Characterizing the transmembrane domains of ADAM10 and BACE1 and the impact of membrane composition

All-atom proteins in implicit membranes

For each protein, transmembrane congener models with JMDs of varying length were simulated using REMD in implicit membranes with thicknesses of 30, 35, and 40 Å described by the generalized Born with simple switching (GBSW) potential in the CHARMM simulation program (41,42,43). This approach follows that taken in our previous work exploring the structure of APP-C99 (44). The congener models used for BACE1 included residues M440–K501, Q449–K501, and Q449–H490; and the congener models for ADAM10 included residues M643–P715, L654–L710, and E665–T703 (Fig. 1 B–C). Each model was constructed by placing the residues in a completely α-helical structure and centered in the GBSW membrane. The proteins were represented with the CHARMM36m force field (45). The GBSW membranes used a 6-Å smoothing length with 38 angular integration points and a maximum distance for radial integration of 20 Å. The nonpolar surface tension coefficient was set to 0.04 kcal/mol/Ų. The membrane/solvent interface on each side of the membrane was set to 2.5 Å, leaving the hydrophobic cores of the 30-, 35-, and 40-Å membranes as 25, 30, and 35 Å, respectively. Replica exchange was attempted every 50 ps between 16 linearly spaced windows from 310K to 460K. Frames were saved every 10 ps for around 200 ns, and the frames from the last 100 ns from the 310K window were used for analysis.

Coarse-grained simulations

The Martini2.2 coarse-grained protein force field is dependent on a predefined secondary structure (46,47,48). To identify a suitable secondary structure, the final 100 ns of frames from the 310K window of the REMD simulations of the selected congener models were clustered with agglomerative clustering using Ward’s minimum variance method (49). The optimal number of clusters was taken to be that with the highest Silhouette Score (50). This clustering scheme was applied using Julia with the Clustering.jl package and is further described in supporting methods I (51). For each protein and membrane width, the medoid of the largest cluster was mapped to the Martini2.2 model using the martinize.py program with a manually defined secondary structure (46,47,48). The secondary structures of the α-helical residues in the proteins were defined based on the averaged backbone hydrogen bond pattern of the frames in the largest cluster, and for the rest of the residues, they were defined using the DSSP result for the input structure (52).

For each protein, its Martini2.2 congener model was simulated in five single-component lipid membranes: 100 mol% DPPC, 100 mol% DIPC, 100 mol% dilauroylphosphatidylcholine (DLPC), 100 mol% dioleoylphosphatidylcholine (DOPC), and 100 mol% 1-palmitoyl-2-oleoylphosphatidylcholine (POPC). The models were also simulated in DIPC and DPPC membranes with four different concentrations of CHOL: 10 mol% CHOL, 20 mol% CHOL, 30 mol% CHOL, and 40 mol% CHOL. The congener model for each protein and membrane was selected based on the width of a pure lipid membrane composed of the lipid included in the membrane (Table S1). The systems were built using the insane.py program (53).

Five replicates were simulated for each system in GROMACS 2021.5 using the new-rf input parameters recommended by the Martini developers (54,55). Each replicate was first minimized by steepest descent and then equilibrated in the isobaric-isothermal (NPT) ensemble using the velocity rescaling thermostat and a semiisotropic Berendsen barostat with a compressibility of 4.5 $\times$ 10⁻⁵ bar⁻¹ and an integration timestep of 20 fs (56). Production simulations were performed for 5 μs in the NPT ensemble using a semiisotropic Parrinello-Rahman barostat with a compressibility of 3.0 $\times$ 10⁻⁴ bar⁻¹(57). Coordinates were saved every 1 ns, and the last 4 μs of each simulation was used for analysis. All simulations were performed at 310K with an ion concentration of 0.15 M and with 10% of water represented with Martini’s antifreeze water model.

All-atom proteins in explicit membranes

Final configurations from each 5-μs Martini2.2 replicate simulation of each protein in 7:3 DPPC:CHOL and 9:1 DIPC:CHOL membranes were backmapped to the CHARMM36 force field using the backwards.py program (58,59). Note that due to Martini2.2’s one-to-four bead per atom representation, individual Martini lipids can represent multiple all-atom lipids. Here, the DIPC lipids were backmapped to di-C18:2 PC lipids and the DPPC lipids were backmapped to di-C16:0 PC lipids. Also note that di-C18:2 PC lipids are named DUPC in the CHARMM36 force field, but in this paper, we will continue to refer to them as DIPC for clarity. Backmapping was completed using a revised version of the initram-v5.sh program. The initram-v5.sh program is available with backwards.py on Martini’s web site. The only revisions to initram were to use the Verlet cutoff scheme for neighbor searching and soft-core potentials to make it compatible with our single-precision version of GROMACS. Similar revisions are discussed and utilized in Martini’s backmapping tutorials. An additional 2 nm of water was added along the z axis (perpendicular to the membrane) to the ADAM10 in the 9:1 DIPC:CHOL system using Packmol to prevent interperiodic contact between the JMDs (60). Each system was equilibrated for 5 ns in the NVT ensemble with the velocity rescaling thermostat and a 1-fs timestep and for 10 ns in the NPT ensemble with a 1-fs timestep, the velocity rescaling thermostat, and the semiisotropic Berendsen barostat with a compressibility of 4.5 $\times$ 10⁻⁵ bar⁻¹(56,61). Production simulations were performed using the Nose-Hoover thermostat and the semiisotropic Parrinello-Rahmen barostat with a compressibility of 4.5 $\times$ 10⁻⁵ bar⁻¹ for 1 μs (57,62,63). Coordinates were saved every 100 ps and the last 900 ns from each simulation was used for analysis. All simulations were performed using GROMACS 2021.2 at 310K with an ion concentration of 0.15 M (54,55).

Development of BACE1 and ADAM10 TMD and JMD models

REMD simulations were performed to sample the conformational space of several length congener models of ADAM10 and BACE1 in 30-, 35-, and 40-Å GBSW implicit membranes. These simulations were used to identify minimal length congener sequences having consistent structural features in the three membranes to be used in explicit membrane simulations. Model validation principally focused on two metrics: residue α-helix propensity, $p_{\alpha }$, and insertion depth, $D_{ins}$. Residue α-helix propensity was calculated as the probability that the $i^{th}$ residue forms an amine-carbonyl backbone hydrogen bond with residue $i+4$ and/or residue $i-4$. Hydrogen bonds were defined as having a donor-acceptor distance of $\le 3.5$ Å and a donor-hydrogen-acceptor angle $\le {150}^{\circ }$. Insertion depths were calculated as the nearest distance from the edge of the GBSW membrane.

For ADAM10, we evaluated model congener sequences including residues M643–P715, L654–L710, and E665–T703. The shortest model (E665–T703) features an extended α-helix and does not include reentrant residues within the N-terminal JMD (Fig. 2 A). We selected the L654–L710 model for further simulation to account for effects of interactions between the JMD and the membrane. In our REMD simulations of the ADAM10 congener models, residues H677 and W673 can form a sidechain-backbone hydrogen bond and induce a hinge in the TMD α-helix. From this point on, the L654–L710 model of ADAM10 will be referred to as ADAM10.

For BACE1, we evaluated three model congener sequences: M440–K501, Q449–K501, and Q449–H490. Helicity and insertion depths appear consistent across all three models in all membranes (Fig. 2 B). Therefore, the Q449–K501 model is the optimal model sequence, as including the last 11 residues of BACE1’s C-terminal domain provided a more complete model congener sequence with minimal added computational cost. This BACE1 model contains four cysteine residues known to be S-palmitoylated (C474, C478, C482, and C485), five sulfur containing residues believed to contribute to BACE1 trimerization (M462, C466, M470, C474, and C478), and P472, which induces a kink in the α-helix. From this point on, the Q449–K501 model of BACE1 will be referred to as BACE1.

Agglomerative clustering was performed on each REMD simulation, and the largest cluster was used to define the secondary structures of the selected models for Martini2.2 simulations (Figs. 2, 3, and 4). The medoid structures of the largest clusters (Fig. 2) were used as the input structure for mapping from all-atom to Martini representation.

Characterization of the ADAM10 TMD and JMD

The ADAM10 congener model was simulated in a variety of coarse-grained and all-atom lipid and lipid/CHOL bilayers. Insertion depths and α-helix propensities were calculated to determine the TMD residues. Insertion depths were calculated as the distance along the normal of the nearest leaflet between a residue’s α-carbon (or BB bead for Martini) and the leaflet (supporting methods II). The position and normal of each leaflet were determined using the lipid phosphate phosphorus atoms within 2 nm of the any protein atom. Residue α-helix propensities were calculated in the same manner as with the REMD simulations. Here, we define the TMD as the continuous chain of residues passing through the bilayer with $D_{ins}\le 0$ and the TMD α-helix as the continuous chain of residues with $p_{\alpha }\ge 0.5$.

To understand how the ADAM10 TMD and JMDs interact with and anchor to the membrane, the average number of hydrogen bonds, $⟨N_{HB}⟩$, formed between each residue and each kind of membrane component were computed for each of the all-atom explicit membrane simulations. Potential hydrogen bond acceptors were identified as atoms whose charge $\le -0.5e$, potential hydrogen atoms involved in hydrogen bonds were identified as those with a charge $\ge 0.3e$, and potential hydrogen bond donors were inferred from the potential hydrogens. Like with backbone hydrogen bonds, protein-membrane hydrogen bonds were defined as having a donor-acceptor distance of $\le 3.5$ Å and a donor-hydrogen-acceptor angle $\le {150}^{\circ }$.

For ADAM10 in a 9:1 DIPC:CHOL all-atom bilayer, the TMD extends from residues I670 to H702, and the α-helix includes residues N669–C699 (Fig. 3 A). In a 7:3 DPPC:CHOL bilayer, the ADAM10 TMD extends from residues I670 to P704, and residues N669–C699 make up the the α-helix. Similar results were found for the TMD in coarse-grained simulations with DIPC and DPPC membranes with various CHOL concentrations and in pure DOPC, DLPC, and POPC membranes (Fig. S5). These results show that the length of the ADAM10 TMD and TMD α-helix are not substantially impacted by membrane lipid and CHOL concentrations.

The tilt angle, θ, between ADAM10’s α-helix axes and the membrane normal and the residues’ mean orientational angles, $⟨{\sigma }_i⟩$, were calculated for coarse-grained DIPC:CHOL and DPPC:CHOL systems with 0%, 10%, 30%, and 40% CHOL and for the all-atom systems with 9:1 DIPC:CHOL and 7:3 DPPC:CHOL membranes. The ${\sigma }_i$ angles describe the orientation of a helical residue’s α-carbon in the membrane about the axis of the α-helix, and the θ angle is measured as the angle between the α-helix axis and the membrane normal (Fig. 4 D). Further details about the θ and ${\sigma }_i$ calculations are provided in supporting methods II.

For ADAM10 in DIPC membranes, the tilt angle decreases linearly with increasing membrane width. In contrast, for DPPC membranes with 30% and 40% CHOL (consistent with concentrations found in L_o membrane domains), ADAM10’s tilt angle distribution becomes bimodal, indicating two distinct configurational states (Fig. 4 A). This bimodality also appears in all-atom 7:3 DPPC:CHOL membrane simulations of ADAM10 and can accounted for by the number of hydrogen bonds between ADAM10’s C-terminal JMD and membrane components and the presence of a hydrogen bond between the sidechain of H677 and the backbone of W673 (Fig. 4 B). Higher tilt angles tend to be associated with configurations with more than two hydrogen bonds between the C-terminal JMD and the membrane ($\overline{\theta }\approx {36}^{\circ }$) and/or with a sidechain-backbone hydrogen bond between H677 and W673 ($\overline{\theta }\approx {37}^{\circ }$). Lower tilt angles are most strongly associated with configurations lacking any hydrogen bonds between the C-terminal JMD and the membrane ($\overline{\theta }\approx {26}^{\circ }$). We observe the impact of hydrogen bonds between the protein and the membrane on the tilt angle in the 7:3 DPPC:CHOL membrane but not in the 9:1 DIPC:CHOL membrane. This does not appear to be a product of hydrogen bonding with CHOL as, relative to lipid-protein hydrogen bonds, few of these interactions occur in either membrane (CHOL-protein hydrogen bonds account for only 5.7% and 9.6% of membrane-protein hydrogen bonds in the 9:1 DIPC:CHOL and 7:3 DPPC:CHOL membranes, respectively). Despite the observed change in tilt angle between DIPC and DPPC membranes, the difference in $⟨{\sigma }_i⟩$ from the 9:1 DIPC:CHOL all-atom to the 7:3 DPPC:CHOL all-atom membrane is unchanged (10 $\pm$ 36°; Fig. 4 C). On the N-terminus side, the TMD is anchored to the membrane primarily by residues L654, R656, and K658 (Fig. 3). On the C-terminus side, the TMD is anchored to the membrane by K697 and nonspecifically by various residues in the C-terminal JMD (Fig. 3).

To understand whether specific interactions with CHOL, other than just hydrogen bonding, could impact the character of ADAM10 in L_o domains, we analyzed residue-CHOL contacts with a residence time of greater than 25 ns in the all-atom 9:1 DUPC:CHOL and 7:3 DPPC:CHOL systems (Figs. S10A and S11). Within the C-terminal side leaflet, strong interactions occur between CHOL and F695. It is possible that in CHOL-rich L_o domains, these interactions lead to an unfavorable decrease in interactions between the C-terminal JMD and lipids and cause the previously described low-tilt state. Meanwhile in the N-terminal leaflet, interactions between H677, W678, and W679 and CHOL could impact the H677–W673 sidechain-backbone hydrogen bond, and interactions between F662 and CHOL could cause the slight decrease in hydrogen bonding between N-terminal JMD residues L654 and R656 observed in the 7:3 DPPC:CHOL system compared with the 9:1 DUPC:CHOL system.

Principal component analyses were performed for the all-atom 9:1 DIPC:CHOL and 7:3 DPPC:CHOL membrane simulations using the RMSD-fitted α-carbon cartesian positions of ADAM10. Principal component analyses have been applied for similar systems to identify critical global protein motions and can be easily applied using the Prody python package and other programs (64). Here, we instead used in-house code described in supporting methods III (65,66,67). Cross correlation matrices were constructed for each membrane using enough principal components to account for 90% of the cumulative variance. For the 9:1 DIPC:CHOL membrane and the 7:3 DPPC:CHOL membrane, the cumulative variances of the first 11 principal components are 90.8% and 91.2%, respectively. The cross correlation matrices of these principal components and the first two principal components for each system are shown in Fig. 5.

In both 9:1 DIPC:CHOL and 7:3 DPPC:CHOL membranes, the principal component analysis reveals strong correlations between helical residues and residues one helical turn later and a break in the helix between residues W673 and H677. This kink in the helix is more significant in the 7:3 DPPC:CHOL membrane than in the 9:1 DIPC:CHOL membrane. Further, in the 9:1 DIPC:CHOL membrane, residues in helical groups 2 (V675–A687) and 3 (L688–C699) are well correlated, whereas in the 7:3 DPPC:CHOL membrane, no significant correlation is observed. Finally, the principal component analysis further substantiates our previous TMD α-helix definition for ADAM10 of N669–C699 in both membranes.

Characterization of the BACE1 TMD and JMD

Following the protocol applied for ADAM10, the BACE1 congener model was simulated in a variety of coarse-grained and all-atom lipid and lipid/CHOL bilayers, and residue insertion depths and α-helix propensities were calculated. We define the TMD as the continuous chain of residues passing through the bilayer with $D_{ins}\le 0$. The TMD α-helix was defined as the continuous chain of residues with $p_{\alpha }\ge 0.5$. To understand how BACE1’s TMD and JMDs interact with the membrane, the average number of hydrogen bonds between each residue and each kind of membrane component were computed for each all-atom simulation. Hydrogen bonds were identified in the same manner as with ADAM10.

For BACE1 in a 9:1 DIPC:CHOL all-atom bilayer, the TMD extends from residues E452–C485 and the α-helix is made up of residues Q449–R484 (Fig. 6 A). In a 7:3 DPPC:CHOL all-atom bilayer, residues D451–Q479 make up the α-helix, and the TMD includes residues E452–R481 (Fig. 6 A). In this membrane, residues C482–L486 are also a part of the continuous chain of residues passing through the membrane with $D_{ins}\le 0$. However, as they are not a part of the α-helix and lie flat along the lipid headgroups, they are better classified as part of the C-terminal JMD. These all-atom results are in agreement with corresponding coarse-grained simulations with DIPC and DPPC membranes containing 0%, 10%, 20%, 30%, and 40% CHOL and in pure DOPC, DLPC, and POPC membranes (Fig. S6), suggesting that the length of BACE1’s TMD and α-helix is dependent upon membrane composition due to structural changes in the membrane rather than specific interactions with individual membrane components.

The tilt angle, θ, and residue orientational angles, ${\sigma }_i$, were calculated for BACE1 in 9:1 DIPC:CHOL and 7:3 DPPC:CHOL all-atom bilayers, and θ was calculated for coarse-grained DIPC and DPPC bilayers with a variety of CHOL concentrations. For BACE1, the tilt angle decreases linearly with increasing membrane width for both DIPC and DPPC membranes (Fig. 7 A). BACE1’s shorter TMD α-helix in DPPC membranes leads to a more pronounced decrease in tilt angle compared with DIPC membranes than what was observed for ADAM10 (Fig. 4 A). In both membranes, BACE1 is anchored to the membrane by hydrogen bonds between the membrane and R481, R484, and R487, belonging to an RCLRCLR motif. However, compared with in the 9:1 DIPC:CHOL membrane, hydrogen binding propensity with the membrane is shifted from R481 to R484 in the 7:3 DPPC:CHOL membrane (Fig. 6). This change in anchoring residue provides a plausible explanation for the shorter TMD α-helix observed in the wider 7:3 DPPC:CHOL membrane and results in a change in the residue orientational angles (Fig. 7 B). The average change in $⟨{\sigma }_i⟩$ from the 9:1 DIPC:CHOL membrane to the 7:3 DPPC:CHOL membrane is −58 $\pm$ 13°, corresponding to the helical twist between the anchoring residues R481 and R484.

To further understand how CHOL impacts the character of BACE1, we examined specific residue-CHOL contacts lasting longer than 25 ns in the all-atom 9:1 DUPC:CHOL and 7:3 DPPC:CHOL systems (Figs. S10B and S12). Within the N-terminal leaflet, these contacts appear to be facilitated by Y460, whose α-carbon is positioned on the side of the helix pointing slightly into the membrane in the 9:1 DUPC:CHOL system ($⟨{\sigma }_i⟩\approx$ −109°) and more significantly into the membrane in the 7:3 DPPC:CHOL system ($⟨{\sigma }_i⟩\approx$ −161°). On the C-terminal side of the helix, protein-CHOL contacts appear to be facilitated by F469 and W480. In the 9:1 DUPC:CHOL membrane, F469 ($⟨{\sigma }_i⟩\approx$ 100°) and W480 ($⟨{\sigma }_i⟩\approx$ 107°) are both on the side of the helix pointing slightly toward the outside of the membrane. In the 7:3 DPPC:CHOL system, F469 points in toward the center of the membrane ($⟨{\sigma }_i⟩\approx$ 35°), and W480 becomes a part of the C-terminal JMD. It is possible that the interactions between these residues and CHOL drive the reorientation of BACE1 in L_o domains.

Principal component analyses were performed for both all-atom systems using the RMSD-fitted α-carbon cartesian positions of BACE1. Cross correlation matrices were constructed using the first 11 principal components from the 9:1 DIPC:CHOL system (90.5% cumulative variance) and the first 10 principal components from the 7:3 DPPC:CHOL system (90.5% cumulative variance). The cross correlation matrices and the first two principal components for each system are shown in Fig. 8. The cross correlation matrices reaffirm our previous observation of a shortened helical structure in the 7:3 DPPC:CHOL bilayer compared with the 9:1 DIPC:CHOL bilayer. In the 7:3 DPPC:CHOL bilayer, group 1 (Q449 and T450) and group 4 (W480–C485) lack the helical character seen in the 9:1 DIPC:CHOL bilayer. In the 7:3 DPPC:CHOL bilayer, the essential motions of the residues from L483 to Q488 are better correlated with one another than they are in the 9:1 DIPC:CHOL membrane. This is a consequence of having a membrane anchor within the C-terminal side of the TMD α-helix and another within the C-terminal JMD.