Impact of pathogenic mutations of the GLUT1 glucose transporter on solute carrier dynamics using ComDYN enhanced sampling

Starting structures

The crystal structures of both I_O and O_O states are available in the PDB; 5EQI for GLUT1 I_O state (Deng et al., 2015), and 4ZW9 for GLUT3 O_O state (Kapoor et al., 2016); these were used as starting points for the simulations, as we focus on the first part of the transport mechanism. The next step, i.e., to elucidate the complete pathway, would be to extend our work to the outward-open conformation which is also available in the PDB (4ZWC GLUT3 with maltose bound), and to study the effect of ligand binding on the transporter dynamics. However, this is beyond the scope of the present study. Based on the reported pathogenic mutations of GLUT1 in the conserved RXGRR motif region, we searched the corresponding positions of GLUT3 for additional mutations and included them in our study. Due to the high sequence identity (~70%) between the proteins, we intentionally did not build a homology model for one or the other protein. This is justified, as our main aim is to characterize the global opening and closing mechanism rather than to look into atomistic details such as protonation states. Moreover, we now avoid additional uncertainties about details of the structure as would be inevitably introduced during the homology modelling process. Our composite scheme using coarse-grained molecular dynamics with common constraints elastic network ‘ComDYN’, that keeps part of the unchanged position constraints between the two protein structures, is explained in the approach below and further details are supplied in the supporting methods, available in the deposited code (Feenstra, 2019b).

Molecular dynamics simulations

In this study, we employed molecular dynamics (MD) simulations using the GROMACS 4.0.5 programme package (Hess et al., 2008). For efficiency reason, we investigated the applicability of the MARTINI coarse-grained (CG) force field (Arnarez et al., 2015; de Jong et al., 2013; Hsu et al., 2017; Monticelli et al., 2008; Periole et al., 2009), which is about 500-fold faster than the full-atomistic GROMOS (May et al., 2014). This speed-up is obtained at the expense of explicit description of hydrogen bonds in the protein, which necessitates the addition of an elastic network to maintain secondary structure and other tertiary contacts and thus overall protein stability; the elastic network is thus tailed to a particular conformational state of the protein. Here, we propose to modify the elastic network that is used in MARTINI based on our starting conformations, to include only common elastic network constraints that differed less than 1Å between the I_O and O_O states. Increasing the distance threshold for the common constraints would result in even fewer constraints in the ComDYN, increasing the risk of instability and possibly unfolding of the protein. Therefore, our chosen constraints allow transition between both states, while maintaining protein structure stability. This composite scheme will henceforward be referred to as ComDYN (COMmon constraints DYNamics). The detailed computational set-up is provided in the Supporting Methods (Feenstra, 2019b).

Analysis of transporter dynamics

Essential dynamics analysis was performed on the GLUT1 and GLUT3 simulations using the built-in analysis tools of GROMACS. To allow this comparison between these two homologous proteins, and allow for focusing on overall motions of the transporter region, we selected the structurally conserved helical segments, as summarized in Supporting Table S2, available as extended data (Feenstra, 2019a). Then, the covariance and eigenvalue calculation were performed on the ensemble of both wild-type systems, using the coordinates of the C-alpha atoms for the full atomistic (AT) simulations, and the backbone particle at the C-alpha position for the coarse-grained (CG) and ComDYN simulations.

To analyse the transporter dynamics from the ComDYN, we defined several order parameters as previously proposed by Nagarathinam et al. (2018) by measuring distances between adjacent TM helices, at the intracellular (in)- and extracellular (out)sides of the protein (see Figure 1B, C). For each of the ring of six central helices that make up the solute transporter region, TM2, TM1, TM5, TM8, TM7, TM11 (and back to TM2), we defined an ‘inside’ and ‘outside’ segment of ten residues (See Figure 1 and extended data, Supporting Figure S3 (Feenstra, 2019a)). Comparing the distances of the mutations to both wild types allows us to capture abnormal behaviour and identify the mutations that have the highest impact on the opening and closing mechanism of the apo-form of the transporter protein (see Results and Discussion).

To compare the distributions of sampled distances during the simulations, two metrics were employed: overlap and shift. ‘Overlap’ is the fraction of overlap between both distributions, as the integral of the minimum of both functions. ‘Shift’ quantifies the direction of change, and is obtained by taking the difference in the position of the peak of the two distributions from the ensemble of the simulations (typically, mutation version wild-type). Negative indicates a ‘closing’ motion, positive is ‘opening’. This analysis was performed using the script ‘calc_overlap.py’, which may be found in the extended data.

Prediction of mutation impact

Table 1 lists the mutants of GLUT1 and GLUT3 that we will consider here, which are situated in the conserved RXGRR-motif distal to the transporter region. The table lists the predicted impact upon mutation obtained from SIFT (Sim et al., 2012) and PolyPhen-2 (Adzhubei et al., 2010). Most mutations are classified as likely pathogenic by both methods, with the exception of GLUT1 R93Q, and GLUT3 R91C and R91H. However, these methods are trained on the dbSNP database which also includes these known mutations, so this should be no surprise. Moreover, these predictions do not allow us to gain any insights into the mechanism by which these mutations may affect transporter function.

Verification of constraining approach

Firstly, we want to verify if our common constraints-based approach for coarse-grained MD simulations (ComDYN) is able to sample the intermediate states between the I_O and O_O states. Including only elastic network constraints that differed less than 1Å between the I_O and O_O states in ComDyn, resulted in 1025 constraints (39.9% of the 2568 in the Martini elastic network) for the O_O state and 978 (42.2% of 2315) for the I_O state. We performed essential dynamics (ED) analysis (Amadei et al., 1993; Van Aalten et al., 1997) to compare AT, traditional CG MARTINI, and ComDYN simulations, as described in the Methods. Figure 2 shows 2D plots of the first two (largest) ED eigenvectors, representing the extracted correlated motions over the ensemble of our simulations. The first eigenvector (EV1) represents the major conformational changes between the O_O and I_O states, as also indicated by the RMSD values between starting states. The changes observed along EV2 may also be genuinely part of the state transition, however we cannot exclude the possibility that some of these conformational changes may relate to differences between the forcefields used. The sampling of the different states, inward-open and outward-occluded, in AT simulation hardly converges due to their limited time scales. The regular CG simulations reach much longer timescales, and already sample more intermediate conformations, but there is no overlap. The ComDYN simulations, on the other hand, also sample many states intermediate to the inside-open and outside-open starting states, compared to the other simulations. This shows that improved sampling of large conformational transitions may be attained using this approach of CG and ComDyn simulations.

Figure 2. Two-dimensional essential dynamics plot of the simulations.

In this projection, eigenvector 1 corresponds to changes from O_O (left) to I_O (right), showing also the overlap and differences between the AT (full atomistic), the CG (coarse-grained), and the ComDYN (common constraints CG) simulations. Note that there is considerable overlap in the sampling, but that the time-scale of the AT simulations only samples conformations around the I_O and O_O states and the elastic network in the CG simulations also limits the visited conformations, while the ComDYN samples a large number of conformations between both states.; I_O, inward open state (PDB-ID: 5EQI); O_O, outward occluded state (PDB-ID: 4ZW9). Spheres indicate the respective starting conformations for each method, triangles to show the corresponding crystal structures; black arrows and labels indicate the RMSD between the starting conformations of AT and ComDYN simulations.

Probing conformational changes

To probe the degree of the conformational changes during ComDyn simulations of the wild types and the mutants in more detail than done with the ED analysis, two distances were used to describe the opening and closing of the periplasmic and cytoplasmic sides of the transporter. Nagarathinam et al. (2018) studied a bacterial homolog of GLUT1 and GLUT3, and analysed the movement between TM5 and TM8. In the extended data, Figure S4 (Feenstra, 2019a), we can see that the distributions obtained from our ComDYN simulations, resemble those reported by Nagarathinam et al. (2018), providing an independent validation that our ComDYN approach is able to sample biologically relevant conformational states for large scale motions, such as those involved in the glucose transporter mechanism. Nevertheless, there are differences between the distance distributions in our work and that of Nagarathinam et al. (2018), which could, apart from obvious differences comparing human glucose transporters with a bacterial multidrug transporter (see also extended data, Figure S4 and Table S3 (Feenstra, 2019a)), also arise from differences in the sampling protocols applied in the two studies.

Therefore, in addition to the TM5-TM8 distances, we extended the analysis to other helices along and across the rim that make up for the entire SLC transporter architecture, allowing us to monitor changes in their position (see Figure 1C, D). For each of these order parameters, we calculated the distance at the inside and outside of the protein with respect to the membrane. Using this analysis, we can immediately observe the changes occurring between the inward-open and outward-occluded states. We see several distances changing significantly during this process: TM1/TM2(in), TM1/TM5(out), TM1/TM7(out), TM1/TM8(in), and TM5/TM11(in) are all closing, while TM1/TM2(out), TM1/TM5(in), TM1/TM7(in), TM2/TM11(in), TM2/TM8(out), and TM5/TM11(out) are opening (see extended data, Table S3 and Figure S4 for more details (Feenstra, 2019a)); these motions are also schematically summarised in extended data, Figure S3 (Feenstra, 2019a). Table 2 summarises the overlap and shift between the sampled distributions of distances between the wild type and each of the mutants for TM5 and TM11 that exhibited the strongest effects and conformational changes during the simulations (the complete table of the distributions for all order parameters are available in the extended data, Table S3 (Feenstra, 2019a)). The corresponding conformational distributions from the ComDYN simulations, calculated as a function of the inner and outer distances between TM5 and TM11 are given in Figure 3.

Figure 3. Distance plots of the inner and outer distances along the order parameter TM5-TM11 in nm over the complete time span of the CG ComDYN simulations (see Figure 1D).

Colour code: wild types I_O in purple (PDB-ID: 5EQI), O_O in green (PDB-ID: 4ZW9), mutants in orange. Pathogenic mutants are highlighted in red. It should be noted that in contrast to the benign R93Q mutant, the pathogenic mutants do not sample the I_O and O_O states during the simulation, which strongly indicates that the mutation blocks the proper opening and closing mechanism. Corresponding plots for all mutations are in the extended data, Figure S4. The corresponding quantification of these plots provided as shift and overlap are given in Table 2.

Impact of mutations on dynamics

Not all mutations have a high impact on the overall dynamics (extended data, Table S3 and Figure S4 (Feenstra, 2019a)). However, in GLUT1, the reported pathogenic mutation G91D has a profound effect on the dynamics of the protein (i.e., a low overlap and large shift, see Table 2). Also when we consider the distance of TM5-TM11, as shown in Figure 3, the strongest effect is observed for the pathogenic G91D mutant: its distribution varies strongest from the wild types, and hardly visits the inward-open and outward-occluded states. Furthermore, Figure 3 shows that for the pathogenic R92W and R333W mutations, only one state or small parts from both can be accessed. For the benign mutant R93Q, in contrast, it can be seen that both states, inward-open and outward-occluded, are sampled thoroughly during the simulations. Assuming that the relative distance between the two helices is crucial for the correct functioning, this strongly suggests that the pathogenic mutations directly affect the opening and closing mechanism of the GLUT1 transporter.

Mutations in GLUT3 show similar behaviour in TM dynamics compared to those in GLUT1. Here, two mutants with strong abnormal behaviour can be identified: G89V and R91H (Table 2; extended data Figure S4 (Feenstra, 2019a) shows the corresponding distance distribution plots). Additionally, similar to the observations for pathogenic mutations on GLUT1, these mutations no longer sample intermediate states associated with the transport function, unlike the wild type and many of the other mutations. This strongly suggests that the corresponding mutations between GLUT1 and GLUT3 also have the same direct blocking effects on the opening and closing mechanism of the GLUT3 transporter. However, it should be noted that we cannot make any conclusions about the clinical significance of these GLUT3 mutants, as none have been reported to be pathogenic. A next step to elucidate the complete pathway would be to extend the analysis to the outward-open conformation, and include ligand-bound states (Delemotte, 2019), which may shed further light on some of the currently still unexplained pathogenic mutations.

Underlying data

Crystal structures for GLUT1 I_O state (Deng et al., 2015) and for GLUT3 O_O state (Kapoor et al., 2016) were obtained from the Protein Data Bank, under accession numbers 5EQI and4ZW9, respectively.

Extended data

Open Science Framework: ComDYN. https://doi.org/10.17605/OSF.IO/F82H5 (Feenstra, 2019a).

The following extended data are available:

Extended data are available under the terms of the Creative Commons Zero "No rights reserved" data waiver (CC0 1.0 Public domain dedication).