Structural dynamics of the active HER4 and HER2/HER4 complexes is finely tuned by different growth factors and glycosylation

Human Epidermal growth factor Receptor 4 (HER4 or ERBB4) carries out essential functions in the development and maintenance of the cardiovascular and nervous systems. HER4 activation is regulated by a diverse group of extracellular ligands including the neuregulin (NRG) family and betacellulin (BTC), which promote HER4 homodimerization or heterodimerization with other HER receptors. Important cardiovascular functions of HER4 are exerted via heterodimerization with its close homolog and orphan receptor, HER2. To date structural insights into ligand-mediated HER4 activation have been limited to crystallographic studies of HER4 ectodomain homodimers in complex with NRG1β. Here we report cryo-EM structures of near full-length HER2/HER4 heterodimers and full-length HER4 homodimers bound to NRG1β and BTC. We show that the structures of the heterodimers bound to either ligand are nearly identical and that in both cases the HER2/HER4 heterodimer interface is less dynamic than those observed in structures of HER2/EGFR and HER2/HER3 heterodimers. In contrast, structures of full-length HER4 homodimers bound to NRG1β and BTC display more large-scale dynamics mirroring states previously reported for EGFR homodimers. Our structures also reveal the presence of multiple glycan modifications within HER4 ectodomains, modeled for the first time in HER receptors, that distinctively contribute to the stabilization of HER4 homodimer interfaces over those of HER2/HER4 heterodimers.


Introduction
Human Epidermal growth factor Receptor 4 (HER4) is a ubiquitously expressed receptor functioning in heart, mammary, and neural development [1][2][3][4][5]. Binding of extracellular growth factors leads to HER4 receptor homodimerization or heterodimerization with one of three other HER receptor family members, EGFR, HER2 or HER3, and subsequent activation of their intracellular kinase domains [1][2][3]. While HER4 activation is linked to signaling pathways activated by other HER receptors, including Ras/MAPK and PI3K/Akt, HER4 is the only HER with documented growth inhibitory effect on cells [4,6,7]. Consistent with this observation, and in contrast to other HER receptors for which genetic alterations are widely linked to oncogenesis [8], HER4 is more commonly observed to be lost or downregulated in human cancers [4,[8][9][10]. More rarely, HER4 activating mutations and overexpression have been observed in lung, melanoma and gastric cancers [10,11].
HER4 plays distinct roles in the nervous and cardiovascular systems from other HER receptors, which are underscored by the pathological consequences of dysregulated HER4 signaling [5,12,13]. Aberrant activation of HER4 is associated with neurological diseases including amyotrophic lateral sclerosis (ALS), schizophrenia and other psychological disorders, where inhibitory missense mutations in HER4, and either increased or decreased levels of the HER4 ligand NRG1 can lead to various disease phenotypes [12,14,15]. In cardiomyocytes, HER4 heterodimerization with HER2 is particularly important for survival under acute stress conditions [3,13,16]. HER2 and HER4 signaling are both essential for embryonic and postnatal heart development [5,16,17].
As an orphan receptor, HER2 does not undergo ligand-induced homodimerization and relies on HER4 for activation [18,19] when HER4 is bound to NRG1 produced by the cardiac endothelium [17] (Figure 1a). Disruption of the HER2/HER4 signaling has been attributed to cardiotoxic effects of HER2-targeting cancer therapeutics, such as Herceptin [20].
The ectodomains of all HER receptors comprise of four domains (I-IV), which in the ligand-free state in EGFR, HER3 and HER4 adopt a tethered conformation around a beta hairpin protrusion known as the dimerization arm [21,22]. Early crystal structures of HER4/NRG1b, EGFR/EGF and EGFR/TGFa ectodomain (ECD) homodimers revealed that ligand binding between extracellular domains I and III causes a substantial conformational change that exposes the dimerization arm in domain II, allowing for formation of active dimers stabilized through dimerization arm exchange between the monomers (Figure 1a) [23][24][25][26]. In these symmetric ectodomain dimers, most of the interaction surface between the two receptors falls within the dimerization arm regions.
The orphan HER2 adopts an extended conformation in its apo state, thus being dimerizationcompetent without ligand binding [18,27]. However, HER2 does not form stable homodimers under physiological expression levels and relies on heterodimerization with another ligand-bound HER receptor for activation [18]. Inability to efficiently homodimerize might be encoded in the nonoptimal manner with which HER2 engages a dimerization arm of a partner receptor, as illustrated in the recent structures of the NRG1b-bound HER2/HER3 and EGF-bound HER2/EGFR ectodomain heterodimers [28,29]. When complexed with NRG1b-bound HER3, HER2 fails to engage the HER3 dimerization arm leaving only the HER2 arm engaged at the dimer interface [28]. The dispensability of the HER3 arm at the interface is corroborated by the observation that its deletion does not impact HER2/HER3 dimerization and signaling [28]. In the EGF-bound EGFR/HER2 structure, the EGFR dimerization arm binds HER2 but in a non-canonical manner characterized by increased dynamics and interactions of the arm with HER2 domains II and III instead of domain II and I observed in most other HER ECD dimers. As in the HER2/HER3 complex, the dimerization arm of the HER2 partner (in this case EGFR) is not required for heterodimerization and activation [29].
Together, the HER2-containing heterodimer structures reveal both a dynamic mode with which HER2 engages a dimerization arm from a partner receptor [28]. These dynamics also suggest that HER homodimers might preferentially form over HER2-containing heterodimers or heterodimers in general. This is consistent with repeated findings in cells expressing EGFR and HER2 in which a strong preference for EGFR homodimerization is observed over heterodimerization with HER2 upon EGF treatment [29][30][31]. Further evidence comes from biophysical studies in which isolated HER ectodomains were shown to form strong homodimers and only weakly detectable heterodimers in the presence of their cognate growth factors [32].
However, this is not always the case and EGFR was reported to form heterodimers more favorably when stimulated with another ligand, betacellulin (BTC) [33]. In addition, HER4 seems to engage equally as homodimers or as heterodimers with HER2, at least when interactions between isolated receptor ectodomains were measured [32]. These receptor and dimer-specific idiosyncrasies highlight the importance of investigating HER receptor complexes with different ligands and a particular need for understanding how HER2 engages with HER4 -the final structure missing among HER2-containing HER complexes.
HER4 is activated by a diverse set of growth factor ligands including the neuregulin 1-4 family (NRG1-4), amphiregulin, epiregulin (EREG) and BTC [3,7,34]. These ligands differ widely in their tissue expression and biological function [35,36]. For example, NRG1b plays essential roles in the development and functioning of the cardiovascular system and nervous system [35,37], while BTC is implicated in the differentiation of pancreatic b-cells [36]. Even in the same cells, these ligands induce distinct signaling outputs. In the human T lymphoblastic CEM cells stably expressing HER4, NRG1b (and NRG2b) are the most potent activators of AKT signaling, while BTC induces the strongest activation of ERK1/2 [7]. These differential effects are likely due to a combination of factors. First, ligands are cross-reactive: NRG1b is also a ligand for HER3 while BTC also binds to EGFR [35,36]. Second, they might form structurally different HER4 ectodomain dimers, which in turn will affect dimer stability and downstream signaling, as observed for EGFR [7,38]. Third, the ligands might differentially modulate the degree of HER4 heterodimerization versus homodimerization [33,39]. In particular, BTC appears to be uniquely poised to promote a wide range of HER heterodimers, including HER2/HER3 [36,40,41], EGFR/HER3 and the HER2/HER4 [33,36,40,41]. The mechanism for BTC-based dimerization of HER receptors remains unknown without structures of their complexes.
Whether HER4 adopts different conformation(s) while bound to different ligands, as seen in EGFR, has remained an open question as only crystal structures of NRG1b-bound HER4 dimers have been reported [25]. A spectrum of ligand-bound EGFR structures, including high affinity (EGF and TGFa) or low affinity (EREG), revealed different dimerization interfaces and underscored that the dimerization arm plays an important role in communication between the ligand binding pocket and EGFR dimer interface [38,42]. In this study, we investigated these relationships for HER4, and its complexes with HER2. We focused on comparison of NRG1b with BTC due to lack of structural insights into interactions of BTC with HER receptors, and its documented aptitude for promoting receptor heterodimerization in contrast to other ligands. Both ligands are known to promote HER4-dependent activation of HER2 [36,40,41]. We used cryoelectron microscopy (cryo-EM) to determine the first high-resolution structures of the NRG1band BTC-bound HER4/HER2 and HER4/HER4 ectodomain dimers in a full-length receptor context. Our analysis shows that there are no major differences between NRG1b-and BTCbound complexes, but surprisingly that in each case HER4 homodimers displayed large-scale dynamics compared with HER2/HER4 heterodimers. We also show that glycan modifications within HER4 ectodomain extensively contribute to the HER4 homodimer interface, a feature previously not recognized in any other HER receptor complexes.
To reconstitute the active HER2/HER4 complex for high-resolution structural analysis by cryo-EM, we introduced a G778D mutation in HER2 that prevents Hsp90 binding to the HER2 kinase domain and promotes HER2 heterodimerization as previously described [28,43,44]. Both HER2 and HER4 receptors were truncated to remove their long, presumably unstructured, tails located C-terminal to the kinase domains and were transiently expressed in Expi293F cells individually.
HER2 was expressed in the presence of canertinib, a covalent type-I kinase inhibitor that stabilizes active conformation of the HER2 kinase. Cell lysates were pooled and in the first purification step, HER4 was affinity-purified via FLAG-tagged NRG1β or BTC [28,45]. In the second step, growth factor-bound HER4 complexes that interact with HER2 were enriched via a HER2-specific MBP-tag using amylose affinity resin ( Figure S1a-b). Eluted proteins were further purified by size exclusion chromatography ( Figure S1c) and dimeric fractions were frozen on graphene-oxide coated (GO) grids for cryo-EM analysis ( Figure S1d).
We observed that HER2/HER4 dimers constituted only a small fraction of complexes purified using both ligands, indicating that in each case HER4 favored self-association ( Figure S1b). HER receptor kinases asymmetrically dimerize in an active receptor complex, with one kinase adopting the function of an allosteric activator of the second kinase (receiver) [46]. To increase the yield of HER2/HER4 heterodimers vs HER4 homodimers, we introduced specific mutations that render the kinases activator only (N-lobe IQ mutation) or receiver only (C-lobe VR mutation). These mutations disrupt kinase homodimers but do not interfere with heterodimers in which the N-lobe mutant combined with the C-lobe mutant reconstitutes the asymmetric dimer [46]. By introducing the relevant mutations, we designed HER2 and HER4 mutants to be compatible with two opposite activator /receiver configurations: HER4 activator (IQ)/HER2 receiver (VR) and HER2 activator (IQ)/HER4 receiver (VR) (Figure 1a). We first tested signaling competency of these combinations, by transiently transfecting full-length HER2 and HER4 carrying respective mutations in COS7 cells and assessing receptor phosphorylation upon growth factor stimulation by Western blot analysis of cell lysates ( Figure S1e). Strikingly, the active heterodimer was only reconstituted in the HER4 activator (IQ)/ HER2 receiver (VR) configuration pointing to stereotyped roles that these two receptors play in the active complex irrespective of the activating growth factor (Figure 1a, Figure S1e). We used this set of interface mutations to enrich for the fraction of functional HER2-G778D-V956R/HER4-I712Q heterodimers in large scale purification for structural studies. We will refer to these complexes simply as HER2/HER4.
The HER2/HER4/NRG1β structures complete the panel of recently reported HER2 heterodimeric complexes. As in the HER2/HER3/NRG1β, HER2/EGFR/EGF and HER2/EGFR/EREG structures [28,29], in the HER2/HER4 complexes the ligand-free HER2 enforces an asymmetric geometry within the heart-shaped ectodomain complex (Figure 1b-c, Figure S5a). The conformation adopted by HER2 is nearly identical in all complexes ( Figure S5b, pairwise RMSD within HER2s 1.1-1.6 Å across different complexes), and the same as observed in structures of an isolated HER2 ectodomain alone or in complex with therapeutic antibodies with only minor variations at the tip of the dimerization arm ( Figure S5b). Thus, our structures are consistent with previous findings that HER2 does not undergo observable conformational changes upon heterodimerization with other HER receptors. The conformation adopted by HER4 in our structure is identical to a previously observed conformation in the crystal structure of isolated HER4 extracellular domain bound to NRG1β (RMSD 1.7 Å -4.4 Å across different structures). This conformation also closely matches the extended state of HER3 and EGFR in their respective heterodimers with HER2 (RMSD 2.2 Å, RMSD 2.6 Å, respectively) ( Figure S5b) [28,29].
Despite the diverse sequences of the NRG1β and BTC ligands, the conformation of the HER2/HER4 heterodimers stabilized by each ligand is identical with only small differences in the ligand binding pockets (Figure 1d). Recently published structures of EGFR homodimers induced by binding of the two high affinity ligands, EGF and TGFa, revealed that binding of these two different ligands results in distinct ensembles of EGFR dimer conformations, specifically within domains IV, seemingly coupled to scissor-like movements around the dimerization arm region [42]. To investigate whether NRG1β and BTC might lead to similar effects in the HER2/HER4 structures, we have performed an equivalent analysis. Extensive 3D variability and 3D classification analysis of the HER2/HER4/NRG1β and HER2/HER4/BTC datasets did not reveal any defined conformational heterogeneity within domains IV or dimerization arm regions ( Figure   1e).
Given the non-canonical engagement of the dimerization arm of the HER2 partner receptor in previously solved HER2-containing heterodimer structures, we analyzed the HER4 dimerization arm in our structures. HER2 and HER4 dimerization arms are resolved in both NRG1β and BTCbound HER2/HER4 complexes (Figure 1b-c, Figure 2a). The HER2 dimerization arm is stabilized by several polar and Van-der-Waals interactions with the dimerization arm-binding pocket of HER4, which involve two conserved aromatic residues, specifically Y274 and F279 in HER2 that interact with HER4 G286, C304 and R306 (Figure 2b box B). The equivalent residues in the HER4 dimerization arm, Y268 and F273, are engaged in reciprocal interactions with the backbone atoms of HER2 G292, C311 and L313 via a network of hydrogen bonds (Figure 2b box C). These aromatic dimerization arm residues are strictly conserved as phenylalanine or tyrosine residues in all HER receptors ( Figure S6a) and participate in the same interactions in the structures of the symmetric EGFR/EGF and HER4/NRG1b ectodomain homodimers ( Figure S6b). In the EGFR/HER2 and HER3/HER2 heterodimers, only the HER2 dimerization arm makes these interactions ( Figure S6b). The EGFR dimerization arm is rotated out of the canonical dimerization arm binding pocket of HER2, preventing such interactions, and the dimerization arm of HER3 is not even resolved (Figure 2c-d, Figure S6b) [23,25]. Thus, in this regard HER2/HER4 heterodimers are more similar to known structures of HER homodimers (NRG1β-bound HER4 and EGF-bound EGFR homodimers) than to heterodimers.
These similarities are also reflected in the interactions that the tips of both dimerization arms at the HER2/HER4 interface make with domains I of their respective dimerization partners. The  Figure S6b). Other HER2 heterodimers and, incidentally also the EGFR/EREG dimer, exhibit an asymmetric dimerization arm configuration with one dimerization arm being less engaged, evidenced by increased Bfactors ( Figure 2d). Thus, the relative orientation of two HER monomers varies among all HER heterodimer structures ( Figure S5c). EGFR and HER3 exhibit a hinging motion in the direction of the HER2 dimerization arm in comparison to HER4, which is rotated slightly away from it ( Figure   S5c). Interestingly, the only other instance when a HER2-containing heterodimer is observed to make symmetric interactions is when HER2 carries an oncogenic mutation, S310F, in the heterodimeric complexes with HER3 ( Figure S5c) [28]. This suggests that HER2/HER4 heterodimers are the most stable among HER2 heterodimers. Consistent with this notion, previous studies showed that the recombinant HER2 and HER4 ECDs form the most stable heterocomplex among all other HER heterodimers, with efficiency similar to HER4 homodimers [32].

The dimerization arm of HER2, but not HER4, is required for HER2/HER4 activation
The EGFR and HER3 dimerization arms are dispensable for signaling within their respective heterodimers with HER2, a property attributed to their high flexibility and disengagement from the HER2 dimerization arm binding pocket [28,29]. The canonical binding mode of the HER4 dimerization arm in the HER2/HER4 dimer structures raises the question whether it is required for signaling by this complex. To test the role of HER4 arm, we transduced full-length HER2-VR (V956R) and HER4-IQ (I712Q) constructs into murine NR6 cells. Dimerization arm sequences were replaced either in HER2 or HER4 with a flexible loop of alternating glycine and serine residues as previously described (GS-arm) [28,29]. NRG1b-induced phosphorylation of HER2, HER4, ERK and AKT was not notably affected by substitution of the HER4 dimerization arm to a GS-arm relative to wild type receptors, indicating that the HER4 dimerization arm is not required for assembly and activation of HER2/HER4 heterodimers ( Figure 2e). In contrast, substitution of the HER2 dimerization arm sequence fully abolished activation of the heterocomplex, as previously reported (Figure 2e) [29]. Thus, despite full engagement at the interface of both dimerization arms in the HER2/HER4 complexes, the HER4 arm is still dispensable for activation, and it is the HER2 arm that potentiates formation of the active complex.

HER4 homodimers display higher large-scale conformational flexibility than HER2/HER4 heterodimers
Our cryo-EM structures of the full-length HER2/HER4 complexes bound to either NRG1β or BTC, did not reveal discernible differences at the receptor dimerization interface (Figure 1d). In other cryo-EM structures of the full-length HER2 heterodimers, EGFR/HER2 bound to high-affinity EGFR ligand, EGF, or a low-affinity EREG, any differences are also imperceptible [29]. Ligandspecific differentiation of structural states becomes only evident in EGFR homodimers. The most drastic example is breaking of C2 symmetry in the crystal structures of EGFR ectodomain homodimers bound to EREG vs symmetric structures of EGFR with EGF or TGFa (see Figure   S5a for symmetry axes in HER receptor dimers) [38]. However, even in the symmetric crystal structures of EGFR bound to two high affinity ligands, EGF and TGFa there are differences in intermonomer EGFR angles between the two ligand complexes [23,26]. As mentioned above, cryo-EM analysis of the full-length EGFR homodimers, extensive 3D classification and variability analysis revealed that that both EGF and TGFa stabilize a range of EGFR dimer shapes with different intermonomer angles, but they differ in their ability to stabilize conformations with large intermonomer angles in which membrane-proximal domains IV are separated [42].
These comparisons raise the question of whether HER homodimers explore a wider range of conformations compared to heterodimers, specifically those singly-liganded heterodimers that contain the orphan HER2 receptor. To test this hypothesis for HER4 complexes, we determined the cryo-EM structures of full-length HER4 homodimers bound to NRG1β or BTC (Figure 3a  Figure S8). In both structures, HER4 kinase domains were bound to afatinib to enable high-resolution reconstruction in the ECD module (3.4 Å for NRG1β, and 3.7 Å for BTC) ( Figure   S8). AMP-PNP/Mg 2+ -bound or apo HER4/NRG1b complexes resulted in a similar overall reconstruction, albeit at lower resolution (4.2 Å and 3.9 Å, respectively) ( Figure S7f-g). As observed in a previous HER4/NRG1b crystal structure of isolated ECDs, liganded HER4 assembles into homodimers with near perfect C2 symmetry (Figure 3a, Figure S9). However, while we observed that applying C2 symmetry in the final refinement step nominally improved resolution ( Figure  vs TGFa bound EGFR homodimers [42] but persisted through multiple different processing methodologies (see methods). Such observations are indicative of the increased dynamics present in homodimeric HER4 receptor assemblies compared to their HER2-bound heterodimer counterparts. This difference in intermonomer angles was maintained even with C2 symmetry applied to final refinement steps and 3D classification.
The structural origins for the differences in intermonomer angles within the NRG1b and BTCbound HER4 homodimers are challenging to explain. Due to the differences in intermonomer movement, overlays between NRG1b and BTC-bound HER4 differ slightly, however overall, the two homodimers are almost identical (Figure 3d

HER4 glycosylation reveals structural stabilization via glycans that bridge extracellular subdomains and receptor dimers
Notably, our cryo-EM HER4/NRG1β structure is more symmetric than the crystal structure of the HER4/NRG1β ectodomain homodimer in which each monomer adopts a different orientation of the domain IV relative to the rest of the ectodomain ( Figure S9, RMSDs for monomers within the three homodimers per asymmetric unit in the 3U7U crystal structure are 1.67 Å, 5.76 Å and 2.38 Å compared to 1.42 Å in the cryo-EM HER4/NRG1b homodimer) [25]. Several reasons could account for this variation, including consequences of crystal packing or lack of intracellular and transmembrane domains, which are present in our constructs, albeit not resolved in cryo-EM density. Another explanation is differences in HER4 glycosylation in our cryo-EM sample purified from human cells as compared to deglycosylated HER4 ectodomains used for crystallography, which only maintain the first NAG on asparagine residues that are N-glycosylated [25]. As in the case of HER4, some glycans on HER2 mediate direct interdomain contacts within HER2, similar to the ones previously observed in the crystal structure of HER2 with pertuzumab, albeit more sugar moieties can be built in our structure [53]. The first three sugar moieties on N259 in domain II are particularly well-resolved and appear to directly engage the domain I polypeptide chain (Figure 4b box B). However, in contrast to the HER4 homodimers, we do not observe continuous density connecting the two heterodimer monomers indicating that glycan-mediated interfaces seen in HER4 homodimers cannot be established in HER2/HER4 heterodimers. This is because HER2 does not have glycosylation consensus sites equivalent to HER4 N358, N548 and N576 (Figure 4b box C). Based on these observations, it is tempting to speculate that the higher propensity for HER4 to homodimerize rather that heterodimerize with HER2 observed in our pull downs ( Figure S1b) is at least partially rooted in stabilization of the homodimer by glycanmediated interactions.

Discussion
First structures of HER2/HER4 and BTC complexes. We present here the first cryo-EM reconstructions of both the homodimer and heterodimer complexes of HER4 receptor in its fulllength form, bound to two different high affinity HER4 cognate growth factors, NRG1β and BTC. This is also the first time that a betacellulin growth factor has been resolved bound to a HER receptor. The HER2/HER4 heterodimer structures now complete the ensemble of possible HER receptor heterodimer structures that involve the orphan HER2 receptor. Only the ectodomains are resolved in our structures, as repeatedly has been the case for any full-length receptor tyrosine kinase reconstructions [28,42,[48][49][50][51]. While not resolved, interactions contributed by the intracellular domains appear to be essential for stabilization of the receptor complexes in our cryo-EM reconstructions. In our previous analysis of the HER2/HER3/NRG1β complex, introduction of oncogenic mutations in the HER3 pseudokinase that increase its dimerization affinity with HER2, and presence of HER2 kinase inhibitors were essential for efficient heterodimer reconstitution and improved resolution [28,45]. Similarly, for HER4 structures reported here, selective enrichment of kinase heterodimers via the introduction of activator/receiver mutations in HER2/HER4 and introduction of kinase inhibitors to homo-and heterodimer complexes improved the resolution of cryo-EM reconstructions.

Conserved features of HER2 heterodimers.
Across the family, the three HER2-containing heterodimers adopt an asymmetric heart-shaped ectodomain structure and in each one of them the HER2 conformation is identical while the dimerization interface is unique. The main difference centers on the engagement of the dimerization arm extended to HER2 by the partner receptors.
The HER2/HER3 interface is most dynamic with the HER3 dimerization arm not being resolved at all [28]. In the HER2/HER4 and the HER2/EGFR structures, HER4 and EGFR dimerization arms are resolved but make unique interactions with HER2 [29]. The EGFR dimerization arm engages HER2 via non-canonical interactions with domain III that resemble those only observed in the crystal structure of the EGFR/EREG ectodomain complex [29,38]. In comparison, the HER4 dimerization arm in the HER2/HER4 heterodimer presented here is engaged with HER2 via several canonical interactions, observed across most of the HER receptor homodimer structures [23,25,26,52]. This binding mode might explain HER2/HER4 heterodimer seems to be most stable among HER2-containing heterodimers, as measured by studying associations between isolated HER ectodomains [32].
While the positioning of the HER2 and HER4 dimerization arm appears almost symmetric, the number of hydrogen bonds formed by the HER4 dimerization arm is reduced compared to that of HER2. In addition, HER2 fails to engage its partner receptors via a conserved cation-p interaction that is exchanged by both monomers in all symmetric EGFR and HER4 homodimers, which in HER4 involves a dimerization arm phenylalanine (F273) and a domain II arginine (R306). The arginine is a leucine in HER2 (L313). It had been speculated previously that the inability of HER2 to form this interaction may be the reason for the non-canonical placement of HER3 and EGFR dimerization arms in their respective heterodimers with HER2 [28,29]. However, our HER2/HER4 heterodimer structure shows that even without the cation-p interaction, the HER4 dimerization arm can be placed in a canonical position. Lastly, the overall weaker interactions that HER2 makes with dimerization arms of its partner receptors are likely the reason why these partner arms are not needed for stabilization of the active signaling HER2 heterodimers. This has been observed for the HER2/HER3 and HER2/EGFR complexes [28,29], and we show here that the same is true for the HER2/HER4 heterodimer.

Homodimerization vs heterodimerization.
The extent of HER receptors propensity to form homodimers versus heterodimers, and their functional significance, are topics of ongoing debate. Some ligands, like EGF, are well documented to favor homodimers of their cognate receptors (EGFR in this case), while others, like BTC, have been shown to more readily promote heteroassociation [33,39]. While in a cellular context there might be many factors that shape these equilibria, including relative levels of receptor expression, their localization within membrane microdomains and/or interaction with other, yet unknown, factors that might stabilize certain dimer combinations, our studies bring insights into these interactions in a simplified in vitro system. We note that both NRG1b and BTC favor HER4 homo-association, and only a small fraction of complexes purified using pull downs with these ligands immobilized on beads yielded HER2/HER4 heterodimers. This was the case even when HER2 and HER4 kinase domains carried mutations designed to prevent their homo-associations and to favor heterodimerization. This phenomenon has been previously observed for the EGFR/HER2 system, where coexpressed receptors stimulated with EGF formed almost exclusively EGFR homodimers upon detergent extraction, with limited formation of EGFR/HER2 heterodimers [29]. The same study also observed only a small fraction of heterodimerization (<10%) by live-cell single molecule imaging of EGFR and HER2 on the plasma membrane of EGFR/HER2 positive SUM159 cells after EGF stimulation. Altogether, these findings raise questions about the conditions under which HER receptor heterodimers form in vivo, especially for HER receptors, which are not obligate heterodimers, namely EGFR and HER4. Is their reluctance to form such heterodimers an important part of the regulation of their signaling specificity, a current lack of knowledge about the conditions under which they form, or both? For example, one of the consequences of HER2 overexpression in cancer could be elevation of the otherwise non-optimal heterodimers with EGFR, resulting in potentiation of oncogenic signaling.
Biased agonism. The degree of symmetry between ectodomains of EGFR in the active liganded dimer has been correlated with the strength of its signaling output. Ligand binding is allosterically coupled to positioning of the dimerization arm and depends on how the ligand engages domains I and III. In EGFR, this allosteric path is differentially engaged by low affinity EGFR ligands (EREG, Epigen) vs high-affinity ligands (EGF, TGFa), resulting in asymmetric and dynamic dimers vs symmetric and stable dimers, respectively [23,26,38]. The weak asymmetric EGFR dimers have been shown to correlate with more sustained activation of ERK and AKT pathways leading to differentiation, while the more stable symmetric dimers induce more transient activation resulting in proliferation [38]. Recent cryo-EM studies of EGFR bound EGF and TGFa revealed that even the high affinity dimers induce a range of EGFR homodimer conformations differing at receptor intermonomer angles, which might explain distinct functional outcomes downstream from these receptor complexes [42]. These analyses directly link structural differences to functional EGFR outputs, posing a question how this regulation looks for other ligands and other HER receptor combinations.
Different HER4 ligands were reported to induce unique signaling outputs via activation of HER4 homodimers. Specifically, the two high affinity HER4 ligands BTC and NRG1β were shown to be different, with BTC more efficiently activating the ERK pathway while NRG1β activated AKT pathway more potently [7]. Our structures of HER4 ectodomain dimers bound to NRG1β and BTC presented here show that in both homodimers there are notable scissor-like movements around the dimerization arms with different intermonomer angles between the two ligands. It is possible that these different dimer conformations influence the stability and consequently signaling outputs emanating from these HER4 homodimers. While these structural differences are seemingly small, they are reminiscent of the ones observed in EGFR homodimers, bound to its two high affinity ligands, EGF and TGFa [42].
In contrast to the effects that BTC or NRG1β have on stabilizing more diverse conformational ensembles of the HER4 homodimers, their complexes with the HER2/HER4 heterodimers showed no discernable structural differences. Strikingly in the HER2/EGFR heterodimer, even more diverse set of growth factors: high affinity EGF and low affinity EREG, also failed to stabilize different dimer conformations [29]. Altogether, these structural analyses show that EGFR and HER4 receptors sample a wider selection of active homodimers states that can be exploited by different ligands and might be better conduits for ligand-specific signaling responses that their respective HER2 heterodimers.

Receptor glycosylation. N-linked glycosylation of receptor tyrosine kinases plays a crucial role
in their maturation, stability and regulation of their interaction with ligands, the extracellular matrix and other membrane proteins [54,55]. HER receptors are heavily glycosylated and their aberrant glycosylation patterns have been associated with diseases such as cancer and promoting drug resistance [56][57][58][59]. Glycosylation patterns on HER receptors can modulate their dimerization propensity. For example, a mutation of N418 glycosylation site on HER3 promotes its ligandindependent association with HER2 [60]. Likewise, mutation of N579 on EGFR drives its ligandindependent activation as well as increases its affinity for ligands [61]. However, the molecular mechanisms behind most of these effects are poorly understood, mostly because the majority of HER ectodomain structures have been solved by X-ray crystallography using heavily deglycosylated receptor fragments [23,25,26]. Recently published cryo-EM analyses of HER receptor samples purified with intact glycosylation, also did not reveal insights into glycanmediated interactions, perhaps due to their flexible and/or heterogeneous nature in these complexes [29,42].
To our knowledge, the cryo-EM maps of HER4 homodimers and HER2/HER4 heterodimers offer the first glimpse into extensive glycan-mediated contacts in an active HER receptor dimer. The observed interactions might explain some of the stabilizing effects of HER receptor glycosylation previously suggested [62]. We identified glycans in both HER2 and HER4 that directly connect their domains I and II and the HER4-specific interdomain glycan connections between domains II and IV. Such glycosylation modifications would be expected to stabilize the extended conformations of HER2 and HER4 receptors, although their effect on tethered states cannot be excluded. Most remarkably, our structures of HER4 homodimer ectodomains reveal reasonably well-resolved glycans between N548 of one receptor monomer and N358 of the other, pointing to the potential importance of these interactions in stabilizing the homodimer. In contrast, we have not observed a direct inter-receptor connection for HER2/HER4 heterodimers due to the absence of respective glycosylation sites in HER2. It is tempting to speculate that the particularly strong propensity for HER4 homodimerization over heterodimerization with HER2 that we see in our reconstitution experiments is, at least partially, rooted in the missing glycan-mediated stabilization of the heterodimer.
In recent years several studies of the effects of HER receptor glycosylation on their structure and signaling has been conducted using molecular dynamics (MD) generating models on how glycans contribute to receptor stability and its interactions with the membrane [63][64][65][66]. Most recently, MD simulations conducted on the HER4/EGFR heterodimer models have suggested that the glycans present on HER4 N358 and N548, as well as EGFR N361 (which is equivalent to HER4 N358), form a connection in the dimerization interface that effectively stabilizes the heterodimer [62]. Our current study provides first direct experimental evidence that these glycan interactions are operative at the level of HER4 homodimers. Moreover, we analyzed the published EGFR cryo-EM maps [42] and noticed that the inter-receptor glycans also appear in EGFR homodimers between N361 and the more membrane-proximal N603. Altogether, our analysis points to an important role, and potential conserved mechanisms by which glycosylation contributes to the HER dimer interfaces ( Figure S12c-d).
In summary, our structural analysis provides new knowledge on HER receptor activation by their different growth factor ligands and mechanistic distinctions of their homodimeric versus heterodimeric pairings. Through this, our findings reveal greater aptitude of HER homodimers to differentiate between biased agonists, at least as compared to HER2-containing heterodimers.
Our structures for the first time reveal extensive intra and interdomain glycan contacts at the active HER dimer interface and have potential to further understanding of how glycosylation can be leveraged for the design of better HER-targeted therapeutics, and how it can contribute to drug resistance.        in cryoSPARC4 using a strategy in which particles are picked generously using template picker, selected by 2D classification to remove bad picks (<10% of particles) and then sorted via 2 rounds of heterogenous refinement into a HER receptor dimer template volume and 3 "junk" classes created from the impure particle stack. Picked particles were subjected to ab initio reconstruction to eliminate bias and further processed as shown. Figure S4. Processing workflow for the HER2/HER4/BTC structure. Data were processed in cryoSPARC2 using a strategy in which particles are picked generously using template picker, selected by 2D classification to remove bad picks (<10% of particles) and then sorted via 2 rounds of heterogenous refinement into a HER receptor dimer template volume and 3 "junk" classes created from the impure particle stack. Picked particles were subjected to ab initio reconstruction to eliminate bias and further processed as shown.     were processed in cryoSPARC2 using a strategy in which particles are picked generously using template picker, selected by 2D classification to remove bad picks (<10% of particles) and then sorted via 2 rounds of heterogenous refinement into a HER receptor dimer template volume and 3 "junk" classes created from the impure particle stack. Picked particles were subjected to ab initio reconstruction to eliminate bias and further processed as shown.

Materials and Methods
NRG1b and BTC expression and purification. NRG1b and BTC were expressed and purified as described previously for NRG1b [25,28,45]. was concentrated with a 3K cutoff and applied on an S200 10/300 increase column (GE Healthcare). Protein content of the major peak was stored in aliquots at -80 °C for subsequent receptor purifications.

Receptor expression.
Human HER2 was expressed as previously described [28]. HER2 with a C-terminal tail truncation (D1030-1255) followed by maltose binding protein (MBP) and twin-strep tags was cloned into pFastBac1 with a CMV promoter (Thermo Fisher Scientific preparation. For purification of liganded HER4 homodimers, the same purification protocol for the ligand-mediated receptor pulldown was followed. After elution from anti-Flag resin, the receptor was concentrated, crosslinked with glutaraldehyde and subjected to gel filtration as described above. Peak fractions corresponding to homodimer sample were pooled and concentrated with a 100 kDa concentrator to 0.1 µM for EM grid preparation or flash frozen in liquid nitrogen and stored at -80 °C.

Electron microscopy sample preparation and imaging. For negative stain EM, fractions
corresponding to heterodimer were applied to negatively glow-discharged carbon coated copper grids, stained with 0.75% uranyl-formate, and imaged on an FEI-Tecnai T12 with an 4k CCD camera (Gatan). The resulting negative stain micrographs were assessed for particle homogeneity and particle density. This analysis was used to determine the target concentration for cryo-EM with graphene oxide grids which typically required 2-5x negative stain concentrations.
poor classes and picks containing lines from visible graphene-oxide flakes. More than 90% of picks were selected and subjected to ab initio reconstruction into three classes. In all datasets, this resulted in "junk" classes without recognizable HER receptor features. To purify this particle set, all 2D-selected particles were subjected to two rounds of heterogenous refinement containing a HER receptor dimer volume (imported from previous datasets or obtained from this dataset in a previous round of processing using the same overall workflow) and 3 "junk" classes. Particles sorted into the HER receptor dimer volume were subjected to ab initio reconstruction into one or two classes, depending on which resulted in better resolution downstream, followed by heterogenous refinement (in two classes) and non-uniform refinement (see processing flow charts for sample-specific details). Once reasonable reconstructions were obtained (as judged by the FSC (Fourier Shell Correlation) curve shape), unbinned particles were re-extracted and subjected to ab initio reconstruction, heterogeneous refinement or 2D classification/selection and finally nonuniform refinement to achieve reconstructions with the highest resolution. Rosetta protocol in torsion space, refined once with PHENIX [70] real-space refinement and further modeled using iterative rounds of ISOLDE [71] and the FastRelax Rosetta protocol in torsion space [72][73][74]. Per atom B-factors were assigned in Rosetta indicating the local quality of the map around that atom. Glycans were built into the density onto a well-refined model using the Carbohydrate module in Coot [75] for mammalian proteins and refined with the Rosetta glycan refinement protocol [76]. After glycan addition, the model was once more refined in ISOLDE and the Rosetta FastRelax protocol in torsion space and main and side chains for domains I-III were inspected for final corrections in Coot. Model statistics were routinely assessed in PHENIX [70] and glycan geometries were cross validated in Privateer [77]. For the HER4/NRG1b homodimer, an initial model was created by placing HER4/NRG1b models from a crystal structure (PDB ID: 3U7U, HER4 chains C + D, NRG1b chains I + K) into the cryo-EM density and running the FastRelax Rosetta protocol in torsion space. The model was further refined using iterative rounds of ISOLDE and the FastRelax Rosetta protocol in torsion space.
Per atom B-factors were assigned in Rosetta indicating the local quality of the map around that atom. Glycans were built into the density onto a well-refined model using the Carbohydrate module in Coot [75] for mammalian proteins and refined with the Rosetta glycan refinement protocol [76]. After glycan addition, the model was once more refined in ISOLDE and the Rosetta indicating the local quality of the map around that atom. Model statistics were routinely assessed in PHENIX [70] and glycan geometries were cross validated in Privateer [77].
3D Classification Analysis. 3D classification analysis was performed using the heterogenous refinement function in cryoSPARC 4. For each hetero-or homodimer, the particle stacks from their final reconstructions were used as input for heterogenous refinement together with 4 identical respective volumes as initial models. The same respective atomic model was fit into the four resulting volumes from the classification using the Rosetta FastRelax protocol in torsion space resulting in four different models for each 3D classification. Models were aligned on the same chain for visualization and intermonomer angles for each torsion-relaxed model were measured as described below. To ensure robustness, the analysis was repeated using the same particles stacks for HER4/NRG1b and HER4/BTC homodimers but using starting volumes of the other homodimer (HER4/NRG1b particles, HER4/BTC starting volumes and vice versa). This yielded similar results indicating that 3D classes are being determined by the particles, not the starting volumes.
Structure analysis. UCSF ChimeraX was used to determine the interface residues, H-bonds and interface area between two chains of a model. The command to measure buried area between model 1 chain A and chain B is: measure buriedarea #1/A withAtoms2 #1/B). This interface area was then multiplied by 2 to obtain the total buried surface area (BSA) of both proteins. Prior to the measurements, hydrogens were added to all models (addh). BSA for HER receptor dimerization interfaces was done for residues 1-450. Polypeptide backbone overlays and determinations of RMSDs between two chains/models was performed using UCSF ChimeraX using the matchmaker command. RMSD values reported are across all pairs of the sequence alignment.
Intermonomer angles were determined using UCSF ChimeraX by defining an axis through monomer 1A and 1B and measuring the angle between the two axis (Commands: define axis #1/A, define axis #1/B, angle #1.2 #1.3). Glycans were removed from the polypeptide chain for this analysis.
Cell-based assays. Untagged full-length human HER4 was cloned into a pcDNA4 TM TO expression vector. Full-length human HER2 tagged with the C-terminal 3xFLAG tag in a pcDNA4 TM TO expression vector was kindly provided by Mark Moasser. HER2 I714Q, HER2 V956R, HER4 I712Q and HER4 V954R were introduced into pcDNA vectors by site-directed mutagenesis. Untagged HER2 and HER4 constructs, and their mutants, were further cloned into a pMSCV retroviral vector, kindly provided by James Fraser, using standard PCR methods and