For building KCNQ1-KCNE1 models, we collected residue contact restraints from previously published biophysical experiments on the KCNQ1-KCNE1 interaction: disulfide crosslinking (Chung et al., 2009; Wang et al., 2011; Chan et al., 2012; Wang et al., 2012), Cd(II)-cysteine bridging (Tapper and George, 2001), and double mutant cycle analysis (Strutz-Seebohm et al., 2011; Li et al., 2015). Restraints were compiled in a state-dependent manner based on the data informing whether a crosslink or mutation favored the open or closed channel state. Most restraints were available for the region N-terminal to the KCNE1 TMD (S37-A44) whereas a smaller number of restraints fell within the TMD (L45-L71). To obtain additional contact information for the KCNE1 TMD, we performed KCNQ1-KCNE1 disulfide trapping experiments and tested if KCNQ1 V141 and I274 are close to L45, V47, or L48 in KCNE1. Residues V141 and I274 are the location of two of five known KCNQ1 gain-of-function mutations within the putative KCNE1-binding region between the VSD and PD: S140G (Chan et al., 2012; Chen et al., 2003a; Peng et al., 2017), V141M (Chan et al., 2012; Peng et al., 2017; Hong et al., 2005), I274V (Arnestad et al., 2007), A300T (Bianchi et al., 2000), V307L (Bellocq et al., 2004). For three of those mutations (V141M, I274V, V307L) a gain-of-function phenotype is observed only when KCNE1 is present, suggesting a physical interaction. Indeed, in cysteine exchange experiments, V141C was previously shown to crosslink with residues flanking the KCNE1 TMD on the extracellular side (Chung et al., 2009; Wang et al., 2011; Chan et al., 2012). We expanded upon these earlier studies and tested for disulfide bond formation between cysteines introduced at V141 or I274 in KCNQ1 and L45, V47, or L48 in KCNE1 using oxidation-state dependent electrophysiology measurements (Figure 2, Figure 2—figure supplements 1–3). L45, V47, and L48 were selected because their sidechains are oriented toward KCNQ1 in an earlier KCNQ1-KCNE1 model by Kang et al., 2008.

The dual mutation KCNQ1 I274C+KCNE1 L45C led to a channel that was highly conductive under reducing conditions, but displayed lower peak current under oxidizing conditions (Figure 2A). This result indicates that disulfide bond formation between these cysteine-substituted residues trapped the KCNQ1-KCNE1 channel in a conformation that favors the closed state. In contrast, channels formed by co-expression of KCNQ1 V141C with KCNE1 L48C exhibited smaller current amplitude under reducing conditions (Figure 2B), indicating that disulfide bond formation between V141C and L48C favored the open state. These results are consistent with the prediction that L45 and L48 are in spatial proximity to KCNQ1 I274 and V141, respectively. The mutant KCNQ1 channels alone (Figure 2—figure supplement 1) or in the presence of WT KCNE1 (Figure 2—figure supplement 2) were insensitive to the addition of DTT or Cu-phenanthroline. Other KCNQ1-KCNE1 residue pairs tested (V141C-V47C, I274C-V47C, I274C-L48C) (Figure 2—figure supplement 3) showed no changes under reducing or oxidizing conditions. Together, these results show that the effect of DTT and Cu-phenanthroline is dependent on the presence of introduced cysteines at KCNQ1 residues V141, I274 and KCNE1 residues L48 and L45.

The above experimental information and the structural data collected from the literature were used as contact restraints to develop structural models for the KCNQ1-KCNE1 complex by molecular docking. Separate restraint lists for building KCNQ1-KCNE1 channel models in closed and open states were compiled (Supplementary file 1 – Table 2 and 3). As input for docking, we used models of human KCNQ1 (Kuenze et al., 2019) with the VSD and PD in resting/closed (RC) or fully activated/open (AO) conformations. Those models were previously developed based on homology modeling with the structures of X. laevis KCNQ1 (Sun and MacKinnon, 2017) and the K_V1.2/2.1 chimera (Long et al., 2007). The models have been recently confirmed by cryo-electron microscopy (EM)-determined structures of human KCNQ1 (Sun and MacKinnon, 2020), to which the model-predicted VSD and PD conformations are highly similar (Figure 3—figure supplement 1; Cα-RMSD for VSD and PD less than 1.9 Å and 2.0 Å, respectively). In the putative KCNE1-binding region used for docking, the homology models agree well with the cryo-EM structures; surface-exposed residues have a sidechain RMSD less than 2.5 Å and 4.0 Å in the RC and AO model, respectively (Figure 3—figure supplement 1). In addition, homology modeling provided a conformation for the VSD in the resting state, a state for which no experimental structure exists.

Using Rosetta protein-protein docking (Gray et al., 2003; Gray, 2006) and Rosetta Membrane potentials (Yarov-Yarovoy et al., 2006; Barth et al., 2007), the ensemble of ten models of the NMR-determined KCNE1 TMD structure (S37-L71) (PDB: 2K21) (Kang et al., 2008) was docked to the KCNQ1 models (Kuenze et al., 2019). We focused our structural studies on the isolated KCNE1 TMD (including a short stretch of the N-terminal TMD-flanking region (S37-A44)) because previous studies had demonstrated that this domain alone is sufficient to produce the slow activation kinetics and increased current amplitude expected for KCNQ1-KCNE1 channels (Melman et al., 2001). Models were generated by iterative rounds of protein-protein docking (Figure 3—figure supplement 2), each with a rigid-body docking phase and an all-atom flexible backbone and sidechain refinement phase, resulting in a steady optimization of the model restraint score and minimization of the Rosetta-calculated binding energy (ΔG_Binding) for the KCNQ1-KCNE1 interaction (Figure 3—figure supplement 3). The most favorably scoring models of the KCNQ1-KCNE1 complex in the RC and AO conformation that exhibit the best combined experimental restraint, Rosetta ΔG_Binding, and MolProbity scores (Supplementary file 1 – Table 4) are illustrated in Figure 3. The atomic coordinates for these models are included in the supporting material for this paper (Supplementary files 2 and 3) and can be obtained from PDB-Dev (PDBDEV: 00000042) (Vallat et al., 2018).

Within the KCNQ1-KCNE1 models, the KCNE1 TMD is bound in a cleft formed by S6 from one KCNQ1 subunit, S5, the P helix from a second subunit, and S1 and S4 from a third (Figure 3B+C). The N-terminal end of the KCNE1 TMD leans towards S1, S5, and S6, and its C-terminal end contacts S1 and the bottom of S4. The FTL motif resides deep in the membrane and is oriented toward KCNQ1 (Figure 3B+C). In the AO model, the KCNE1 TMD C-terminus forms additional interactions with the cytosolic end of S6, which is kinked toward the membrane. Those interactions are absent in the RC model, in which S6 extends more vertically into the cytoplasm. Other than these small differences, the KCNE1 TMD-binding mode within the RC and AO KCNQ1 channel is deemed similar.

The KCNQ1-KCNE1 models satisfied the experimental restraints remarkably well. Many of the restrained KCNQ1-KCNE1 Cα-Cα distances were below the upper cutoff (12 Å) employed in docking (Figure 4) and came close to the expected maximal Cα-Cα crosslinking distance (disulfide: 7–8 Å, cysteine-Cd(II)-cysteine: 10–11 Å) when considering dynamics in the protein model by conducting MD simulations (Supplementary file 1 – Table 2 and 3). Only one medium restraint violation (3.5 Å) and two large restraint violations (>5 Å) were observed in the RC model. The restraint with the medium violation involved residues KCNE1 L45 and KCNQ1 I274. It is possible that crosslinking between these cysteine-substituted sites slightly perturbed the KCNQ1-KCNE1 structure leading to a low-conductance closed-like state, which could explain why in the WT channel model these residue sidechains have suboptimal geometry for disulfide bond formation. The two largest violations involved pairs of residues originally believed to be in direct contact based on results of double mutant cycle experiments: KCNE1 T58–KCNQ1 F340 (Strutz-Seebohm et al., 2011) and KCNE1 Y65–KCNQ1 A344 (Li et al., 2015). In our KCNQ1 RC channel model (Kuenze et al., 2019) as well as in the experimental KCNQ1 structure (Sun and MacKinnon, 2020), F340 and A344 are deeply buried within the PD and inaccessible to KCNE1 T58 and Y65. Satisfying either of the two restraints would require hard-to-rationalize conformational changes in KCNE1 and/or KCNQ1. Moreover, Xu and coworkers (Xu et al., 2013) confirmed previously that the FTL motif residues are unable to form disulfide bonds with S338, F339, or F340 in KCNQ1, suggesting that energetic coupling between those KCNQ1-KCNE1 residue pairs is mediated by allosteric networks rather than a direct interaction.

To assure the robustness of our structure prediction protocol, control docking calculations were performed with the cryo-EM-determined AO state structure of human KCNQ1 (Sun and MacKinnon, 2020), which became available only after our KCNQ1-KCNE1 models were completed. These control calculations arrived at a model that was very similar to the one developed by docking KCNE1 to the Rosetta homology model of KCNQ1 (interface RMSD (I-RMSD) = 3.5 Å, Figure 3—figure supplement 4A). Additional control calculations were carried out with KCNE3, starting either with the cryo-EM-determined or Rosetta-predicted KCNQ1 model. Guided by a set of published experimental restraints for the KCNQ1-KCNE3 complex (Kroncke et al., 2016), this procedure was able to reproduce the experimental KCNE3 binding pose with an accuracy of I-RMSD = 2.5 Å or 4.2 Å, respectively, for these two structures (Figure 3—figure supplement 4B and C).

In summary, using iterative protein-protein docking and model filtering with experimental restraints, we extensively probed the KCNQ1-KCNE1 interaction and developed molecular models of the channel complex that favorably agree with the available experimental data. These models were tested by the subsequent experimental and MD analysis.

To gain further mechanistic insight into how KCNE1 and KCNQ1 interact, we analyzed the location and degree of their contacts by performing MD simulations of our structural models (Figure 5A and B, Figure 5—figure supplement 1) coupled with site-directed mutagenesis. For MD simulations, KCNQ1-KCNE1 models were prepared with a stoichiometry of 4:2 KCNQ1:KCNE1 subunits. This appears to be the predominant stoichiometry on the surface of mammalian cells (Plant et al., 2014), although the possibility for multiple ratios ranging from 4:1 to 4:4 has been discussed (Morin and Kobertz, 2008; Nakajo et al., 2010; Murray et al., 2016). In our modeling procedure, we did not expect to find changes in the KCNE1 interaction mode for different KCNQ1:KCNE1 ratios, because we first docked one KCNE1 molecule to tetrameric KCNQ1 and subsequently created 4:2 and 4:4 complexes by imposing C2 or C4 symmetry, respectively. In our final MD analysis, we focused on the 4:2 stoichiometry and observed no significant changes in the interaction mode between the two KCNE1 subunits and with respect to the model obtained by Rosetta docking.

The first residues in the KCNE1 TMD and TMD-flanking region (S37-G50) interact with S1, the S1-S2 loop, the P helix, and S6 in both the RC and AO channel models (Videos 1 and 2, Supplementary file 1 – Table 5). Among those residues, the largest number of contacts with KCNE1 is made by W323 at the N-terminal end of S6 (Figure 5C, left panel). Mutations of W323 to Ala and Leu led to KCNQ1-KCNE1 channels with faster activation (Figure 5—figure supplement 2A) and a significantly hyperpolarized activation curve compared to the WT channel (change in ‘apparent’ activation V_1/2 (see Materials and methods): ΔV_1/2app,W323A = -6.4 mV, ΔV_1/2app,W323L = -11.3 mV) (Figure 5D+E). Mutations of other residues in S6 (V324, V334 [Nakajo et al., 2011]) and in the nearby P helix (A300 [Bianchi et al., 2000], V307 [Bellocq et al., 2004]) also caused channel opening at more negative voltages, likely via destabilization of the closed state. This region has been implicated with the positive G(V) shift by KCNE1 (Nakajo et al., 2011). Mutation of W323 to Phe resulted in a WT-like channel (ΔV_1/2app,W323F = 0.5 mV) suggesting that an aromatic or large hydrophobic moiety is an important structural component required for interaction of site 323 with KCNE1. Our structural model shows that the indole ring of W323 caps the sidechain of KCNE1 Y46, which is tucked in between S6 and the P helix (Figure 5C).

The middle part of the KCNE1 TMD (L51-Y65) interacts with S1 and S5 in both the RC and AO channel models (Figure 5A and B, Videos 1 and 2, Supplementary file 1 – Table 5). KCNE1 interacts in slightly different ways with the C-terminal end of S4 in the RC and AO models owing to the movement of S4 when the VSD becomes activated (Video 3). Among the KCNQ1 residues interacting with the middle part of the KCNE1 TMD, Y267 in S5 had the largest number of MD contacts with KCNE1. Our structural models suggest Y267 engages in an H-bond contact with the sidechain of T58 (Figure 5C, middle panel). We therefore used electrophysiology to experimentally confirm an interaction with KCNE1. Substitution of Y267 with Phe, which maintains the aromatic sidechain character but lacks a hydroxyl group, resulted in a significant alteration of the voltage-dependence of KCNQ1-KCNE1 activation (ΔV_1/2app,Y267F = 13.5 mV) (Figure 5D+E). We also noted that channel activation was previously shown to be altered by Ala mutations at Y267 and other residues in S5, mostly L266 and F270 (Strutz-Seebohm et al., 2011), which underscores the importance of the S5-KCNE1 interface.

At its cytosol-proximal end, the KCNE1 TMD interacts with the S0-S1 loop, S4, and, in the AO channel model, additionally with the loop connecting S4 and S4-S5L and with the C-terminal end of S6 (Figure 5C right panel, Videos 1 and 2). The latter interaction may specifically contribute to the stability of the open state by reducing the dynamics of S6 and locking the S6 gate open. To test this hypothesis, we introduced Ala mutations at K362 and N365 in S6, and determined the effect on the voltage-dependence of KCNQ1 activation. Both mutations resulted in significantly depolarized activation V_1/2app (ΔV_1/2app,K362A = 14.7 mV, ΔV_1/2app,N365A = 14.7 mV) (Figure 5D+E) and significantly faster deactivation (Figure 5—figure supplement 2A), indicating that these mutant channels required more energy to open. A proximity between the C-terminal ends of the KCNE1 TMD and S6 is supported by the results of cysteine-crosslinking experiments (Lvov et al., 2010), which showed that H363C in KCNQ1 formed disulfide bonds with H73C, S74C, and D76C in KCNE1. While these residues were not included in our docking because of their location in a flexible linker region, they are close to the last KCNE1 residue in our models. Extending the models by five additional KCNE1 residues can bring H73, S74, and D76 in contact with H363 (Figure 5—figure supplement 3). Furthermore, using double mutant cycle analysis, Chen et al., 2020 recently suggested that H73, S74, and D76 can interact with another residue on S6, I368. This result can also be explained by our structural models (Figure 5—figure supplement 3). Interestingly, the sequence R360-Q361-K362-H363 has been observed to undergo a major conformational change during KCNQ1 gating (Sun and MacKinnon, 2020). This structural change has been associated with PIP2 binding (Sun and MacKinnon, 2020). Analysis of our KCNQ1-KCNE1 model suggests interaction with KCNE1 may play an additional stabilizing role to this channel-lipid interaction.

Taken together, these results have identified several important interactions in the transmembrane KCNE1-binding site of KCNQ1. We next focused on identifying and validating the KCNE1 sites that interacted with KCNQ1 in the MD simulations.

The KCNE1 residues observed in MD simulations to make many contacts with KCNQ1 include Y46, V47, L51, F54, F57, T58, I61, M62, and Y65 in both the RC and AO channel models (Figure 5A and B). Additionally, in the AO model, S64, I66, and K69 interact with KCNQ1 S6 (Figure 5C, right panel). Residues Y46 and Y65, which border the KCNE1 TMD on the extra- and intracellular side, respectively, make the most contacts (Figure 6, Figure 6—figure supplement 1). Y46 is tucked in between S6 and the P helix. Y65 occupies the space below S4 and the loop connecting S4 and S4-S5L. Interactions at those two outermost points appear to anchor KCNE1 in its binding cleft and define its helical orientation (Figure 6A, Videos 1 and 2). The FTL residues are in the interface with KCNQ1; F57 is packed against S5, T58 is deeply buried between S1, S4, and S5, and L59 directly interacts with S1 (Figure 6B, Videos 1 and 2).

To validate this model-predicted binding mode, we correlated the pattern of KCNQ1-contacting positions in KCNE1 with site-directed mutagenesis data for KCNE1 (Figure 7). Exhaustive mutational scans of the KCNE1 TMD with Cys (Wang et al., 2012) or Trp and Asn (Chen and Goldstein, 2007) were previously reported. In addition, we tested Ala mutations at selected positions across the KCNE1 TMD and studied the mutational effects on voltage-dependent activation and gating kinetics of the resulting KCNQ1-KCNE1 channels (Figure 7A; Supplementary file 1 – Table 6). Aromatic residues were also mutated to Leu, and residues Y46 and Y65 were additionally changed to Phe.

We defined KCNQ1-contacting positions in KCNE1 by calculating the relative change in solvent-accessible surface area (ΔSASA) owing to KCNQ1-KCNE1 binding at every position in KCNE1 (Figure 7B). Residues with >20% ΔSASA are partially or fully buried in the KCNQ1-KCNE1 interface and expected to be more sensitive to mutation. We observed a fairly good match between the pattern of KCNE1 interface positions (in both the RC and AO model) and the location of high-impact mutation sites (according to a |ΔV_1/2| threshold criterion; see legend to Figure 7). 87% (14/16) of the high impact mutation sites in KCNE1 were located at the interface with KCNQ1 (ΔSASA >20%), and 71% (5/7) of the positions with low impact on channel activation were solvent-exposed (ΔSASA <20%) (Figure 7C). Thus, there is a clear dependence between KCNQ1-contact sites and high-impact mutation sites in KCNE1 for the orientation proposed by the structural models (p<0.05, χ²-test). However, no significant correlation was found when we simulated other hypothetical orientations for KCNE1 by rotation around its helical screw axis (Figure 7—figure supplement 1). This result and the following in-depth mutational analysis of selected KCNE1 residues support the model-predicted binding mode for KCNE1.

Mutation of Y46 to Ala and Leu produced channels that were activated at more negative potentials (ΔV_1/2app,Y46A = -36.3 mV, ΔV_1/2app,Y46L = -27.7 mV) and had faster activation kinetics (Figure 7A, left panel). These data suggest that mutating Y46 destabilized the closed state. This observation resembles the mutational phenotype for KCNQ1 W323 (Figure 5E). In our structural model, W323 and Y46 are in direct contact, which can explain why mutations at either position perturb KCNQ1-KCNE1 function in a similar manner. Simultaneous substitution of both W323 and Y46 with Ala led to a channel with almost complete loss of current, which prohibited us from confirming this interaction by double mutant cycle analysis. Substituting W323 or Y46 with Phe, however, maintained WT-like channel properties (Figures 5E and 7A), which suggests that an aromatic or large hydrophobic sidechain is a sufficient structural property required for interaction between these residues. In this regard, it is worth mentioning that the sidechain properties at Y46 also influence ion conduction through the KCNQ1-KCNE1 pore, as demonstrated by Xu et al., 2013. Small amino acids (Gly, Cys) and positively charged sidechain modifications (Cys-MTSET) increased the conductance of Cs⁺ relative to K⁺, whereas aromatic amino acids (Phe, Trp) did not significantly change the Cs/K conductance ratio compared to WT KCNE1. The position of Y46 in our structural models seems well suited to control this effect, and sidechain volume may influence the steric pressure on the nearby P helix to increase or restrict the conductance of larger Cs⁺ ions through the selectivity filter.

Mutations at F57 and Y65 also had a significant impact on KCNQ1-KCNE1 function (Figure 7A). Mutations F57A and F57L produced channels that opened at more negative voltages (ΔV_1/2app,F57A = -5.8 mV, ΔV_1/2app,F57L = -17.7 mV) with faster kinetics (Supplementary file 1 – Table 6). Mutations at Y65 led to a shift of V_1/2app to more positive voltages (ΔV_1/2app,Y65A = 4.3 mV, ΔV_1/2app,Y65L = 12.3 mV, ΔV_1/2app,Y65F = 9.4 mV), and Y65L and Y65A showed faster activation and deactivation kinetics (Figure 7A, right panel). This is consistent with the observation of Y65 forming sidechain packing and hydrogen bond interactions with multiple KCNQ1 residues in the MD simulations. Mutation I66A, while failing to meet our threshold criterion (|ΔV_1/2app| >10 mV) to be considered a high impact mutation site, still induced a significant change in activation voltage-dependence compared to WT (ΔV_1/2app,I66A = 8.2 mV, p<0.001), consistent with the interface location of this residue. In contrast, mutations at F56 and L63 failed to produce significant functional perturbations, as expected from their lipid-exposed positions.

Taken together, these results have identified several KCNE1 residues that interact with KCNQ1. We next focused on the interactions of the KCNE1 FTL motif and compared them with those of the TVG motif in KCNE3.

KCNE1 FTL (F57-T58-L59) is essential to induce slow activation of KCNQ1 (Melman et al., 2001; Melman et al., 2002). A hydroxylated amino acid at the middle position of this motif was previously found to be necessary for this effect (Melman et al., 2002). Replacement of FTL with TVG from KCNE3 shifts the G(V) curve of KCNQ1-KCNE1 channels towards that of KCNQ1-KCNE3 and removes KCNE1-specific effects on the gate and S4 movement (Barro-Soria et al., 2017). The proposed mechanism responsible for this effect is direct binding of FTL to the PD in KCNQ1 (Melman et al., 2004; Panaghie et al., 2006), and changed binding upon mutation to TVG. However, later studies have failed to confirm a direct interaction (Xu et al., 2013). Our refined models of the KCNQ1-KCNE1 complex together with recently released structures of the KCNQ1-KCNE3 complex (Sun and MacKinnon, 2020) provide new insight into the binding and mode of action of the activation motifs from KCNE1 and KCNE3.

The TMDs of KCNE1 and KCNE3 bind in the same location in KCNQ1 in both open and closed states and share the same overall orientation (Figure 8). When superimposing the TM segments of KCNQ1 from both complexes with each other, KCNE1 and KCNE3 deviate by a Cα-RMSD of less than 3.1 Å (RC model, Figure 8A) and 5.9 Å (AO model, Figure 8B), respectively, along their TM region. Small observable differences between KCNE1 and KCNE3, such as the degree of helix curvature, are expected and are likely related to different intrinsic conformational properties of the TMDs of KCNE1 and KCNE3. Previous NMR (Kang et al., 2008; Kroncke et al., 2016) and EPR (Sahu et al., 2014) studies revealed that the TMDs of unbound KCNE1 and KCNE3 are curved and that the degree of helix curvature can vary. Importantly, despite these small differences, we found that homologous residues in KCNE1 and KCNE3 occupy the same spatial position and point in the same direction toward KCNQ1. For instance, KCNE1 Y46 and Y65, and their corresponding residues in KCNE3, Y60 and Y79, make similar interactions with KCNQ1 (Figure 8C+E). The FTL and TVG motifs also share a common binding site, between S1 and S4 from one subunit and S5 from a second subunit, but make different specific interactions with KCNQ1 (Figure 8D). Most strikingly, we noticed an H-bond between KCNE1 T58 and KCNQ1 Y267 in the MD simulations (Figure 8F), an observation that offers an explanation for why a hydroxylated amino acid in the middle of the activation motif is required for KCNE1 function. In order to test if Y267 and T58 are interacting, we performed a double mutant cycle experiment by substituting Y267 with Phe and T58 with Val, either separately or in combination, and determined the changes in the activation energy of the resulting channel complexes (Figure 8—figure supplement 1). The energy changes for KCNQ1 Y267F–KCNE1 (ΔG = 0.66 kcal/mol), KCNQ1–KCNE1 T58V (ΔG = −0.36 kcal/mol) and KCNQ1 Y267F–KCNE1 T58V (ΔG = 0.90 kcal/mol) were not additive, however, the net energy change (|ΔΔG| = 0.60 kcal/mol) was smaller than 1.0 kcal/mol, which is commonly used as lower cutoff to identify two residues as interacting. Thus, we were not able to experimentally confirm an interaction between Y267 and T58. However, we note that the free energy changes at T58 were previously observed to have a pronounced sidechain volume dependency (Strutz-Seebohm et al., 2011) and that substitutions to amino acids involving a more drastic change in sidechain size could reveal a stronger energetic coupling between Y267 and T58 than determined in this work. In the KCNQ1-KCNE3 complex, V72 and Y267 are not interacting. The Y267 sidechain is oriented differently and is in H-bond distance to D242 (or M238) on S4 (Figure 8D). Different interactions are also observed for the first and third motif residues, F57 (T71) and L59 (G73), which are bulkier in KCNE1, probably causing different steric effects on the surrounding residues in KCNQ1. Thus, our structural analysis of the KCNQ1-KCNE1 and KCNQ1-KCNE3 channel complexes suggests that the different functions of KCNE1 and KCNE3 are the consequence of distinct interactions involving their activation motif sites, as discussed below, while interactions with the top and bottom of the TMD help to preserve the same overall KCNE TMD-binding mode.

Molecular models of the KCNQ1-KCNE1 complex were developed by computational docking with Rosetta (version 3.9) (Leaver-Fay et al., 2011; Figure 3—figure supplement 2). KCNE1 was docked to human KCNQ1 models of the closed state with the VSD and PD in resting and closed (RC) conformations and to models of the open state with the VSD and PD in fully activated and open (AO) conformations. Molecular modeling was guided by contact restraints derived from the results of disulfide crosslinking and site-directed mutagenesis experiments, which were obtained in this study or collected from previous literature reports (Tapper and George, 2001; Chung et al., 2009; Wang et al., 2011; Chan et al., 2012; Wang et al., 2012; Strutz-Seebohm et al., 2011; Li et al., 2015). Two separate restraint lists for docking of KCNE1 to either the closed or open KCNQ1 model were compiled based on the functional annotation of the channel state under the experimental conditions. A crosslink restraint was assigned the closed or open conformation if it trapped the channel in the closed or open state, respectively, or both conformations if the channel state was unclear. For example, Chung and coworkers (Chung et al., 2009) reported that disulfide crosslinks KCNE1 K41–KCNQ1 I145 and KCNE1 L42–KCNQ1 V324 favored the open state, whereas crosslinks KCNE1 K41–KCNQ1 V324 and KCNE1 L42–KCNQ1 I145 stabilized the closed state. Similarly, Wang et al., 2011 concluded, in the closed state, KCNQ1 residues T144, I145, Q147 are preferably crosslinked to KCNE1 R36-E43, whereas in the open state, those positions in KCNQ1 crosslink with KCNE1 G40, and KCNQ1 Q147 can be disulfide-bonded to KCNE1 R36, G38, and K41. Complete restraint lists for development of the closed and open KCNQ1-KCNE1 model can be found in Supplementary file 1 – Table 2 and 3, respectively.

Prior to docking, the TMD and a short stretch of the N-terminal juxtamembrane domain of the ten deposited models of the KCNE1 NMR structure (PDB: 2K21) (Kang et al., 2008), hereupon termed KCNE1 TMD (residues S37-L71), were energy-minimized with Rosetta using the RosettaMembrane (Yarov-Yarovoy et al., 2006; Barth et al., 2007) energy function. KCNE1 TMD was then placed near the transmembrane region of our previously published KCNQ1 homology models of the RC and AO state (Kuenze et al., 2019), and a total of 40,000 KCNQ1-KCNE1 models were generated for each state using the Rosetta protein-protein docking algorithm (Gray et al., 2003; Gray, 2006). The disulfide crosslinks and additional contact information were implemented as Cα-atom pair distance restraints with an upper bound of 12 Å, which corresponds to the length of an extended disulfide crosslink (7 Å) plus an additional 5 Å padding to account for the effect of protein flexibility. For Cd(II)-cysteine crosslinks (Tapper and George, 2001), the upper bound distance was increased to 15 Å (corresponding to 10 Å theoretical distance plus 5 Å padding). Distance restraints were evaluated with a harmonic penalty function that was zero below 12 Å (15 Å for Cd(II)-bridged crosslinks) and grew quadratically beyond that distance. The KCNQ1-KCNE1 docking models were filtered by a combination of score criteria (interface score <0 REU, ΔG_Binding <0 REU, atom pair restraint score <350 REU) and the remaining models were sorted by the binding energy between KCNQ1 and KCNE1 (ΔG_Binding). The 1000 best-scoring KCNQ1-KCNE1 models were then used as input structures for a subsequent round of docking to generate additional 40,000 model complexes. This alternating docking-filtering procedure (Figure 3—figure supplement 2) was iterated until the change in ΔG_Binding averaged over the ten lowest-energy models from iteration to iteration converged to less than 5%. In subsequent iterations, the atom pair restraint score cutoff was gradually decreased from 350 REU to 100 REU in the last docking round in order to apply a more restrictive experimental filter. In addition, the allowed range of translational and rotational perturbations at the beginning of each docking run was decreased from initially 3 Å / 8° to 1 Å / 3° to enable a more fine-grained conformational sampling. Docking calculations converged after five to seven iterations, and representative models were chosen by RMSD-based clustering of the 5000 best-scoring models, followed by visual inspection and MolProbity (Davis et al., 2007) analysis of the 10–20 lowest-energy models from the ten largest clusters. The model with the best combined ΔG_Binding and MolProbity scores from the largest model cluster as well as two additional low-energy models from the same cluster were selected for further MD analysis.

For MD simulations, KCNQ1-KCNE1 models were prepared with a stoichiometry of 4:2 KCNQ1:KCNE1 subunits. This appears to represent the predominant stoichiometry on the surface of mammalian cells (Plant et al., 2014), although the possibility of multiple stoichiometries ranging from 4:1 to 4:4 has been discussed (Plant et al., 2014; Morin and Kobertz, 2008; Nakajo et al., 2010; Murray et al., 2016; Chen et al., 2003b). The docked conformation of KCNE1 was duplicated and aligned with the opposite half of KCNQ1 by 180° rotation around the central channel axis. Subsequently, the interface of both KCNE1 molecules with KCNQ1 was relaxed by sidechain rotamer repacking and energy minimization of all backbone and sidechain degrees of freedom in Rosetta. Representative KCNQ1-KCNE1 complex models (with 4:2 stoichiometry) are provided with the supporting material to this paper (Supplementary file 2 and 3) and can be obtained from PDB-Dev under accession number PDBDEV_00000042. Restraints used in docking and starting model coordinates can be obtained from https://doi.org/10.5281/zenodo.3598943.

MD simulations of KCNQ1-KCNE1 models were performed in explicit phospholipid membranes at 310 K with AMBER16 (Case et al., 2016) employing the ff14SB (Maier et al., 2015) force field for proteins and the Lipid17 force field (Gould IR, Skjevik AA, Dickson CJ, Madej BD, Walker RC, 2018, "Lipid17: A Comprehensive AMBER Force Field for the Simulation of Zwitterionic and Anionic Lipids", manuscript in preparation). As starting conformations for MD, the Rosetta model with the best combined ΔG_Binding and MolProbity scores and two additional low-energy models from the ensemble of Rosetta docking models were selected, and prepared with a 4:2 KCNQ1:KCNE1 stoichiometry, as described above. Models were aligned to the membrane normal using the PPM webserver (Lomize et al., 2012) and embedded into bilayers of POPC (palmitoyloleoyl-phosphatidylcholine) and PIP2 (phosphatidyl-4,5-bisphosphate) (~280 lipids per leaflet) using the membrane builder tool of the CHARMM-GUI website (Wu et al., 2014). A TIP3P water layer with 24 Å thickness containing 150 mM of KCl was added on either side of the membrane. In addition, four K⁺ ions were placed in the channel selectivity filter at positions inferred from the X-ray structure of K_V1.2–2.1 (PDB: 2R9R). Bilayers contained 10 mol% of PIP2 in the inner leaflet which comprised equal numbers of C4-PO₄- and C5-PO₄-mono-protonated PIP2 molecules with stearoyl and arachidonoyl conjugations at the sn-1 and sn-2 position. The geometry of PIP2 was optimized with Gaussian 09 (Gaussian, Inc, Wallingford CT) on the B3LYP/6–31G** level of theory, and assignment of AMBER atom types and calculation of RESP charges was done with Antechamber (Wang et al., 2006). Bond and angle parameters of the protonated C4-PO₄ or C5-PO₄ group in PIP2 were adjusted to values previously reported for phosphorylated amino acids (Homeyer et al., 2006) to avoid simulation instabilities. SHAKE (Ryckaert et al., 1977) bond length constraints were applied to all bonds involving hydrogen. Nonbonded interactions were evaluated with a 10 Å cutoff, and electrostatic interactions were calculated by the particle-mesh Ewald method (Darden et al., 1993).

Each MD system was first minimized for 15,000 steps using steepest descent followed by 15,000 steps of conjugate gradient minimization. With protein and ions restrained to their initial coordinates, the lipid and water were heated to 50 K over 1000 steps with a step size of 1 fs in the NVT ensemble using Langevin dynamics with a rapid collision frequency of 10,000 ps⁻¹. The system was then heated to 100 K over 50,000 steps with a collision frequency of 1000 ps⁻¹ and finally to 310 K over 200,000 steps and a collision frequency of 100 ps⁻¹. After changing to the NPT ensemble, restraints on ions were gradually removed over 500 ps and the system was equilibrated for another 5 ns at 310 K with weak positional restraints (with a force constant of 1 kcal mol⁻¹ Å⁻²) applied to protein Cα atoms. The protein restraints were then gradually removed over 10 ns, and production MD was conducted for 450 ns using a step size of 2 fs, constant pressure periodic boundary conditions, anisotropic pressure scaling and Langevin dynamics. Four independent simulations were carried out for the RC and AO KCNQ1-KCNE1 model yielding 1.80 µs of total MD data for each state. Representative snapshots from the MD simulations can be obtained from https://doi.org/10.5281/zenodo.3598943.

Analysis of MD trajectories with CPPTRAJ (version 18.0) (Roe and Cheatham, 2013) included calculation of Cα-atom root-mean-square deviations (Cα-RMSD), enumeration of protein-protein hydrogen bonds, measurement of residue pair distances in the VSD, and counting of intermolecular contacts between KCNQ1 and KCNE1. Residue contact numbers were calculated by counting within a 4 Å radius of a given KCNQ1 or KCNE1 residue the number of heteroatoms from the other protein binding partner (i.e. the contact number of a KCNE1 residue was calculated by counting the number of atoms from KCNQ1 that were within 4 Å and vice versa). The residue contact number was then averaged over both KCNE1 molecules in all MD trajectories of the RC or AO KCNQ1-KCNE1 model, respectively.

Measurement of the channel pore radius was carried out with the HOLE program (Smart et al., 1996) using snapshots of KCNQ1 taken at one ns intervals during the last 400 ns of MD. Calculation of the solvent-accessible surface area (SASA) of KCNE1 was conducted using NACCESS (Hubbard and Thornton, 1993). In addition, computation of the binding free energy (ΔG_Binding) between KCNQ1 and KCNE1 was carried out using the MMPBSA.py program (Miller et al., 2012). A total of 2660 KCNQ1-KCNE1 conformations sampled at 150 ps intervals from the last 400 ns of a MD trajectory were processed to compute the molecular mechanics potential energies and solvation free energies in the MMPBSA procedure (Kollman et al., 2000). The solvation free energy contribution to ΔG_Binding was calculated using a continuum Poisson-Boltzmann (PB) model for channel proteins as described in Xiao et al., 2017. The entropic contribution to ΔG_Binding was estimated by applying the quasi-harmonic approximation (QHA) (Karplus and Kushick, 1981), and 26,600 KCNQ1-KCNE1 conformations were used for this analysis. ΔG_Binding of KCNQ1-KCNE1 was then calculated as the difference between the free energy of the KCNQ1-KCNE1 complex and the sum of the KCNQ1 and KCNE1 free energies.

Dynamical network analysis was performed with the Network View plugin (Eargle and Luthey-Schulten, 2012) in VMD (Humphrey et al., 1996). A node in the network was assigned to every amino acid in KCNQ1 and KCNE1 centered at their Cα atom. Network edges were defined between nodes whose residues were within 4.5 Å distance for at least 75% of the MD trajectory. The last 400 ns of simulation were used for the analysis. Edge weights were derived from the pairwise residue correlation matrix calculated with the program Carma (Glykos, 2006). Network communities were determined with the Girvan-Newman algorithm (Girvan and Newman, 2002) implemented in the program gncommunities as part of the Network View plugin (Eargle and Luthey-Schulten, 2012).

To check whether use of homology models of KCNQ1 as input for docking affected structure prediction of the KCNQ1-KCNE1 complex, docking calculations were also carried out with the cryo-EM-determined structure of human KCNQ1 (Sun and MacKinnon, 2020), which was released after the KCNQ1-KCNE1 models of this work were completed. In addition, docking calculations were performed with KCNE3 to assure that our computational protocol could recapitulate the experimentally observed structure for the KCNQ1-KCNE3 complex.

Prior to docking, the cryo-EM-determined open state structure of human KCNQ1 bound to calmodulin and KCNE3 (PDB: 6V01) (Sun and MacKinnon, 2020) was minimized with Rosetta using the FastRelax protocol (Conway et al., 2014) guided by the cryo-EM density map (Wang et al., 2016). C4-symmetry was enforced with the help of a symmetry definition file (DiMaio et al., 2011), and positional restraints on the protein backbone atoms were gradually ramped down during five repeats of FastRelax. Subsequently, 10 models of the NMR-determined KCNE1 TMD structure (S37-L71) (PDB: 2K21) (Kang et al., 2008) were docked to the cryo-EM AO model as described above using experimental restraints for the open channel state. The KCNQ1-KCNE1 model with the best ΔG_Binding score was deemed the final model and compared to the model developed in this work by docking KCNE1 to the Rosetta homology model of KCNQ1. Model similarity was evaluated by calculating the heavy-atom RMSD for the residues in the KCNQ1-KCNE1 interface within 10 Å of any residue on the other protein, and by computing the fraction of recovered contacts. A contact was defined as every residue in KCNQ1 (KCNE1) that was within 5 Å of a residue in KCNE1 (KCNQ1) in the reference model.

Similarly, ten models of the KCNE3 TMD NMR structure (P51-V85) (PDB: 2NDJ) (Kroncke et al., 2016) were docked to both the cryo-EM and Rosetta KCNQ1 AO model using the same computational protocol and published experimental restraints for the KCNQ1-KCNE3 complex (Kroncke et al., 2016). Similarity between the docking model and experimental KCNQ1-KCNE3 structure was assessed based on the all-atom interface RMSD and fraction of native contacts recovered.

Chinese hamster ovary cells (CHO-K1, CRL 9618, American Type Culture Collection, Manassas VA, USA) were grown in F-12 nutrient medium (GIBCO/Invitrogen, San Diego, CA, USA) supplemented with 10% fetal bovine serum (ATLANTA Biologicals, Norcross, GA, USA), penicillin (50 units/mL), streptomycin (50 μg/mL) at 37°C in 5% CO₂. The identity of CHO-K1 cells was certified by American Type Culture Collection using Cytochrome C Oxidase (COI) assay testing. Cells were negative for mycoplasma contamination and are regularly tested using the MycoAlert PLUS Mycoplasma Detection Kit (Lonza, Rockville, MD, USA). Unless stated otherwise, all tissue culture media was obtained from Life Technologies, Inc (Grand Island, NY, USA). CHO-K1 cells constitutively expressing human KCNE1 (designated CHO-KCNE1 cells) were generated using the FLP-in system (Thermo Fisher Scientific, Waltham, MA, USA) and maintained under selection with hygromycin B (600 μg/mL) as described previously (Vanoye et al., 2018).

KCNQ1 cDNA (GenBank accession AF000571) was engineered in the pIRES2-EGFP expression vector (BD Biosciences-Clontech, Mountain View, CA, USA) or a modified pIRES2-mScarlet vector, and KCNE1 cDNA (GenBank accession L28168) was cloned into a pIRES2-DsRed-MST vector as described previously (Vanoye et al., 2018; Lundquist et al., 2005; Manderfield and George, 2008). These vectors allowed co-expression of KCNQ1 and KCNE1 with fluorescent proteins as means for tracking successful cell transfection. Mutants of KCNQ1 and KCNE1 were generated using the QuikChange II XL system (Agilent technologies, Santa Clara, CA, USA). Correctness of the KCNQ1 and KCNE1 coding region was checked by DNA sequencing (Eurofins Genomics, Louisville, KY, USA), and plasmid DNA was amplified using an endotoxin-free plasmid preparation method (Nucleobond Xtra Maxi EF, Macherey-Nagel Inc, Bethlehem, PA, USA). Transfection of plasmid DNA encoding KCNQ1 and KCNE1 WT or variants into CHO-K1 cells for manual patch clamp experiments was performed using Fugene as previously described (Vanoye et al., 2009). Transfection of KCNQ1 WT and variant cDNA into CHO-KCNE1 cells (i.e. CHO-K1 cells with stable expression of KCNE1) and the transfection of KCNQ1 WT and KCNE1 WT or mutant cDNA into CHO-K1 cells for automated patch clamp recordings were done by electroporation using the Maxcyte STX system (MaxCyte Inc, Gaithersburg, MD, USA) as described previously (Vanoye et al., 2018).

Automated patch clamp experiments were performed using the Syncropatch 768 PE platform (Nanion Technologies, Munich, Germany) equipped with single-hole, 384-well recording chips with medium resistance (2–4 MΩ). Pulse generation and data collection were carried out with PatchController384 V.1.3.0 and DataController384 V1.2.1 software (Nanion Technologies, Munich, Germany). Whole-cell currents were filtered at 3 kHz and acquired at 10 kHz. The access resistance and apparent membrane capacitance were estimated using built-in protocols. Whole-cell currents were recorded at room temperature in the whole-cell configuration from −80 to +60 mV (in 10 mV steps) at 1990 ms after the start of the voltage pulse from a holding potential of −80 mV. The external bath solution contained: 140 mM NaCl, 4 mM KCl, 2 mM CaCl₂, 1 mM MgCl₂, 10 mM HEPES, 5 mM glucose, pH 7.4. The internal solution contained: 60 mM KF, 50 mM KCl, 10 mM NaCl, 10 mM HEPES, 10 mM EGTA, 2 mM ATP-K₂, pH 7.2. Whole-cell currents were not leak-subtracted. The contribution of background currents was determined by recording before and after addition of 20 μM of the I_Ks blocker HMR1556. Recordings with measurable outward current were examined to verify block by HMR1556. Only HMR1556-sensitive currents and recordings meeting the following criteria were used in data analysis: seal resistance ≥0.5 GΩ, series resistance ≤20 MΩ, capacitance ≥1 pF, voltage-clamp stability (defined as the standard error for the baseline current measured at the holding potential for all test pulses being <10% of the mean baseline current). Current-voltage (I-V) relationships were derived for all cell recordings meeting these criteria.

Oxidation-state dependent electrophysiological recordings of cysteine mutants of KCNQ1 (V141C, I274C) and KCNE1 (L45C, V47C, L48C) as well as the corresponding control experiments were done by manual patch clamp measurements. Whole-cell currents were recorded at room temperature (20–23°C) using Axopatch 200 and 200B amplifiers (Molecular Devices Corp., Sunnyvale, CA, USA) in the whole-cell configuration of the patch clamp technique (Hamill et al., 1981). Pulse generation was performed with Clampex 10.0 (Molecular Devices Corp., Sunnyvale, CA, USA). Whole-cell currents were filtered at 1 kHz and acquired at 5 kHz. The access resistance and apparent membrane resistance were estimated using an established protocol (Lindau and Neher, 1988). Whole-cell currents were not leak-subtracted. Whole-cell currents were measured from −80 to +60 mV (in 10 mV steps) at 1990 ms after the start of the voltage pulse from a holding potential of −80 mV. The external bath solution contained: 132 mM NaCl, 4.8 mM KCl, 1.2 mM MgCl₂, 1 mM CaCl₂, 5 mM glucose, pH 7.4. The internal solution contained: 110 mM K⁺-aspartate, 1 mM CaCl₂, 10 mM HEPES, 11 mM EGTA, 1 mM MgCl₂, 2 mM ATP-K₂, pH 7.3. The pipette solution was diluted 5–10% to prevent induction of swelling-activated currents. Patch pipettes were pulled from thick-wall borosilicate glass (World Precision Instruments, Inc, Sarasota, FL, USA) with a multistage P-97 Flaming-Brown micropipette puller (Sutter Instruments Co., San Rafael, CA, USA) and heat-polished with a Micro Forge MF 830 (Narashige, Japan). After heat polishing, the resistance of the patch pipettes was 3–5 MΩ in the control recording solution. As a reference electrode, a 2% agar-bridge with a composition similar to the control bath solution was utilized. Junction potentials were zeroed with the filled pipette in the bath solution. Unless otherwise stated, all chemicals were obtained from Sigma-Aldrich (St. Louis, MO, USA).

The possibility for disulfide bond formation between KCNQ1 and KCNE1 cysteine mutants was tested by perfusing cells with 10 mM 1,4-dithiothreitol (DTT, reducing) or 100–350 μM Cu(II)-phenanthroline (Cu-phen, oxidizing) in the external bath solution and measuring whole-cell currents as described above. Because the I_Ks current when analyzed by whole-cell voltage clamp in mammalian cells exhibits rundown (see Supplemental Figure 2 in Vanoye et al., 2018), which varies in both rate and magnitude, thus confounding experiments on a single cell in which there is a time lapse between conditions and treatments, we recorded from multiple different cells for each treatment (Control, +DTT, +Cu-phen) to average the cell-to-cell variability. In addition, cells were recorded for all treatments from a specific cell transfection batch and from at least three distinct transfections for each KCNQ1-KCNE1 combination.

Data were collected for each experimental condition from at least three transfections and analyzed and plotted using DataController384 V1.2.1 (Nanion Technologies, Munich, Germany), Clampfit V10.4 (Molecular Devices Corp.), Excel (Microsoft Office 2013, Microsoft), SigmaPlot 2000 (Systat Software, Inc, San Jose, CA, USA) and OriginPro 2016 (OriginLab, Northampton, MA, USA) software. Whole-cell currents were normalized for membrane capacitance and results expressed as mean ± SEM. The number of cells used for each experimental condition and the threshold for statistical significance (p<0.001) are given in the figure legends or table footnotes. Additional custom semi-automated data handling routines were used for rapid analysis of current density, voltage-dependence of activation, and gating kinetics. The voltage-dependence of activation was determined only for cells with mean current density greater than the background current amplitude. Since normalized KCNQ1-KCNE1 tail currents do not saturate at the potentials tested nor would they saturate at more depolarized potentials (see Figure 3C in Wang et al., 2020 and Figure 3C in Wang et al., 2012), we refer to the V_1/2 value determined by curve fitting as the ‘apparent’ activation V_1/2 (V_1/2app).