele}}^{\mathrm{per}-\mathrm{residue}}+\mathrm{\Delta }{G}_{\mathrm{pol}}^{\mathrm{per}-\mathrm{residue}}+\mathrm{\Delta }{G}_{\mathrm{nonpol}}^{\mathrm{per}-\mathrm{residue}}$$ 2

$\mathrm{\Delta }{E}_{\mathrm{vdW}}^{\mathrm{per}-\mathrm{residue}}$, $\mathrm{\Delta }{E}_{\mathrm{ele}}^{\mathrm{per}-\mathrm{residue}}$ and $\mathrm{\Delta }{G}_{\mathrm{pol}}^{\mathrm{per}-\mathrm{residue}}$ are the same terms as depicted in Eq. (1). However, ΔSASA was obtained from the icosahedron (ICOSA) method, rather than the LCPO method, which was employed to calculate $\mathrm{\Delta }{G}_{\mathrm{nonpol}}^{\mathrm{per}-\mathrm{residue}}$.

Hierarchical clustering analysis of per-residue binding free energy

The hierarchical clustering analysis was performed based on the per-residue energy contribution to recognize the common features of the ligand–receptor complexes. Additionally, a clustering tree was generated using R statistical analysis software [80] with the similarity levels among the vectors measured by the Manhattan distance:

$$\mathrm{Distance}\left(a,b\right)={\sum }_i∣a_i-b_i∣,$$ 3

where i represents each dimension of per-residue energies a and b. the Cluster algorithm applied here was based on the Ward’s minimum variance method, which was developed to minimize the total within-cluster variance. The hierarchical tree graph was generated by the online tree generator iTOL [81].

Computational alanine scanning analysis

In order to verify the credibility of the critical residues identified by MD simulations, the computational alanine scanning (CAS) was calculated in this study [82, 83]. The last 50 ns trajectories were also used to produce mutated snapshots, and then the binding free energies of 500 mutated snapshots were calculated. More specifically, the alanine mutation was generated by cuting off the selected residue at Cγ atom and replacing it with a hydrogen atom at a 1.09 Å distance from Cß along the direction of Cγ–Cß bond. And the topology files for all the mutation complexes were regenerated by the LEaP module in AMBER16 [68, 69]. Next, the MM/GBSA method was also used to calculate the mutated binding free energy, and the relative binding free energy (ΔΔG_CAS) was used to describe the difference between the WT and MUT systems, as shown below:

$$\mathrm{\Delta }\mathrm{\Delta }{G}_{\mathrm{CAS}}=\mathrm{\Delta }{G}_{\mathrm{MUT}}-\mathrm{\Delta }{G}_{\mathrm{WT}},$$ 4

where ΔG_WT and ΔG_MUT in Eq. (4) are the binding free energies of the wild-type and mutated systems, respectively.

Molecular interaction fingerprint analysis

Decomposition analysis of the binding free energy ($\mathrm{\Delta }{G}_{\mathrm{cal}}^{\mathrm{per}-\mathrm{residue}}$) describes the energy contribution of per residue, and then molecular interaction fingerprint analysis was conducted to further clarify the interactions between ligand and receptor in this study. In each simulation system, 500 snapshots extracted from the last 50 ns MD simulation trajectories were used to implement the molecular interaction fingerprints based on IChem with the default parameters [79, 84, 85]. Residues within 5.0 Å of the ligand mass center were extracted from trajectories to form MOL2 files. Then, molecular interaction fingerprints were calculated, and seven interactions (hydrophobic, aromatic face-to-face, aromatic edge-to-face, hydrogen bond accepted by ligand, hydrogen bond donated by ligand, ionic bond with ligand negatively charged, ionic bond with ligand positively charged) were successively considered in every snapshot [84]. If a particular interaction between receptor and ligand was detected, “1” was output, otherwise “0” was recorded. Finally, the calculated results were displayed in a radar chart to clearly represent the ligand–receptor interactions.

Results and discussion

Initial conformations of NAMs binding to mGlu₁ and mGlu₅

In this work, molecular docking was carried out to predict the initial binding poses of NAMs to the mGlu₁ and mGlu₅ receptor. To guarantee the reliability of the docking protocol, re-docking and cross-docking were first performed based on four crystal structures (PDB ID: 4OR2, 4OO9, 6FFH, and 5CGD) (Supplementary Fig. S3a). The RMSD values between the crystal conformations and re-docking poses for FITM, mavoglurant, fenobam and HTL14242 are 0.48, 0.17, 0.64, and 0.35 Å, respectively. The obtained cross-docking poses of mavoglurant (binding to 5CGD) and HTL14242 (binding to 4OO9) are also consistent with the crystal conformations, and the RMSD values are 0.46 and 1.09 Å, indicating that the docking protocol is suitable for obtaining the receptor–ligand complexes in this study.

The 20 top-ranked poses were generated from SP Glide Docking and IFD refinement for each studied NAM. According to their conformations in the binding pockets, they were defined as two clusters (cluster 1 and cluster 2 in Supplementary Fig. S3b). The docking results are consistent with the previous studies showing that NAMs represent two completely different conformations when binding to mGlu receptors [35, 39, 86–88]. The docking poses of FITM and its analog BDBM50301534 were selected according to the binding mode of FITM reported by the Gloriam group, in which the N-methylamide moiety is flipped compared with the co-crystalized structure 4OR2 (Supplementary Fig. S3a) [28, 34]. For the other NAMs (EMQMCM, JNJ16259685, VU0467558, RGH-618, and VU0424238), as none of them show significant chemical similarity to crystal ligands (the Tanimoto coefficients of the NAMs were calculated and the values are shown in Supplementary Table S4), two representative poses from two docking conformation clusters for each ligand (Supplementary Fig. S3b) were selected as the initial poses for further MD simulation and MM/GBSA analysis.

Molecular mechanism of NAMs binding to mGlu₁ and mGlu₅

MD simulation of NAM–mGlu₁ and NAM–mGlu₅ complexes

For each NAM–mGlu_1/5 complex, multiple 100 ns MD simulations with different initial velocities were carried out (4.5 μs in total), starting from the initial pose picked out from molecular docking. The simulation stabilities were assessed by the RMSD values of protein backbone atoms, backbone atoms of binding site residues (within 5.0 Å of ligand mass center) and ligand heavy atoms as a function of time (Supplementary Fig. S4). As shown, the complexes were stable during the simulation inferring from a small variation of the RMSD values for binding sites and ligands. Compared with the initial pose, the calculated RMSD values indicate that the NAM and receptor have undergone appropriate conformational changes to accommodate each other during the simulation and maintain the equilibration state.

Interaction and binding free energy analysis of mGlu₁–NAM complexes

To interpret the molecular mechanism of NAM binding to mGlu₁, the representative structure of each complex was extracted from the equilibrated MD trajectory, and the MM/GBSA method [89, 90] was applied to calculate the total binding free energies (ΔG_cal) (Table 1, Supplementary Table S5 and Fig. 2). As Fig. 2 depicts, the nonpolar residues distributed around the binding pocket, such as L648^2.56, V664^3.36, V753^5.40, P756^5.43, L757^5.44, I797^6.49, W798^6.50, F801^6.53, I812^7.29, A818^7.35, and V819^7.36 interact with the NAMs through hydrophobic contacts. In the binding pocket, residues Q660^3.32, R661^3.33, T748^ECL2, G752^5.39, N760^5.47, and Y805^6.57 form hydrogen bonds with NAMs directly or mediated through water molecules. For instance, all four NAMs contain an aldehyde that forms a hydrogen bond with N760^5.47, and the aldehyde connects to the monocyclic and polycyclic substituents on both sides of the molecule, respectively. In addition, water mediated polar interactions have also been found in the binding modes of NAM to the members of mGlu family [13, 35, 39, 91, 92]. Water networks play an important role in understanding binding modes, exploring selective features and designing new ligands targeting mGlu members.

Table 1.

The calculated and experimental data of NAMs binding to mGlu₁ and mGlu₅ receptors (ΔG is in kcal/mol).

NAMs	IC50/Ki		ΔGexpa	ΔGcalb
	mGlu1	mGlu5
FITM	5.10c	7000c	−11.75	−49.49
BDBM50301534	10.0c	6600c	−11.34	−39.05
EMQMCM	8.10c	>10000c	−11.47	−48.01
JNJ16259685	1.21c	28,000c	−12.64	−52.51
VU0467558	69.0d	129d	−10.15	−38.94
fenobam	–	162.2e	−9.62	−37.73
HTL14242	–	0.501e	−13.18	−45.25
RGH-618	–	4.90e	−11.78	−44.41
VU0424238	–	4.40e	−11.84	−41.44
VU0467558	69.0d	129d	−9.60	−37.55

^aExperimental binding free energies (ΔG_exp) of the NAMs in mGlu₁ and mGlu₅ were estimated using reported IC₅₀ and K_i values by ΔG_exp = RTln(IC₅₀) and ΔG_exp = RTln(K_i), respectively (R = 8.314 J/(mol K) and T = 310.0 K).

^bEstimated MM/GBSA binding free energies in this work.

^cIC₅₀ values (nM) of human mGlu_1/5 receptor based on Ca²⁺ mobilization assay.

^dIC₅₀ values (nM) of rat mGlu₁ receptor based on Ca²⁺ mobilization assay.

^eK_i values (nM) of human mGlu₅ receptor based on competitive radioligand binding assay, the data collected from ChEMBL database and previous studies.

Fig. 2. Representative structures of selective mGlu₁ NAMs binding to the allosteric site from equilibrated MD trajectories.

a, b, c, c1, d, and d1 are the binding conformations of FITM, BDBM50301534, EMQMCM (pose 1), EMQMCM (pose 2), JNJ16259685 (pose 1), and JNJ16259685 (pose 2), respectively. The important residues and NAMs are shown in sticks. The hydrogen bonds are depicted as red dotted lines, and the water molecules are displayed as red balls.

In FITM and BDBM50301534 binding complexes, the N-methylamide moiety flips and forms a hydrogen bond with N760^5.47, which is consistent with the optimized binding mode reported by the Gloriam group (Fig. 2) [34]. The calculated binding free energies of FITM and BDBM50301534 in mGlu₁ are −49.49 and −39.05 kcal/mol (Table 1). The higher binding affinity of FITM can be explained by the additional hydrogen bonds (with T748^ECL2 and Y805^6.57) and the stronger hydrophobic interactions (with L648^2.56 and I812^7.29) to some extent.

For EMQMCM and JNJ16259685, two docking poses representing cluster 1 (pose 1) and cluster 2 (pose 2) were used for MD simulations. The calculated binding free energies for EMQMCM (pose 1), EMQMCM (pose 2), JNJ16259685 (pose 1), and JNJ16259685 (pose 2) are −48.01, −44.82, −52.51, and −49.30 kcal/mol, respectively (Fig. 2 and Supplementary Table S5). Compared to pose 2 of EMQMCM, the pose 1 participates in stronger hydrophobic interactions with residues V664^3.36, V753^5.40, and it also forms additional π–π interaction with F801^6.53. Likewise, stronger interactions between JNJ16259685 (pose 1) and residues V664^3.36, V753^5.40, T815^7.32 are also observed. Therefore, the equilibrated conformations from MD simulation of pose 1 (Fig. 2c, d) were proposed as the binding modes of EMQMCM–mGlu₁ and JNJ16259685–mGlu₁ complexes.

“Hot Spots” and “Warm Spots” in mGlu₁ for NAM binding

To quantify the contribution of each residue in mGlu₁ to NAM binding, the total binding free energy was decomposed on a per-residue basis. Fig. 3 shows a collection of 19 residues whose absolute energy contribution value is ≥0.5 kcal/mol in any complex of the four selective NAMs binding in mGlu₁. Among them, five residues V664^3.36, V753^5.40, L757^5.44, N760^5.47, and F801^6.53 contributed significantly to the binding of NAMs (absolute energy contribution value ≥ 1.5 kcal/mol) and were defined as “Hot Spots”. In addition, ten residues (G665^3.37, S668^3.40, T748^ECL2, P756^5.43, I797^6.49, W798^6.50, Y805^6.57, T815^7.32, A818^7.35, and V819^7.36) that were necessary for at least two binding complexes (0.5 kcal/mol ≤ absolute energy contribution value < 1.5 kcal/mol), were defined as “Warm Spots”. As a result, 15 residues including 5 “Hot Spots” and 10 “Warm Spots” were considered as the major energy contributors shared by the four selective NAM–mGlu₁ complexes, and they are mainly located in the TM3, TM5, TM6, and TM7 domains of mGlu₁. It should be noted that P756^5.43, N760^5.47, F801^6.53, Y805^6.57, and T815^7.32 have been implicated in mutagenesis-based studies, and the mutations can reduce the binding affinity of FITM [28, 93].

Fig. 3. Per-residue binding free energies of four NAMs–mGlu₁ complexes.

Five residues V664^3.36, V753^5.40, L757^5.44, N760^5.47, and F801^6.53 contributed significantly to the binding of NAMs (absolute energy. contribution value ≥1.5 kcal/mol) were defined as “Hot Spots”, and ten residues (G665^3.37, S668^3.40, T748^ECL2, P756^5.43, I797^6.49, W798^6.50, Y805^6.57, T815^7.32, A818^7.35, and V819^7.36), which were necessary at least two binding complexes (0.5 kcal/mol ≤ absolute energy contribution value < 1.5 kcal/mol), were defined as “Warm Spots”.

Interaction and binding free energy analysis of mGlu₅–NAM complexes

In our previous study, the binding mechanism of five clinical NAMs (mavoglurant, dipraglurant, basimglurant, STX107, and fenobam) in the mGlu₅ receptor was explored by the same computational approach. In this study, three new non-alkyne-based clinical candidates (RGH-618, VU0424238, and HTL14242) of mGlu₅ were picked out to improve the shared binding mode of NAMs in mGlu₅. Furthermore, the binding mode of fenobam in mGlu₅ was optimized based on the crystal structure (PDB ID: 6FFH) reported in 2018 [39]. For each complex, the representative structures extracted from the equilibrated MD trajectories and the estimated ΔG_cal values are shown and labeled in Fig. 4. The results of MD simulation show that the novel non-alkyne-based clinical candidates (fenobam, RGH-618, VU0424238, and HTL14242) of mGlu₅ located in the same binding pocket as alkyne-based clinical NAMs [53] defined by residues (G624^2.45, I625^2.46, I651^3.36, S654^3.39, P655^3.40, S658^3.43, Y659^3.44, V740^5.40, L744^5.44, N747^5.47, T781^6.46, W785^6.50, F788^6.53, M802^7.32, S805^7.35, V806^7.36, S809^7.39, and A810^7.40) from the TM2, TM3, TM5, TM6, and TM7 domains. The aromatic ring of non-alkyne-based NAMs (corresponding to the alkyne group of alkyne-based NAMs) binds to the “narrow linker” in a similar way for the side instead of lying flat. Moreover, the conformations of HTL14242 and fenobam in the binding pocket are similar to the crystal structures (PDB ID: 5CGD and 6FFH) [24, 39], making direct or water mediated hydrogen bonds with residues I651^3.36, S654^3.39, Y659^3.44, N747^5.47, T781^6.46, W785^6.50, and S809^7.39. For the fenobam–mGlu₅ complex, a difference from the crystal structure is that the distance between the ligand and S809^7.39 decreases from 2.9 to 2.7 Å, resulting in additional hydrogen bonds between Y659^3.44 and W785^6.50 with fenobam. In 2017, a novel NAM RGH-168 was discovered with its preclinical characterization. Galambos et al. reported that the phenyl rings and one sulfone oxygen of RGH-168 form aromatic interactions and hydrogen bond with W785^6.50 and S805^7.35 [23]. In this work, two docking poses of RGH-168 were sampled by MD simulation (Fig. 4c, c1). The better binding conformation (Fig. 4c) shows that, except for the aromatic interactions with W785^6.50, an additional interaction is observed between the sulfone oxygen and residue N747^5.47, and another sulfone oxygen forms hydrogen bond with S805^7.35 through a water molecule. For the VU0424238, among its two binding conformations in Fig. 4d, d1, the former is proposed to be the better one. In Fig. 4d, VU0424238 has a higher binding affinity and forms an interactional network of hydrogen bonds with several residues (R648^3.33, Y659^3.44, T781^6.46, and S805^7.35) and water molecules.

Fig. 4. Representative structures of mGlu₅ NAMs binding to the allosteric site from equilibrated MD trajectories.

a, b, c, c1, d, and d1 are the binding conformations of HTL14242, fenobam, RGH-618 (pose 1), RGH-618 (pose 2), VU0424238 (pose 1), and VU0424238 (pose 2), respectively. The important residues and NAMs are shown as sticks. The hydrogen bonds are depicted as red dotted lines, and the water molecules are displayed as red balls.

Common features shared by NAM–mGlu₅ recognition

The hierarchical clustering analysis of per-residue binding free energy was exploited to characterize the common features shared by clinical NAMs (fenobam, HTL14242, RGH-618, VU0424238, mavoglurant, basimglurant, dipraglurant, STX107) in the mGlu₅ receptor allosteric site (Fig. 5). A total of 149 residues were picked out for cluster analysis, and these residues have energy contributions in at least one NAM–mGlu₅ system. As shown, the per-residue energy contributions of residues in group A (I625^2.46, I651^3.36, S654^3.39, P655^3.40, L744^5.44, W785^6.50, F788^6.53, M802^7.32, V806^7.36, S809^7.39 and A810^7.40) are consistently higher than those of residues in other groups, and are characterized as the common features shared by the eight clinical NAM–mGlu₅ complexes with different chemical structures. Furthermore, several residues with destabilization contribution to the NAM–mGlu₅ complexes were also identified, which were harmful to ligands binding. Among them, residue G624 in the HTL14242–mGlu₅ complex has the maximum energy contribution of 0.07 kcal/mol.

Fig. 5. Hierarchical clustering trees of 149 residues with energy contribution to at least one of eight clinical NAMs based on their per-residue energy contributions.

Energy contributions favoring ligand binding are displayed in red, with the highest contribution set as exact red and lower contributions gradually fading towards white. Energy contributions hampering ligand binding are shown in blue, with the highest contribution set as exact blue and lower contributions gradually fading towards white.

Verifying the computational models of MD simulation

The ranking ability of the estimated binding free energy ΔG_cal is usually considered to be important evidence for evaluating the accuracy of the computational model [89, 90, 94]. To verify the models of MD simulation in our study, the experimental binding free energies (ΔG_exp) of the NAMs in mGlu₁ and mGlu₅ were calculated using reported activity values by ΔG_exp = RTln(IC₅₀) or ΔG_exp = RTln(K_i) (Table 1) [10, 14, 23, 24, 28–31, 38, 86]. The correlation coefficient (R²) between the values of ΔG_cal and ΔG_exp for five mGlu₁ NAMs (FITM, BDBM50301534, EMQMCM, JNJ16259685, and the dual-target NAM VU0467558) and eight mGlu₅ NAMs (fenobam, HTL14242, RGH-618, VU0424238, mavoglurant, basimglurant, MPEP, and 51D) are 0.71 and 0.69, respectively (Fig. 6). Therefore, the computational model is capable of evaluating the relative binding affinities of NAMs to mGlu₁ and mGlu₅. Nevertheless, the ΔG_cal for each complex is overestimated compared to the experimental results (ΔG_exp), which is caused by omitting the entropy effect from the formula. In this work, the calculation of free energy was used to rank the ability of NAMs binding to mGlu₁ and mGlu₅ and then to identify the key residues, rather than calculating the absolute binding free energy of the ligand–receptor complex. To further ensure the reliability of our computational models, multiple parallel trajectories were conducted, and the results indicate that the computational binding free energy can be reproduced by the parallel trajectories of each NAM–mGlu_1/5 complex (Supplementary Table S5).

Fig. 6. Verification of computational model.

a is the correlation between the calculated (ΔG_cal) and experimental (ΔG_exp) binding free energies for NAM–mGlu1 complexes, and b is the corresponding value for NAM–mGlu5 complexes. The circle, square, and triangle indicate the NAMs targeting mGlu1, mGlu5, and mGlu1/5, respectively. c is the result of computational alanine scanning based on five “Hot Spots” in four NAMs–mGlu1 complexes.

On the other hand, in silico site-directed mutagenesis analysis has also been conducted on the MPEP–mGlu₅ complex based on seven residues (P655^3.40, S658^3.43, Y659^3.44, L744^5.44, T781^6.46, and Y792^6.57), and the calculated results showed a good correlation (R² = 0.90) with the experimental values in our previous study [53]. Herein, to verify the identified important residues, the CAS analysis was performed with the five “Hot Spots” (V664^3.36, V753^5.40, L757^5.44, N760^5.47, and F801^6.53) in NAM–mGlu₁ complexes. As shown in Fig. 6c, the ΔΔG_CAS values of the five “Hot Spots” range from 1.72 to 4.11 kcal/mol, which implies that they are the major energy contributors, and further indicates that the computational model can recognize key residues of NAMs–mGlu_1/5 complexes. In summary, the computational models are reliable for exploring the subtype-selective mechanism of NAMs in mGlu₁ and mGlu₅.

Subtype-selective mechanism of mGlu₁ and mGlu₅ to NAMs

As mGlu₁ and mGlu₅ are transmembrane proteins, NAMs from the outside of membrane enter the allosteric cavity through a narrow entrance [2, 28]. Structural alignment of mGlu₁ and mGlu₅ shows that the allosteric binding pockets of the two receptors do not completely coincide (Fig. 7). The two pockets of mGlu₁ and mGlu₅ are divided into three subsites: the “upper chamber”, “lower chamber”, and “narrow linker”, of which only the upper chamber is an area where the two pockets overlap [13, 24, 92]. The large difference between the two binding pockets may result in significant conformational and energetic variations of the corresponding residues to ligand binding, and the completely different chemical structures of selective mGlu₁ and mGlu₅ NAMs (the Tanimoto coefficients between NAMs are shown in Supplementary Table S4) also allow them to selectively bind to different receptor subtypes with high sequence identity. Based on the per-residue energy analysis, 15 important energy contributors (5 “Hot Spots” and 10 “Warm Spots”) shared by four NAM–mGlu₁ complexes were identified. Sixteen residues in mGlu₅ were identified as important residues for eight clinical NAMs’ binding, including 11 major energy contributors and 5 important residues (S658^3.43, Y659^3.44, N747^5.47, T781^6.46, and S805^7.35). Therefore, a total of 18 pairs of residues, including 12 conserved ones and 6 non-conserved ones were picked out and listed in Table 2. All of them were mapped onto the structures of mGlu₁ and mGlu₅ shown in Fig. 7. Comparison of the structural and energetic information of mGlu₁ and mGlu₅ interacting with diverse NAMs, revealed the subtype-selective mechanism.

Fig. 7. The superimposition of two 7TM domains of mGlu₁ and mGlu₅.

a shows the allosteric binding pockets and 18 pairs of important residues highlighted in orange (mGlu₁) and green (mGlu₅), and b is the image rotated 180° viewed parallel to the membrane.

Table 2.

The 18 pairs of important residues in the allosteric binding pockets of mGlu₁ and mGlu₅.

Conserved residues
mGlu1	G6653.37	S6673.39	Y6723.44	V7535.40	P7565.43	L7575.44
mGlu5	G6523.37	S6543.39	Y6593.44	V7405.40	P7435.43	L7445.44
mGlu1	N7605.47	I7976.49	W7986.50	F8016.53	V8197.36	S8227.39
mGlu5	N7475.47	I7846.49	W7856.50	F7886.53	V8067.36	S8097.39
Non-conserved residues
mGlu1	V6643.36	S6683.40	C6713.43	T8157.32	A8187.35	V8237.40
mGlu5	I6513.36	P6553.40	S6583.43	M8027.32	S8057.35	A8107.40

Different properties of non-conserved residues in mGlu₁ and mGlu₅ determining the binding modes of NAMs

In Fig. 7, the upper chamber residues V664^3.36, T815^7.32, and A818^7.35 of mGlu₁ have short side chains that can provide a larger space than that of residues I651^3.36, M802^7.32, S805^7.35 in the corresponding position of mGlu₅, allowing the FITM to sit in the upper chamber. In 2019, Cong et al. speculated that M802^7.32 may be an important residue for the subtype selectivity of PAMs in mGlus [35]. Our results are consistent with previous speculation and indicate that M802^7.32 directly affects the space of upper chamber by larger side chains in the NAMs–mGlu₅ complexes. For the narrow linker and lower chamber sites, residues S658^3.43, P655^3.40, and A810^7.40 of mGlu₅ form a broad entrance into the deeper part and provide a binding pocket with more space to accommodate the NAMs, while residues C671^3.43, S668^3.40, and V823^7.40 in mGlu₁ hinder the deep binding of NAMs (Fig. 7). The structural differences between the two binding pockets result in energy contribution variations of the corresponding residues, for instance in Fig. 8 the energy contribution of S668^3.40 in mGlu₁ is obviously different from that of P655^3.40 in mGlu₅. Furthermore, ligand–receptor interactions can also be quantitatively calculated based on the presence or absence of interactions by interaction fingerprints, and the NAMs–mGlu_1/5 interaction fingerprints are depicted in Fig. 9. As shown, the key non-conserved residues in mGlu₁ and mGlu₅ determining the selectivity also present different numbers of interactions with NAMs during the simulation.

Fig. 8. Energy contributions of 18 pairs of residues in NAMs–mGlu_1/5 complexes.

a–d are four NAM–mGlu₁ complexes (FITM, BDBM50301534, EMQMCM, and JNJ16259685) shown in orange; e–l are eight NAM–mGlu₅ complexes (fenobam, HTL14242, RGH-618, VU0424238, mavoglurant, dipraglurant, basimglurant, and STX107) shown in green.

Fig. 9. Fingerprint of ligand–receptor interactions in NAM–mGlu_1/5 complexes.

a–d The FITM, BDBM50301534, EMQMCM, and JNJ16259685 bind to mGlu₁, e–l the fenobam, HTL14242, RGH-618, VU0424238, mavoglurant, dipraglurant, basimglurant, and STX107 bind to mGlu₅. The residues within 5 Å of the ligand mass center were extracted from the last 50 ns MD trajectory labeled in the radar charts, and the numbers from 0 to 1.0 indicate the probabilities of interactions between the NAM and the residues based on 500 snapshots in each complex.

In 2015, the Marshall group proposed that the substitution of residues S668^3.40 and C671^3.43 in mGlu₁ with P655^3.40 and S658^3.43 in mGlu₅ can be used to rationalize the subtype selectivity [13]. And the other residues (mGluR₁: V664^3.36, T815^7.32, A818^7.35; mGlu₅: S805^7.35, A810^7.40) were considered key factors in subtype selectivity by mutagenesis studies, and their effects were summarized in the GPCRpd database [28, 34, 93, 95]. In our models, NAMs of mGlu₁ bind in the upper chamber of the pocket, while the NAMs of mGlu₅ further deeply integrate into the narrow linker and lower chamber (Fig. 10). Therefore, it is proposed that the unequal positions of NAMs located in the two receptors determined by the different properties of non-conserved residues in the binding pockets are critical factors in the subtype-selective mechanism of mGlu₁ and mGlu₅. Except the residues reported by previous research, a total of six pairs of non-conserved residues were recognized, and their differences in terms of energy and conformation were clarified to guide further experimental study.

Fig. 10. The binding modes of NAMs in mGlu_1/5.

The NAMs FITM(orange), BDBM50301534 (pink), EMQMCM (magenta), and JNJ16259685 (read) bind in the upper chamber of the mGlu₁ allosteric site, while fenobam (green), HTL14242 (yellow), RGH-618 (cyan), VU0424238 (blue), mavoglurant (lime), dipraglurant (pea), basimglurant (wheat), and STX107 (gray) further deeply integrate into the narrow linker and lower chamber of the mGlu₅ binding pocket.

Different roles of conserved residues in mGlu₁ and mGlu₅ stabilizing the binding modes of NAMs

There are five pairs of conserved residues (mGlu₁: L757^5.44, N760^5.47, W798^6.50, F801^6.53, and V819^7.36; mGlu₅: L744^5.44, N747^5.47, W785^6.50, F788^6.53, and V806^7.36) that play important roles in the binding of NAMs to both mGlu₁ and mGlu₅ (Fig. 8). In NAM–mGlu₁ complexes, L757^5.44 and N760^5.47 (mGlu₅: L744^5.44, N747^5.47) fix the NAMs in the upper chamber by strongly hydrophobic and hydrogen interactions (Figs. 2 and 8). The residue W798^6.50 (W785^6.50) has attracted a lot attention because of its obvious conformational differences in NAM–mGlu_1/5 complexes according to the structural specificity of the NAMs (Figs. 2 and 4). In addition to the important role in ligand binding revealed in this work, W798^6.50 was considered a gatekeeper of water flux in the research of Cong et al. [35]. Residues F801^6.53 and V819^7.36 (mGlu₅: F788^6.53 and V806^7.36) are also related to the selective binding of NAMs to mGlu₁ and mGlu₅. Compared with the NAM–mGlu₅ complexes, tight interactions between F801^6.53 and four selective mGlu₁ NAMs were observed (Fig. 8). In contrast, V806^7.36 facilitates the binding of NAMs to mGlu₅, whose energy contribution is significantly higher than the corresponding residues V819^7.36 in mGlu₁ (Figs. 8 and 9). Therefore, the enhancement of the interaction between NAMs and conserved residues L757^5.44, N760^5.47, and F801^6.53 in mGlu₁ (mGlu₅: L744^5.44, N747^5.47, and F788^6.53) promotes the binding of NAMs to mGlu₁, while NAMs with better selectivity for mGlu₅ should form tight interaction with V806^7.36 (V819^7.36 in mGlu₁). In 2015, the Gloriam group reported that G628^2.49 is a selective hotspot in mGlu₅. Unfortunately, it was not identified in this study, due to the ordinary energy contribution and the similar conformation in mGlu₁ and mGlu₅ (a conserved residue in mGlu₁ and mGlu₅). However, compared with the other mGlu members in groups II and III, Gly628 is a non-conserved residue, which seems to be a key factor of selectivity between group I and other receptors.

Binding mechanism of the dual-target NAM VU0467558

VU0467558 is a novel NAM that exhibited near equipotent activity against both mGlu₁ and mGlu₅ [38]. The binding mechanism of VU0467558 was predicted to verify the selective features identified in this study and provide information for the design of dual-target NAMs. Two docking poses of VU0467558 binding in mGlu₁ and mGlu₅ have been studied by the same approach, respectively. As shown in Supplementary Fig. S5, the binding conformations of VU0467558 in pose 1 (cluster 1) have significant energy differences over those of pose 2 (cluster 2) (6.21 and 4.65 kcal/mol). The positions of VU0467558 in the two allosteric sites are obviously different. In detail, VU0467558 binds in the upper chamber of the mGlu₁ pocket, and is located deeper in the mGlu₅ pocket. The large differences between the two binding pockets result in significant energetic variations of the corresponding residues to VU0467558 binding, such as I638^2.46, S667^3.39, and S822^7.39 in mGlu₁ with the corresponding residues I625^2.46, S654^3.39, and V809^7.39 in mGlu₅. The six identified pairs of non-conserved residues (mGlu₁: V664^3.36, S668^3.40, C671^3.43, T815^7.32, A818^7.35, and V823^7.40; mGlu₅: I651^3.36, P655^3.40, S658^3.43, M802^7.32, S805^7.35, and A810^7.40) shaping the pockets for specific NAM binding were confirmed by structural superimposition of VU0467558 in complex with mGlu₁ and mGlu₅ (Fig. 11). Notably, the conformation of residues V664^3.36 and T815^7.32 in mGlu₁ can provide a larger space than that of residues I651^3.36 and M802^7.32 in the corresponding position of mGlu₅, allowing the tricyclic group of VU0467558 to sit in the upper chamber. The conformation of S805^7.35 moved to the outside of the binding pocket to accommodate the binding of VU0467558 in mGlu₅, compared to A818^7.35 in mGlu₁. On the other hand, S658^3.43, P655^3.40, and A810^7.40 in mGlu₅ form a broad entrance allowing the small moiety of VU0467558 to insert into the deeper part of the binding site and provide the binding pocket more space to accommodate the NAM, but residues C671^3.43, S668^3.40, and V823^7.40 in mGlu₁ hinder this deep binding. In addition, the energy contributions of several residues also demonstrated the binding difference of VU0467558 in mGlu₁ and mGlu₅. The roles of the conserved residues L757^5.44, N760^5.47, F801^6.53, and V819^7.36 in mGlu₁ (corresponding to L744^5.44, N747^5.47, F788^6.53, and V806^7.36 in mGlu₅) were observed in the predicted binding modes of VU0467558 binding to the two receptors. As shown in Fig. 11c, residues L757^5.44, N760^5.47, F801^6.53, and V819^7.36 contributed −3.09, −1.85, −1.71, and −1.06 kcal/mol for VU0467558 binding in mGlu₁, while the energy contributions of the corresponding residues in mGlu₅ were −1.06, −0.59, −1.19, and −2.06 kcal/mol, respectively. Moreover, the energy contributions of nine pairs of residues (mGlu₁: V664^3.36, S668^3.40, V753^5.40, L757^5.44, N760^5.47, W798^6.50, F801^6.53, T815^7.32, and V819^7.36; mGlu₅: I651^3.36, P655^3.40, V740^5.40, L744^5.44, N747^5.47, W785^6.50, F788^6.53, M802^7.32, and V806^7.36) are both important in mGlu₁ and mGlu₅, which are the key factors in designing new dual-target NAMs.

Fig. 11. Representative structures and per-residue decomposition energies of VU0467558-mGlu_1/5 complexes.

a is the representation of VU0467558 binding in mGlu₁ (orange) and mGlu₅ (green), and b is the image rotated 180°. The important residues identified as critical features of the subtype-selective mechanism are shown as sticks. c is the per-residue energy contribution of VU0467558 binding to mGlu₁ (orange) and mGlu₅ (green).

Conclusion

In this work, a computational modeling approach of molecular docking and MD simulation were utilized to explore the binding mechanism of different single- and dual-target NAMs in group I mGlus and clarify the subtype-selective mechanism in detail. The selective mGlu₁ NAMs bind in the upper chamber of the allosteric site, and the binding pocket partially overlaps with mGlu₅ in the upper chamber. In the mGlu₅ receptor, the novel non-alkyne compounds are located in the same binding pocket as the alkyne NAMs, deeply inserted into the narrow linker and lower chamber. By comparing the structural and energetic profiles underlying the binding of twelve NAMs with diverse scaffolds and different activities to mGlu₁ and mGlu₅, it was found that ten pairs of residues including six non-conserved residues (mGlu₁: V664^3.36, S668^3.40, C671^3.43, T815^7.32, A818^7.35, and V823^7.40; mGlu₅: I651^3.36, P655^3.40, S658^3.43, M802^7.32, S805^7.35, and A810^7.40) and four conserved residues (mGlu₁: L757^5.44, N760^5.47, F801^6.53, and V819^7.36; mGlu₅: L744^5.44, N747^5.47, F788^6.53, and V806^7.36), arounding the binding pockets are essential to receptors subtype selectivity. The non-conserved residues mainly affect ligand binding through spatial conformation, and the four conserved residues enhance the selective ligands binding by participating in strong interactions. This information should be considered in the design of selective NAMs. In addition, the same dual-target NAM (VU0467558) adopted completely different conformations in the two receptors verifying the results of our computational model. Moreover, the nine pairs of residues (mGlu₁: V664^3.36, S668^3.40, V753^5.40, L757^5.44, N760^5.47, W798^6.50, F801^6.53, T815^7.32, and V819^7.36; mGlu₅: I651^3.36, P655^3.40, V740^5.40, L744^5.44, N747^5.47, W785^6.50, F788^6.53, M802^7.32, and V806^7.36) were recognized as the key factors in designing new dual-target NAMs of mGlu₁ and mGlu₅. The results of this research can provide theoretical guidance for the development of novel NAMs and also provide valuable information on the mechanisms of class C GPCR allosteric inhibition.

Author contributions

WWX and FZ designed the research. TTF, GT, and MP performed the research. TTF, GXZ, FYY, JYY, YZ, XJY, and WWX, analyzed the data. TTF, WWX, and FZ wrote the manuscript.

Competing interests

The authors declare no competing interests.

Supplementary information

The online version of this article (10.1038/s41401-020-00541-z) contains supplementary material, which is available to authorized users.